An industry technology map generation method based on a large model

Abstract

The application discloses an industry technology graph generation method based on a large model, belongs to the field of knowledge graph and machine learning, and solves the shortcomings of the existing industry technology graph in terms of time sequence credibility, correlation traceability, anti-illusion and emerging technology inclusiveness by introducing source tracing annotation, calibrating entity evolution stages in combination with literature and patent time sequence data, fusing stage calibration confidence to quantize correlation credibility, and jointly determining the adoptability of objects through multidimensional signals, so that a structured industry technology graph with objective evidence support and interpretability can be generated to meet the needs of industry technology evolution deduction.

Description

Technical Field

[0001] This application relates to the fields of knowledge graphs and machine learning, and in particular to a method for generating industry technology graphs based on large models. Background Technology

[0002] The construction of industry technology maps relies on extracting technical entities and their relationships from heterogeneous evidence from multiple sources, such as project documents, papers, and patents, to form a structured knowledge graph that reflects the laws of technological evolution. Existing large-model-based construction methods typically directly output the evolution stage labels of technical entities and cross-domain relationships, and supplement the extraction results with confidence scores to filter out erroneous triples, or use authoritative technology dictionaries as hard constraints to suppress the illusion of large models.

[0003] However, in the methods described above, the evolution stage labels output by the large models lack binding and cross-validation with objective temporal evidence such as annual literature output and annual patent output, resulting in the temporal dimension of the graph being uncalibrable and susceptible to model randomness. Furthermore, the confidence scores for cross-domain associations do not correlate their source evidence with the credibility of the entity's evolution stage, nor do they record the specific evidence supporting the associations as traceable source information, making it difficult to quantify and accept the associations in the graph, and also impossible to verify their evidentiary basis afterward. In addition, the rigid constraints of authoritative dictionaries cannot accommodate emerging technologies not included in the dictionaries, and the confidence screening lacking a hard threshold for source evidence is insufficient to effectively suppress illusory nodes without factual support.

[0004] The aforementioned deficiencies result in significant shortcomings in the existing industrial technology map in terms of temporal reliability, traceability of associations, and the balance between anti-illusion and inclusion of emerging technologies, making it difficult to meet the requirements of objective evidence support and interpretability for industrial technology evolution projection. Summary of the Invention

[0005] This application provides a method for generating an industry technology map based on a large model, the technical solution of which is as follows: On the one hand, a method for generating industry technology maps based on large models is provided, the method comprising: The entity and association structures of multi-source heterogeneous industrial technology evidence are extracted and bound with source tracing labels to obtain a set of technology entities and a set of cross-entity associations; Based on the annual output sequence of documents and the annual output sequence of patents in the evidence fragments bound to each entity in the set of technical entities, the evolution stage calibration index is calculated, and the temporal state and stage calibration confidence of each entity are determined according to the evolution stage calibration index. The stage calibration confidence is used as a weighting factor and fused with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs in the cross-entity association set to generate a quantitative confidence of cross-domain association, and then bound to the association tracing list. The evidence co-occurrence signal weighted by the stage calibration confidence level, the source document support signal composed of the associated tracing list, and the hit signal of the authoritative frontier dictionary are used to jointly determine the admissibility of each object in the set of technical entities and the set of cross-entity associations, and generate a structured industry technology map under the premise of requiring source document support. Attached Figure Description

[0006] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 This is a flowchart of an industry technology map generation method based on a large model provided in an embodiment of this application; Figure 2 This is a flowchart of another method for generating an industry technology map based on a large model, provided in an embodiment of this application; Figure 3 This is a flowchart illustrating another method for generating an industry technology map based on a large model, as provided in this application embodiment. Figure 4 This is a flowchart of another method for generating an industry technology map based on a large model, provided in an embodiment of this application; Figure 5 This is a flowchart of another method for generating an industry technology map based on a large model, provided in an embodiment of this application. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0009] In this application, the terms "first," "second," etc., are used to distinguish identical or similar items with essentially the same function. It should be understood that there is no logical or temporal dependency between "first," "second," and "nth," nor are there any restrictions on quantity or execution order.

[0010] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0011] Traditional methods for constructing industry technology maps suffer from several shortcomings when extracting technological entities and their relationships from heterogeneous multi-source evidence. These include a lack of binding and cross-validation between the evolutionary stage labels output by the large model and objective temporal evidence, leading to an uncalibrated temporal dimension that is susceptible to model randomness. Furthermore, the confidence scores for cross-domain associations fail to correlate the source evidence with the credibility of the entity's evolutionary stage, and do not record the specific evidence supporting the associations as traceable source information. This makes it difficult to quantify and accept the associations in the map, and also makes it impossible to verify their evidentiary basis afterward. In addition, the rigid constraints of authoritative dictionaries cannot accommodate emerging technologies not included in them, and the lack of a hard threshold for confidence screening based on source evidence makes it difficult to effectively suppress illusory nodes without factual support. These deficiencies result in significant shortcomings in existing industry technology maps regarding temporal credibility, traceability of associations, and the balance between preventing illusions and including emerging technologies, making it difficult to meet the requirements of objective evidence support and interpretability for industry technology evolution projection.

[0012] To address this, this application proposes a method for generating industry technology maps based on a large model. Taking the implementing entity as an example, see [link to relevant documentation]. Figure 1 The method includes: 101. Perform structured extraction of entities and associations from multi-source heterogeneous industrial technology evidence and bind source tracing labels to obtain a set of technical entities and a set of cross-entity associations.

[0013] 102. Based on the annual output sequence of documents and the annual output sequence of patents in the evidence fragments bound to each entity in the technology entity set, calculate the evolution stage calibration index, and determine the temporal status and stage calibration confidence of each entity according to the evolution stage calibration index.

[0014] 103. Using the calibration confidence level of this stage as a weighting factor, it is fused with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs in the cross-entity association set to generate a quantitative confidence level of cross-domain association, and then bound to the association traceability list.

[0015] 104. Using the evidence co-occurrence signal weighted by the confidence level of this stage, the source document support signal composed of the related tracing list, and the hit signal of the authoritative frontier dictionary, the admissibility of each object in the set of technical entities and the set of cross-entity associations is jointly determined, and a structured industrial technology map is generated under the premise of requiring source document support.

[0016] For ease of understanding, the following explains some key terms in this embodiment: Multi-source heterogeneous industrial technology evidence: refers to industrial technology-related data from different channels and in different formats, such as project documents, academic papers, patent documents, industry reports, etc.

[0017] Structured extraction of entities and relationships: refers to the identification of technical concepts (entities) with specific meanings and their relationships (relationships) from unstructured or semi-structured text data, and their representation in a structured form.

[0018] Source tracing and annotation: refers to attaching the original evidence and excerpts of the original text to the extracted entity or related information to facilitate tracing and verification.

[0019] Technical entity set: refers to the set of all technical entities obtained after structured extraction.

[0020] Cross-entity association set: refers to the set of relationships between all technical entities obtained after structured extraction.

[0021] Annual output sequence of literature: refers to the statistical sequence of the number of publications or citations of a certain technology entity in different years.

[0022] Patent annual output sequence: refers to the statistical sequence of the number of patent applications or grants of a certain technology entity in different years.

[0023] Evolutionary stage calibration index: refers to a comprehensive quantitative index used to measure the development stage of a technological entity. Its calculation takes into account time-series data such as literature, patent output, and citations.

[0024] Time sequence status: refers to the discrete development stage of a technology entity in its entire life cycle timeline, such as the nascent stage, growth stage, maturity stage, and decline stage.

[0025] Stage calibration confidence level: refers to the degree of reliability of the determination of the timing status of a technical entity, reflecting the reliability of the determination result.

[0026] Evidence co-occurrence strength: refers to the frequency or degree of correlation of two technical entities appearing together in multi-source heterogeneous industrial technology evidence.

[0027] Semantic affinity: refers to the degree of similarity or relevance between two technical entities at the semantic level, which is usually quantified through natural language processing techniques.

[0028] Quantitative confidence level of cross-domain association: refers to the quantitative assessment of the reliability of the association between cross-domain technical entities.

[0029] Linkage traceability list: refers to the collection of original evidence identifiers and text fragments that support a specific relationship, used to provide the evidence chain of the relationship.

[0030] Evidence co-occurrence signal: refers to the signal generated based on the strength of the co-occurrence of technical entities or associations in evidence, used to assess its admissibility.

[0031] Supporting signals from source documents: These are signals generated based on the quantity or quality of source documents associated with a technical entity or related entity, used to assess their adoptability.

[0032] Authoritative or cutting-edge dictionary hit signal: This refers to the signal generated by the preset authoritative or cutting-edge technology dictionary to assess the admissibility of a technical entity name.

[0033] Adoptability: refers to whether a technology entity or related entity meets pre-defined quality and credibility standards, thereby being included in the industry technology map.

[0034] Structured industry technology map: refers to an industry technology knowledge system presented in a graphical way, in which technology entities are represented as nodes, technology relationships are represented as edges, and include attribute information such as time sequence and strategic value.

[0035] This embodiment provides a method for generating industry technology maps based on a large model. It involves structured extraction of entities and associations from multi-source heterogeneous industry technology evidence, followed by source tracing annotation to obtain a set of technical entities and a set of cross-entity associations. Specifically, the structured extraction of entities and associations can employ methods based on manual rules or pattern matching, identifying technical entities and their relationships from text using preset keywords and grammatical structures. Source tracing annotation simply associates the extraction results with identifiers in the original document. For example, when processing a patent document, the technical names and their inventive relationships can be manually identified, and this information can be associated with the patent application number.

[0036] Based on the annual output sequence of literature and the annual output sequence of patents in the evidence fragments bound to each entity in the set of technical entities, an evolutionary stage calibration index is calculated, and the temporal state and stage calibration confidence of each entity are determined according to this index. Specifically, the evolutionary stage calibration index can be calculated by performing a simple linear weighted summation of the number of literature publications and patent applications of the technical entity in different years. Based on the numerical range of this index, the technical entity is divided into several preset temporal states, such as early, middle, and late stages. The stage calibration confidence can be set to a fixed value, or it can be simply mapped according to the absolute size of the index. For example, for a technical entity, the number of papers published and patents applied for each year in the past five years can be counted, and these numbers can be simply added together as its annual output index. The evolutionary stage can then be determined based on the growth or decline trend of this index.

[0037] Therefore, the stage calibration confidence level is used as a weighting factor and fused with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs in the cross-entity association set to generate a quantitative confidence level for the cross-domain association, which is then linked to an association source list. Specifically, the evidence co-occurrence strength can be simply obtained by counting the number of times two technical entities co-occur in the same document. Semantic affinity can be determined by calculating the similarity of the term frequency-inverse document frequency (TF-IDF) vectors of the two entity name texts. A simple linear combination of the stage calibration confidence level, evidence co-occurrence strength, and semantic affinity generates the quantitative confidence level for the cross-domain association. The association source list can record only the identifiers of the original documents supporting the association. For example, when two technical entities frequently co-occur in multiple documents and their names have high semantic similarity, a strong association can be considered to exist between them. In this case, the quantitative confidence level of the association can be adjusted by combining the confidence levels of the evolution stages of each of the two entities.

[0038] The admissibility of the set of technical entities and the objects in the cross-entity association set is jointly determined by the evidence co-occurrence signal weighted by the calibration confidence level at this stage, the source document support signal constituted by the association tracing list, and the authoritative frontier dictionary hit signal, and a structured industry technology map is generated under the premise of requiring source document support. Specifically, the evidence co-occurrence signal can be directly converted from the evidence co-occurrence strength. The source document support signal can be generated based on the number of documents included in the association tracing list. The authoritative frontier dictionary hit signal can be a Boolean value indicating whether the technical entity name is in a preset dictionary. These signals can be assigned preset weights and then weighted and summed, and the total score is compared with an adoption threshold to determine the admissibility of technical entities and associations. When generating the map, only objects that are determined to be admissible and supported by at least one source document are included. For example, if a technical entity is mentioned in a large number of documents, its name is included in an authoritative dictionary, and its evolution stage has a high confidence level, its likelihood of being adopted will increase.

[0039] It should be noted that the above basic embodiments are only used to illustrate the minimum feasible implementation of the method of this application and are not intended to limit the scope of protection of this application. In the preferred embodiment, the structured extraction of entities and associations adopts the above-mentioned prompting engineering method based on a large language model, and traceability is achieved through binding fields. The calculation of the evolution stage calibration index adopts a multi-dimensional weighted fusion method (document growth rate, patent growth rate, citation acceleration, authority ranking), rather than a simple linear summation. The stage calibration confidence is dynamically adjusted based on the consistency comparison between the quantitative index and the semantic judgment of the large language model, rather than a fixed value. The quantitative confidence of cross-domain associations uses the stage calibration confidence as a multiplicative weighting factor, rather than a simple linear combination. Admissibility adjudication uses a hard threshold of source document support as a pre-screening mechanism, rather than relying solely on confidence scores. The following embodiments will be described in detail around the above preferred scheme.

[0040] This application addresses the shortcomings of existing industrial technology maps in terms of temporal credibility, traceability, and anti-illusion and inclusiveness of emerging technologies by introducing source tracing and annotation, calibrating entity evolution stages by combining literature and patent time-series data, calibrating confidence metrics and correlation credibility during the fusion stage, and jointly adjudicating the admissibility of objects with multi-dimensional signals. This enables the generation of structured industrial technology maps with objective evidence support and interpretability, thus meeting the needs of industrial technology evolution projection.

[0041] The technical solution of this application is described in detail below with reference to specific implementation methods. The core inventive concept of this application lies in: deeply integrating the extraction process of multi-source heterogeneous industrial technology evidence with the source tracing and annotation mechanism, using objective time-series data such as annual output sequences of literature and annual output sequences of patents to provide quantitative calibration basis for the evolution stages of technical entities, independent of the large language model, and by using the stage calibration confidence as a multiplicative weighting factor for the cross-domain correlation quantitative confidence, the temporal credibility of the entity itself is transmitted to the credibility assessment of the correlation relationship, thereby forming a four-in-one closed-loop generation framework of "evidence binding → temporal calibration → confidence transmission → multi-signal adjudication". The overall technical process of this closed-loop framework is as follows: performing structured extraction and binding source tracing and annotation on multi-source heterogeneous industrial technology evidence to obtain a set of technical entities with complete traceability information and a set of cross-entity correlations. Using the annual output sequences of literature and annual output sequences of patents contained in the evidence fragments, the evolution stage calibration index is calculated from three dimensions: output growth rate, citation acceleration, and authority score, and the temporal status and stage calibration confidence of each technical entity are determined accordingly. The stage-calibrated confidence level is used as a weighting factor and fused with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs to generate a cross-domain association quantitative confidence level with stage-calibrated confidence weighting, which is then linked to an association traceability list. The admissibility of each object is collaboratively determined by three parties: the evidence co-occurrence signal weighted by the stage-calibrated confidence level, the source document support signal, and the authoritative cutting-edge dictionary hit signal. A structured industry technology map is generated under the constraint of a hard threshold for source document support. The specific implementation methods of each step are explained in detail below.

[0042] In the structured extraction of entities and relationships from multi-source, heterogeneous industrial technology evidence, directly relying on large language models for entity identification and relationship discovery may lead to issues with the accuracy, reliability, and traceability of the extraction results. Especially when dealing with massive and complex industrial technology evidence, ensuring that the entities and relationships output by the large language model can effectively support subsequent graph construction and clearly trace back to their original evidence sources is a significant challenge in generating high-quality industrial technology graphs.

[0043] In response, this application further proposes the above method, which involves structured extraction of entities and associations from multi-source heterogeneous industrial technology evidence and binding them with source tracing annotations to obtain a set of technical entities and a set of cross-entity associations. See [link to relevant documentation]. Figure 2 Specifically, it includes: 201. Construct a prompt message carrying entity category constraints and output format specifications, and input the prompt message and the multi-source heterogeneous industrial technology evidence into the large language model to drive the large language model to perform entity recognition and relationship discovery.

[0044] 202. Parse entity records and associated records from the output of the large language model. For each entity record and each associated record, verify whether it carries a binding field. The binding field is used to record the evidence identifier and original text fragment supporting the corresponding record.

[0045] 203. Form the technical entity set by combining the entity records carrying the binding field, form the cross-entity association set by combining the associated records carrying the binding field, and discard the records that do not carry the binding field.

[0046] For example, before driving a large language model to perform entity recognition and relationship discovery, it is necessary to construct a cue message carrying entity category constraints and output format specifications. This cue message aims to provide the large language model with clear instructions and guidance to ensure that its output meets the expected structure and content requirements. Entity category constraints can explicitly specify the types of entities to be extracted, such as technology names, products, companies, people, events, etc., thereby limiting the recognition scope of the large language model to the key information required for building the industry technology map. The output format specifications define the structured form of the large language model's output, such as JSON, XML, or other easily machine-parseable text formats, ensuring ease of subsequent processing. This cue message, along with multi-source heterogeneous industry technology evidence, is input into the large language model, enabling it to perform entity recognition and relationship discovery according to established rules while understanding the content of the evidence.

[0047] The system parses entity and association records from the output of a large language model, verifying each record for the presence of a binding field. This binding field is crucial for ensuring the traceability of the extracted results; it records the evidence identifier and original text fragment supporting the corresponding record. The parsing process converts the large language model's text output into structured entity and association records according to a predefined output format specification. The verification stage checks for the presence of this binding field in each record and further verifies the validity of its content. For example, the evidence identifier should uniquely point to a specific record in the original multi-source heterogeneous industrial technology evidence, while the original text fragment should be text content directly extracted from the original evidence that directly supports the entity or association. This verification mechanism effectively identifies and filters out any "illusions" or unfounded extraction results that the large language model might produce.

[0048] Entity records carrying the binding field are grouped into a technical entity set, and associated records carrying the binding field are grouped into a cross-entity association set. Records without the binding field are discarded. This step involves data filtering and integration based on the verification results. Only entity records and associated records that have been verified and have clear traceability information are included in the technical entity set and the cross-entity association set. Records that fail to carry a valid binding field, no matter how reasonable their content may seem, will be discarded to avoid introducing unreliable or unverifiable information into the subsequent graph construction process.

[0049] Through the aforementioned technical solution, this application addresses the issues of accuracy, reliability, and traceability of extraction results when using a large language model for structured extraction. By providing refined prompts, the large language model is guided to specific domains and format requirements, improving the quality of entity recognition and relation discovery. More importantly, the introduction of binding fields and rigorous validation ensures that every extracted entity and relation can be traced back to its original evidence source, enhancing the transparency and credibility of the extraction results. This mechanism effectively avoids unfounded outputs that might arise from the large language model, providing a solid and reliable data foundation for subsequent evolutionary calibration, quantitative confidence generation, and the construction of industry technology maps, thereby enhancing the authority and practical value of the entire map.

[0050] When using large language models to extract entities and relationships from multi-source heterogeneous industrial technology evidence in a structured manner, without effective guidance, the large language models may have difficulty accurately understanding the professional context of the industrial technology field, resulting in insufficient professionalism and inconsistent format of the extraction results. Furthermore, it is difficult to effectively trace the source of the extracted entities and relationships, thereby affecting the accuracy and reliability of subsequent graph construction.

[0051] To address this, this application further proposes a step-by-step approach to structurally extract entities and relationships from multi-source heterogeneous industrial technology evidence and bind them with source tracing annotations to obtain a set of technical entities and a set of cross-entity relationships. Specifically, this includes: inputting a role-setting prompt into the large language model, constraining it to interpret the multi-source heterogeneous industrial technology evidence as an industrial technology analyst. Following the role-setting prompt, a format constraint prompt is input into the large language model, requiring it to classify and label the extracted technology names according to circle category labels and quadrant category labels, and to output the relationships in a triple format of head entity identifier, tail entity identifier, and relationship type. A binding instruction is embedded in the format constraint prompt, requiring the large language model to simultaneously attach the extracted evidence identifier and the original text fragment to each output field to constitute the bound field.

[0052] In this process, role-setting prompts are input into the large language model, constraining it to interpret the multi-source heterogeneous industrial technology evidence as an industry technology analyst. This aims to guide the large language model's interpretive perspective and knowledge focus towards the industrial technology field. By explicitly informing the large language model of its expected role, such as "You are a senior industry technology analyst; please identify key technology entities, their development trends, and interrelationships from the following text," the large language model will prioritize drawing upon its knowledge reserves in industry, technology, and innovation when processing multi-source heterogeneous industrial technology evidence. This allows it to more accurately identify entities and connections related to the construction of the industrial technology map, avoiding generalized interpretations or deviations from the professional context.

[0053] After the role setting prompt, a format constraint prompt is input into the large language model, requiring it to classify and label the extracted technology names according to tier category labels and quadrant category labels, and output the relationships in a triple format of head entity identifier, tail entity identifier, and relationship type. This is to ensure that the output of the large language model is highly structured and standardized. Tier category labels can be used to describe the position of a technology in the industrial ecosystem, such as "core technology," "supporting technology," and "application technology." Quadrant category labels can reflect the life cycle stage of a technology, such as "emerging technology," "mature technology," and "declining technology." These classification labels provide the foundation for subsequent graph visualization and analysis. Outputting relationships in a triple format of head entity identifier, tail entity identifier, and relationship type, such as "(artificial intelligence, applied to, intelligent manufacturing)," is a universal and efficient format for graph data storage and processing, simplifying subsequent graph construction and query operations.

[0054] The format constraint prompt embeds a binding instruction, requiring the large language model to simultaneously include the evidence identifier and the original text fragment used for extraction in each output field to form the binding field. This is to enable source traceability of the extraction results. The evidence identifier is a unique identifier for each piece of multi-source heterogeneous industrial technology evidence, and the original text fragment is the original text content on which the large language model extracts the entity or association. By outputting these two as binding fields along with each extracted entity or association, direct evidence can be provided for subsequent verification, validation, and user traceability, ensuring the transparency and credibility of the extraction results.

[0055] The above technical solutions address potential issues in large language models when processing domain-specific data, such as insufficient professionalism, inconsistent output formats, and difficulties in tracing origins. Role-setting prompts enable large language models to more accurately understand industry and technology contexts, improving the accuracy and professionalism of entity and association extraction. Format constraint prompts ensure the standardization and structuring of extraction results, facilitating subsequent industry technology map construction and data analysis. The embedding of binding instructions provides traceable source evidence and original text fragments for each extracted entity and association, enhancing the transparency and credibility of the extraction results. This prompting engineering approach, combining domain role-playing, structured output specifications, and tracing mechanisms, enables large language models to output high-quality, highly reliable, and easily processed structured data in industry technology map generation tasks, laying a solid foundation for constructing accurate and reliable industry technology maps.

[0056] In some embodiments described above in this application, a method is proposed to input format constraint hints and embed binding instructions into a large language model, requiring the large language model to simultaneously extract the evidence identifier and original text fragments on which it is based for each output field, thus forming a bound field. However, in practice, simply requiring the large language model to include this information may not ensure its accuracy and traceability. There is a risk that the large language model may generate inaccurate or difficult-to-verify traceability information, thereby affecting the subsequent data quality and the reliability of the graph construction.

[0057] To address this, this application further proposes setting a mandatory sourcing rule in the format constraint prompt, requiring the large language model to simultaneously reference the evidence identifier corresponding to the original text block generated in the output when it outputs entity names or relationships. A text anchoring rule is also set in the format constraint prompt, requiring the large language model to simultaneously extract the original text fragment supporting the corresponding output when it outputs entity names or relationships as the text fragment. The evidence identifier referenced by the large language model according to the mandatory sourcing rule, and the text fragment extracted according to the text anchoring rule, are combined to form the binding field.

[0058] For example, to address the aforementioned issues, this application proposes setting mandatory source tracing rules in the format constraint prompts. These rules aim to explicitly require the large language model to simultaneously refer back to the evidence identifier corresponding to the original text block that generated the output when outputting any entity name or association. For instance, the prompts can explicitly instruct the large language model to provide a unique identifier following each identified technical entity or discovered association, pointing to its specific source document within the multi-source heterogeneous industrial technical evidence. This can be achieved by requiring the large language model to use a specific format (such as the JSON field "source_id":"doc_XYZ") or by directly referencing the document number in the output text, ensuring that each extraction result has a clear document-level source.

[0059] This application proposes setting source text anchoring rules in format constraint prompts. These rules require the large language model to simultaneously extract the original text fragments supporting the corresponding output when outputting entity names or associations. This means the large language model must not only indicate the source document but also accurately provide the specific text content (e.g., a sentence, a phrase, or a paragraph) from the original document that directly supports the entity or association. For example, the prompt could require the large language model to provide a "text_snippet":..." field after outputting the entity or association, containing text extracted from the original evidence that directly supports the output. This facilitates human verification and understanding of the context of the large language model's extraction results.

[0060] This application combines the evidence identifiers referenced by the large language model according to the mandatory source tracing rule, and the original text fragments extracted according to the original text anchoring rule, to form a binding field. This binding field is a structured data unit that tightly links the evidence identifiers and the original text fragments, forming a complete source tracing information package. For example, the binding field can be a composite object containing two subfields, "evidence_id" and "original_snippet," ensuring that each extracted entity or association is accompanied by comprehensive and accurate source information.

[0061] The above technical solutions address the issues of inaccurate or unverifiable source information during the extraction process of large language models. Mandatory source tracing rules ensure that each extraction result can be traced back to a specific source document, while original text anchoring rules further provide supporting original text fragments, enhancing the accuracy and reliability of source tracing. Combining evidence identifiers and original text fragments into binding fields structures and standardizes source information, facilitating subsequent automated verification and manual checks. This not only reduces the risk of "illusions" generated by large language models and improves the quality of technical entity sets and cross-entity association sets, but also provides a solid and reliable data foundation for subsequent evolutionary stage calibration, association confidence generation, and graph adjudication, thereby enhancing the overall accuracy and reliability of the industry technology graph.

[0062] In the process of using large language models to structurally extract entities and relationships from multi-source heterogeneous industrial technology evidence, while the large language model can identify and output technical entities and relationships, the accuracy, completeness, and reliable binding of its output to the original evidence source are crucial for the subsequent construction of the industrial technology map. If the entity and relationship records output by the large language model fail to effectively carry or correctly link their supporting evidence, the map data will lack credibility, making it difficult to trace and verify its origins, thus affecting the overall quality and application value of the map.

[0063] To address this, this application further proposes specific steps for parsing entity records and associated records from the output of the large language model, and verifying whether each entity record and each associated record carries a binding field. These steps include: traversing the output of the large language model, extracting fields conforming to the entity definition format as entity records, and extracting fields conforming to the association definition format as associated records. For each extracted entity record and each associated record, checking whether the corresponding record's data structure contains the binding field. For records containing the binding field, verifying whether the evidence identifier in the binding field is traceable to the original record in the multi-source heterogeneous industrial technology evidence; marking traceable records as verified, and considering untraceable records as not carrying a valid binding field.

[0064] For example, in the process of traversing the output of the large language model and extracting fields that conform to the entity definition format as entity records and fields that conform to the association definition format as association records, the system scans the text or structured data generated by the large language model one by one according to preset parsing rules. These parsing rules can be based on regular expressions, JSON parsers, XML parsers, or other specific format parsing logic, aiming to identify and extract data fragments that conform to predefined entity structures (e.g., fields containing technology names, categories, etc.) and association structures (e.g., fields containing header entity identifiers, footer entity identifiers, relation types, etc.), and encapsulate them into independent entity records and association records respectively. For example, if the large language model is required to output in JSON format, the parser will identify and extract the parts marked as "entity" or "association" in the JSON object.

[0065] In the step of checking whether the data structure of each extracted entity record and each associated record contains the binding field, the system performs a structural check on each entity record and associated record extracted in the previous step. This check aims to confirm whether the data structure of each record explicitly contains a "binding field" used to store source tracing information. This binding field is usually a predefined key-value pair or data structure. For example, it could be a field named "source_id" to store evidence identifiers, and a field named "text_fragment" to store original text fragments. If a record lacks these predefined binding fields, its tracing information is preliminarily determined to be incomplete.

[0066] In the process of verifying whether the evidence identifier in the binding field of a record containing the binding field can be traced back to the original record in the multi-source heterogeneous industrial technology evidence, and marking traceable records as verified, while considering untraceable records as not carrying a valid binding field, the system further performs in-depth verification of the validity of the binding field for records that are initially confirmed to contain the binding field. Specifically, the system extracts the evidence identifier stored in the binding field and compares it with the original multi-source heterogeneous industrial technology evidence database. This can be achieved by querying a pre-built evidence index (e.g., a mapping from evidence identifier to original record). If the evidence identifier can successfully match a valid record in the original evidence database, it indicates that the source of the entity or association is real and traceable, and the record is marked as "verified." Conversely, if the evidence identifier cannot find a corresponding record in the original evidence database, or the record it points to is invalid, then even if the binding field exists, the record is considered "not carrying a valid binding field" because it cannot provide valid traceability support.

[0067] The above technical solution enables rigorous parsing and verification of entity and relationship records output by the large language model. Through standardized traversal and extraction, it ensures accurate identification of technical entities and relationships from the free-format or semi-structured output of the large language model. A mandatory check is performed on each record to ensure it contains binding fields, effectively filtering out data lacking traceability information. More importantly, by verifying whether the evidence identifiers in the binding fields can be traced back to the original evidence, this application ensures that each adopted technical entity and relationship has reliable source support. This enhances the credibility and transparency of data in the industry technology graph, avoiding data distortion caused by illusions or misreferences generated by the large language model. It provides a solid data foundation for subsequent evolutionary stage calibration, relationship confidence generation, and graph adjudication, thereby constructing a high-quality, verifiable industry technology graph.

[0068] In some embodiments described above in this application, an evolutionary stage calibration index is proposed to be calculated based on the annual output sequence of documents and the annual output sequence of patents in the evidence fragments bound to each entity in the technology entity set. The temporal state and stage calibration confidence level of each entity are then determined based on this evolutionary stage calibration index. However, in its implementation, if a comprehensive and multi-dimensional calculation method for the evolutionary stage calibration index is lacking, relying solely on the original output sequence may not accurately reflect the true development trend and influence of the technology entities, thus affecting the accuracy and confidence level of subsequent temporal state determination. This results in the generated industry technology map not depicting technology evolution with sufficient precision and reliability.

[0069] In response, this application further proposes a method for calculating evolution stage calibration indicators based on the annual document output sequence and annual patent output sequence in the evidence fragments bound to each entity in the technical entity set. See [link to relevant documentation]. Figure 3 Specifically, it includes: 301. For each technical entity in the set of technical entities, aggregate the document count, patent count, and cumulative citation count by year from all source traceability annotations bound to the corresponding technical entity, and form the annual output sequence of the document, the annual output sequence of the patent, and the cumulative citation sequence, respectively.

[0070] 302. Calculate the annual output growth rate of literature based on the output comparison between the recent window and the base window in the annual output sequence of literature, calculate the annual output growth rate of patents based on the output comparison between the recent window and the base window in the annual output sequence of patents, and calculate the citation acceleration based on the cumulative citation sequence.

[0071] 303. Extract the authority score from the source tracing annotation and determine the percentile ranking of the authority score among all technical entities. Normalize the annual output growth rate of the document, the annual output growth rate of the patent, the citation acceleration and the percentile ranking respectively, and then perform a weighted sum according to the preset sub-indicator weights. Use the weighted sum as the calibration index for this evolution stage.

[0072] To calculate the aforementioned evolutionary stage calibration indicators, data aggregation is required for each technical entity in the set of technical entities. For example, the system iterates through all source tracing annotations associated with each technical entity, which contain detailed information about literature or patents. By statistically analyzing these annotations by year, the system can obtain the number of publications (forming an annual literature output sequence), the number of patent applications or grants (forming an annual patent output sequence), and the cumulative number of citations up to that year (forming a cumulative citation sequence) for that technical entity in different years. For instance, for a technical entity called "quantum computing," the system collects all literature and patents mentioning "quantum computing" and calculates its output and cumulative citations for each year, such as 2010 and 2011, thereby constructing time-series data reflecting its historical development trajectory.

[0073] Building upon this foundation, to quantify the dynamic development trends of technological entities, it is necessary to calculate their output growth rate and citation acceleration. The calculation of the annual output growth rate of literature typically involves selecting a recent time window (e.g., the last three years) and a base period time window (e.g., the three years prior to the recent window), and comparing the relative changes in the average annual output of literature within these two windows. Similarly, the annual output growth rate of patents also employs the same or similar time window comparison method to reflect the changing trends in patent activity. The calculation of citation acceleration focuses on the rate of change in the cumulative citation sequence. For example, it can be achieved by calculating the difference between the cumulative citation value of the current year and the previous year, and the difference between the cumulative citation value of the previous year and the year before that, and then comparing these two differences to measure whether the rate of citation growth is accelerating or slowing down. These dynamic indicators can effectively reveal whether a technological entity is in a period of rapid development, stable growth, or a period of slowing decline.

[0074] To further enhance the comprehensiveness and reliability of the evolution stage calibration indicators, this application also introduces an authority score. The authority score extracted from source tracing and annotation can be evaluated based on multiple dimensions such as the journal impact factor of the literature, the citation count of the patent, or the size of the patent family. These authority scores are ranked among all technical entities, and a percentile ranking is determined for each technical entity to reflect its relative importance and influence in the entire technical field. The calculated annual output growth rate of literature, annual output growth rate of patents, citation acceleration, and percentile ranking of the authority score are normalized to a uniform numerical range (e.g., 0 to 1). Based on preset sub-indicator weights, these normalized indicators are weighted and summed; the result is the evolution stage calibration indicator for that technical entity. These preset weights can be adjusted according to actual application scenarios and expert experience to highlight the importance of different indicators in assessing the technological evolution stage.

[0075] The aforementioned technical solution provides a comprehensive and multi-dimensional evolutionary stage calibration index for each technological entity. This index considers not only the entity's historical accumulation and dynamic growth in literature and patent output, but also its influence in academia and industry (reflected by citation acceleration and authority score). This comprehensive calculation method overcomes the limitations of single indicators, enabling the evolutionary stage calibration index to more accurately and precisely reflect the true development trend and life cycle stage of a technological entity. Therefore, the time series status and stage calibration confidence determined based on this index will be more reliable, thereby improving the accuracy and guiding value of the generated industry technology map in depicting technological evolution and providing users with more precise decision support.

[0076] In some embodiments described above in this application, an evolution stage calibration index is proposed based on the annual output sequence of documents and the annual output sequence of patents from the evidence fragments bound to each entity in the technology entity set. However, in its implementation, how to accurately and stably quantify the annual output growth trend of documents and patents, and how to effectively capture the acceleration or deceleration of citations of technology entities, are crucial to ensuring the accuracy and robustness of the evolution stage calibration index. Without a clear calculation method, the index may be overly sensitive to short-term fluctuations or fail to effectively reflect the true dynamics of technological evolution.

[0077] In this regard, this application further proposes specific methods for calculating the annual output growth rate of the document, the annual output growth rate of the patent, and the citation acceleration.

[0078] For example, to accurately measure the growth trend of a technology entity's literature output, this implementation divides the annual literature output sequence into a recent window and a base period window. The recent window typically refers to the most recent few years (e.g., the last 3 or 5 years) to reflect the current activity level of the technology entity. The base period window refers to the few years preceding the recent window (e.g., the last 5 or 10 years) to provide a stable benchmark. By calculating the average annual output of literature within these two windows, and using the relative change between the average annual output of literature in the recent window and the average annual output of literature in the base period window as the annual output growth rate, short-term fluctuations between years can be effectively smoothed, and the long-term growth or decline trend of a technology entity in terms of literature output can be reflected more stably. This method of calculating relative change avoids misjudgments caused by excessively large differences in absolute quantities, making the growth rates of technology entities at different stages of development comparable.

[0079] Similarly, this implementation method also divides the annual patent output sequence into a recent window and a base period window. Patent output is an important indicator for measuring the activity of technological innovation and industrialization potential. By calculating the relative change between the average annual patent output in the recent window and the average annual patent output in the base period window as the annual patent output growth rate, this helps to assess the dynamic changes of technological entities in terms of innovation achievement transformation and intellectual property layout, providing a basis for judging their technological maturity and market competitiveness. This method can also effectively filter out the random fluctuations in the number of annual patent applications, providing a more robust assessment of growth trends.

[0080] Furthermore, citation acceleration aims to capture the "rate of change" of a technology entity's citation trend, i.e., the speed at which its popularity increases or decreases. This implementation reads the first cumulative citation value for the current year, the second cumulative citation value for the year preceding the current year, and the third cumulative citation value for the year preceding the current year from the cumulative citation sequence. A first difference between the first and second cumulative citation values ​​is calculated, representing the growth rate of cumulative citations in the most recent year. A second difference between the second and third cumulative citation values ​​is calculated, representing the growth rate of cumulative citations in the year prior to the previous year. The difference between the first and second differences is used as the citation acceleration. This second-order difference calculation method can accurately reflect the change in the growth rate of a technology entity's influence in academia or industry, i.e., its "acceleration" or "deceleration" state, providing deeper insights into judging technology popularity, development potential, and decline risk.

[0081] The above technical solutions provide a more accurate and robust method for calculating evolution stage calibration indicators. For example, by dividing the annual output sequence of literature and the annual output sequence of patents into recent and base period windows, and using the relative change in average annual output to calculate the growth rate, the interference of single-year data fluctuations on the judgment of growth trends is effectively avoided. This allows the growth rate indicator to more stably and accurately reflect the long-term dynamics of a technology entity in terms of knowledge accumulation and innovation output. By introducing three-point cumulative citation values ​​to calculate citation acceleration, this application can capture the second-order change in the citation trend of a technology entity, i.e., whether its influence growth rate is accelerating or decelerating. This provides a more refined quantitative basis for assessing technology popularity and development potential. These precisely defined indicators make the determination of subsequent evolution stages more reliable, thereby improving the accuracy and effectiveness of the entire industry technology map generation method and providing users with more insightful technology evolution analysis.

[0082] After calculating evolutionary stage calibration indicators based on quantitative metrics such as literature and patent output sequences, a key issue is how to accurately and reliably determine the temporal state of each technological entity and effectively assess its confidence level. Simply relying on quantitative indicators may not fully reflect the complexity of technological evolution, especially for emerging technologies, where data sparsity may lead to judgment biases, thereby affecting the accuracy and reliability of the industry technology map.

[0083] In response, this application proposes a method for determining the temporal status and stage calibration confidence of each entity. For example, based on the changing trends of the annual output sequence of literature and the annual output sequence of patents corresponding to each technical entity relative to their respective base period windows, this application determines whether the evolution trend of each technical entity is growth-oriented or slowing-downward-declining.

[0084] This step aims to identify the development trend of a technology entity by analyzing its historical activities in document publication and patent application. For example, based on the annual document output sequence, if the average annual document output in the recent output window is not lower than the average annual document output in the base period window, the document dimension is determined to be growing. Otherwise, it is determined to be declining. Similarly, the annual patent output sequence is assessed. When both the document and patent dimensions are growing, the evolution trend of the technology entity is determined to be growing. When at least one dimension is declining, it is determined to be a slowing decline. This trend determination provides an important qualitative basis for subsequent time-series state mapping.

[0085] The evolution stage calibration index is compared with a preset first threshold, and combined with the aforementioned evolution trend categories, each technological entity is mapped to a discrete evolution stage in the entire life cycle timeline, with the mapping result serving as the time-series state. This process combines quantitative calibration indicators with qualitative trend categories to more comprehensively determine the life cycle stage of a technological entity, such as the nascent stage, growth stage, maturity stage, or decline stage. For example, a series of threshold ranges can be set. When the evolution stage calibration index falls into a certain range and the evolution trend category is growth-oriented, it may be mapped to the growth stage. If the index is high but the trend is a slowing decline-oriented one, it may be mapped to the maturity stage or the decline stage.

[0086] To further improve the accuracy and confidence of temporal state determination, this application independently invokes a large language model during the structured extraction process to output evolution stage labels based on the semantic content of multi-source heterogeneous industrial technology evidence. This step leverages the powerful semantic understanding capabilities of the large language model to extract qualitative judgments about the evolution stage of technological entities from the original textual evidence, serving as independent verification of the quantitative indicator results. For example, the large language model can analyze textual information such as technology descriptions, research progress, and market applications, outputting labels such as "early exploration," "rapid development," and "widespread application."

[0087] The temporal state obtained by jointly mapping the evolutionary stage calibration index and the evolutionary trend category is compared with the evolutionary stage labels output by the large language model. When the comparison results are consistent, a stage calibration confidence level is calculated by adding an upward increment to a preset base confidence level. The upward increment monotonically increases with the number of source tracing annotations and does not exceed a preset upper limit. This mechanism rewards consistency between different evaluation methods and further enhances the confidence level based on the richness of supporting evidence, but sets an upper limit to avoid overconfidence. For example, if the base confidence level is 0.6, the confidence level can be increased to 0.9 when the two judgments are consistent and there are a large number of source tracing annotations.

[0088] When the comparison results show inconsistency, the time-series state is adopted as the defined evolution stage, and a consistency discount factor is applied to the preset base confidence level to reduce the stage calibration confidence level. Simultaneously, the technical entity is marked as requiring further verification. The consistency discount factor is configured such that the stage calibration confidence level calculated under inconsistency is not lower than a preset minimum confidence level, and the preset minimum confidence level is higher than the fallback confidence level assigned to entities without any source traceability support. This strategy ensures that data-driven quantitative judgment remains the primary method when disagreements exist, but uncertainty is reflected by reducing the confidence level, and entities are marked as requiring further verification for subsequent manual intervention or more in-depth analysis. For example, if the base confidence level is 0.6 and the discount factor is 0.5, the confidence level under inconsistency is 0.3, higher than the fallback confidence level of 0.1 for entities without evidence.

[0089] The aforementioned technical solution enables a more accurate and robust assessment of the evolutionary stages of technological entities by comprehensively utilizing data from two different dimensions: quantitative indicators and semantic analysis. By introducing consistency comparison and confidence adjustment mechanisms, the problem of judgment bias that may arise from solely relying on quantitative indicators is resolved. This is particularly beneficial for emerging technologies with sparse data, providing a valuable supplementary perspective for semantic analysis using large language models. Furthermore, marking inconsistent technological entities as requiring enhanced verification provides clear guidance for subsequent iterative processing or manual intervention, thereby improving the accuracy, reliability, and traceability of the temporal states of various technological entities in the industry technology map, making the generated map more valuable for reference.

[0090] In some embodiments described above in this application, a method is proposed to determine the temporal state and stage calibration confidence of each entity based on evolution stage calibration indicators. However, when a technical entity is marked as requiring enhanced verification, its stage calibration confidence will decrease due to a consistency discount factor. Although this approach can identify potential uncertainties, without further processing of these marked entities, some technical entities with practical value but insufficient initial evidence or biased model judgments may be underestimated or omitted, thereby affecting the comprehensiveness and accuracy of the industry technology map.

[0091] To address this, this application further proposes that after marking a technical entity as requiring enhanced verification, the following steps are also included: In subsequent iterations, for technical entities in this state, supplementary source tracing annotations are preferentially retrieved from multi-source heterogeneous industrial technology evidence. Based on the supplemented source tracing annotations, a consistency comparison is re-executed, and the stage calibration confidence level is updated according to the re-comparison results. When the updated stage calibration confidence level is not lower than a preset confidence threshold, the state requiring enhanced verification is lifted.

[0092] For example, in subsequent iterations, for technical entities in a state requiring enhanced verification, supplementary source tracing annotations are preferentially retrieved from multi-source heterogeneous industrial technology evidence. Here, "subsequent iterations" refers to the cyclical process within the overall workflow of the industrial technology map generation method, where the system processes and optimizes the data again after preliminary adjudication and map construction. For technical entities marked as requiring enhanced verification, the system initiates a priority retrieval mechanism. For instance, the entity's name, existing relationships, and industry domain can be used as search keywords to conduct a deeper and more comprehensive search within multi-source heterogeneous industrial technology evidence (such as patent documents, academic papers, news reports, and industry reports). This priority retrieval can employ various strategies, such as expanding the search scope, adjusting search weights, and utilizing more complex semantic matching algorithms, to discover more evidence markers and original text fragments related to the technical entity, thereby forming supplementary source tracing annotations.

[0093] Based on the supplemented source tracing annotations, a consistency comparison is re-performed, and the stage calibration confidence level is updated according to the re-comparison results. After obtaining the supplemented source tracing annotations, the system re-evaluates the evolution stage of the technical entity. This includes re-aggregating document counts, patent counts, and cumulative citation counts to form updated annual document output sequences and annual patent output sequences. Based on the updated output sequences, the evolution stage calibration index is recalculated, and its evolution trend category is determined again. The re-mapped temporal state is compared with the evolution stage labels output by the large language model. If the comparison results show that the consistency has improved or reached a consensus, the stage calibration confidence level of the technical entity can be increased, for example, by removing the previously applied consistency discount factor or by adding an upward increment based on the number of newly added source tracing annotations. Conversely, if the consistency has not improved, the confidence level may be maintained or fine-tuned, but at least a more comprehensive evidentiary basis is provided.

[0094] When the updated stage calibration confidence level is not lower than the preset confidence threshold, the pending verification status is lifted. The "preset confidence threshold" is a pre-defined value representing the system's minimum acceptable level of confidence in a technical entity. When a technical entity, after supplementary evidence retrieval and reassessment, reaches or exceeds this threshold in its stage calibration confidence level, it indicates that the technical entity has obtained sufficient evidence support and consistency verification, and its uncertainty has been reduced. At this point, the system will lift the pending verification status of the technical entity, treating it as a fully verified and reliable technical entity that can participate normally in subsequent stages such as map construction, visualization rendering, and value assessment, without requiring special attention or processing.

[0095] The above technical solution addresses the issue of some technical entities being marked as requiring enhanced verification and experiencing reduced confidence levels during the generation of industry technology maps due to insufficient initial evidence or biases in large language models. By prioritizing and targeting supplementary evidence retrieval for these entities in subsequent iterations, and re-performing consistency comparisons and confidence level updates based on more comprehensive source tracing annotations, technical entities with low confidence but potential value can receive sufficient verification and support. When their stage calibration confidence level reaches a preset threshold, their "requiring enhanced verification" status is lifted. This ensures that the generated industry technology map not only identifies potential uncertainties but also proactively improves the accuracy and reliability of key technical entities within the map, avoiding the omission or misjudgment of valuable information and enhancing the overall quality and practicality of the map.

[0096] While the above-described embodiments of this application enable the structured extraction of multi-source heterogeneous industrial technology evidence and the determination of the temporal status and stage calibration confidence of each technical entity, accurately assessing the reliability, effectiveness, and traceability of cross-domain relationships between technical entities remains a problem that needs to be solved. Simply identifying the relationships without a quantitative assessment of their confidence and a clear traceability of their sources will make it difficult to support the accurate construction and effective application of subsequent industrial technology maps.

[0097] In response, this application further proposes a method that uses the stage calibration confidence as a weighting factor and integrates it with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs in the cross-entity association set to generate a quantitative confidence of cross-domain association and bind it to the association tracing list.

[0098] For example, for any cross-domain technical entity pair within a cross-entity association set, it is necessary to calculate their evidence co-occurrence strength. Evidence co-occurrence strength aims to measure the frequency or intensity of two technical entities appearing together at the level of original evidence. It is calculated by using the source attribution annotations attached to each of the two technical entities as statistical basis, and calculating the cross-union ratio (CUNR) at the evidence level. The CUNR is a commonly used set similarity measure that objectively reflects how many common pieces of evidence mention two entities, thus providing quantitative support for the association strength based on original data. For instance, if two technical entities appear simultaneously in a large number of identical documents or patents, their CUNR will be high, indicating a strong evidentiary association between them.

[0099] To capture the implicit semantic relationships within the names of technical entities, this application utilizes a pre-trained semantic encoding model to perform word segmentation, token encoding, and context pooling on the name texts of the two technical entities, respectively, to obtain semantic feature representations for each entity. Word segmentation divides the text into meaningful tokens, token encoding converts these tokens into numerical representations, and context pooling, considering contextual information, aggregates the token encodings into fixed-length vectors, thus forming a representation capable of capturing the deep semantics of the entities. The cosine similarity between these two semantic feature representations is calculated as the semantic affinity. Cosine similarity measures the directional similarity between two vectors; a value closer to 1 indicates greater semantic similarity, effectively identifying potential connections based on textual descriptions and compensating for semantic relationships that may be missed by relying solely on co-occurrence statistics.

[0100] Building upon this foundation, to generate a quantitative confidence score for cross-domain association, this application normalizes the calculated evidence co-occurrence strength and semantic affinity to a uniform numerical range, such as 0 to 1. Furthermore, an independent source support number is introduced, representing the number of deduplicated source identifiers in the source tracing annotations bound to the association records involving both technical entities, reflecting the breadth of independent evidence sources for the association. The normalized evidence co-occurrence strength, semantic affinity, and independent source support number are weighted and summed according to a preset fusion weight to obtain a basic confidence score. This basic confidence score comprehensively reflects the evidence support strength, semantic relevance, and richness of independent evidence sources for the association itself. To further improve the accuracy and reliability of the quantitative confidence score, this application multiplicatively weights this basic confidence score with the stage calibration confidence score of the technical entity itself, using the multiplicative weighted result as the quantitative confidence score for cross-domain association. This multiplicative weighting mechanism ensures that the quantitative confidence level of the association not only reflects the strength of the association itself, but also fully considers the maturity and reliability of the technical entities at both ends of the association, thus avoiding the negative impact of low-confidence entities on the reliability of the association.

[0101] To ensure the traceability of cross-domain associations, this application compiles all source tracing annotations and corresponding original text fragments supporting the cross-domain association, forming a tracing list of the associations, and binds it to the cross-domain association. The tracing list of associations provides the original evidence chain for each association, including evidence identifiers and specific original text fragments, enabling users or systems to easily trace back to the original documents or patents supporting the association, thereby enhancing the credibility and verifiability of the graph.

[0102] The above technical solutions address the challenge of accurately assessing the reliability and traceability of cross-domain associations in the construction of industry technology maps. For example, calculating the co-occurrence strength of evidence objectively reflects the frequency of co-occurrence of technical entities at the original evidence level, providing solid data support for association strength. Utilizing a semantic coding model to calculate semantic affinity captures deep-seated semantic connections within the text of technical entity names, effectively identifying potential relationships based on textual descriptions and overcoming the limitations of relying solely on co-occurrence statistics. More importantly, by using the stage-calibrated confidence level of technical entities as a weighting factor to multiplicatively weight the base confidence level, the quantitative confidence level of the association not only reflects the strength of the association itself but also fully considers the maturity and reliability of the technical entities at both ends of the association. This avoids the negative impact of low-confidence entities on the reliability of the association, making the generated cross-domain association confidence level more accurate and reliable. Furthermore, by binding the linked traceability list, a clear chain of original evidence is provided for each cross-domain association, enhancing the credibility and verifiability of the map. Users can easily trace back to the original documents or patents that support the association, thereby improving the quality and practical value of the entire industry technology map.

[0103] When generating quantitative confidence scores for cross-domain associations, multiple dimensions need to be considered, including evidence co-occurrence strength, semantic affinity, and the number of independent sources supporting the data. However, these indicators may have different dimensions and numerical ranges, and directly weighting and summing them may lead to an overestimation or weakening of the influence of some indicators, thus affecting the accuracy and reliability of the confidence scores. Furthermore, how to effectively calibrate the confidence scores using the evolutionary stages of technical entities as weighting factors to more precisely reflect the reliability of the associations is also a problem that needs to be addressed.

[0104] To address this, this application further proposes using the stage calibration confidence level as a weighting factor, fusing it with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs in the cross-entity association set to generate a quantitative confidence level for cross-domain association, and binding it to an association source tracing list. Specifically, this includes: performing a first normalization operation on the evidence co-occurrence strength, a second normalization operation on the semantic affinity, and a third normalization operation on the number of independent source supports, mapping the three to the same numerical range. The normalized evidence co-occurrence strength, normalized semantic affinity, and normalized number of independent source supports are then weighted and summed according to the preset fusion weights to obtain the basic confidence level. Using the stage calibration confidence level as a global scaling factor, the basic confidence level is multiplicatively adjusted, and the adjusted result is used as the quantitative confidence level for the cross-domain association.

[0105] When calculating the quantitative confidence of cross-domain associations, evidence co-occurrence strength, semantic affinity, and independent source support are three important metrics. However, the original numerical ranges and dimensions of these metrics may differ significantly. For example, evidence co-occurrence strength might be a ratio between 0 and 1, semantic affinity might be a cosine similarity value, and independent source support is an integer count. To ensure that these heterogeneous metrics are treated fairly in subsequent fusion calculations and to avoid some metrics from having an unreasonable dominant influence on the results due to their large original numerical ranges, they need to be normalized. The first, second, and third normalization operations aim to linearly or non-linearly transform the original values ​​of their respective metrics into a preset uniform numerical range, such as [0,1] or [-1,1]. Common normalization methods include Min-Max normalization (scaling the data to a specified interval), Z-score normalization (converting the data to a distribution with a mean of 0 and a standard deviation of 1), or using non-linear transformations such as the Sigmoid function or the Tanh function. Through these operations, the values ​​of different indicators are standardized, eliminating the differences in dimensions and making them comparable in the subsequent weighted summation.

[0106] After normalizing the evidence co-occurrence strength, semantic affinity, and number of independent source supports, these indicators are within a unified numerical range and comparable. At this point, to comprehensively assess the preliminary reliability of the cross-domain association, they need to be fused. Weighted summation is a commonly used fusion method that allows assigning different importance to each indicator based on its contribution to the association's reliability. The preset fusion weights are a set of predetermined values, corresponding to the normalized evidence co-occurrence strength, semantic affinity, and number of independent source supports, respectively. These weights can be determined based on domain expert experience, historical data analysis results, or through training machine learning models (such as logistic regression, support vector machines, etc.). For example, if evidence co-occurrence strength is considered the most important factor in measuring association reliability, it can be assigned a relatively large weight. Through weighted summation, a comprehensive value, the basic confidence score, can be obtained, reflecting the preliminary credibility of the cross-domain association at both the evidentiary and semantic levels.

[0107] After obtaining the basic confidence level reflecting the strength of the association itself, this application further introduces an evolution stage calibration confidence level for the technical entities to fine-tune the basic confidence level. This stage calibration confidence level reflects the reliability assessment of the evolution stage of the technical entities at both ends of the association. For example, the reliability of the association between a technical entity in its early nascent stage may be lower than that between a technical entity in its mature development stage. Using this stage calibration confidence level as a global scaling factor means that it will multiplicatively adjust the entire basic confidence level, rather than simply as an additive term. This multiplicative adjustment mechanism is significant: it ensures that the basic confidence level is fully reflected only when both technical entities at both ends of the association have high stage calibration confidence levels. Conversely, if the stage calibration confidence level of either technical entity is low, even if the basic confidence level itself is high, the cross-domain association quantification confidence level will be effectively reduced. This allows the quantification confidence level to consider not only the strength of the association itself but also the reliability of the state of the technical entities involved in the association, thus providing a more comprehensive and robust association confidence assessment.

[0108] By performing first, second, and third normalization operations on the evidence co-occurrence strength, semantic affinity, and number of independent source supports, these heterogeneous indicators are uniformly mapped to the same numerical range, effectively eliminating the dimensional differences between them and avoiding biases caused by inconsistent numerical ranges. Based on this, a weighted summation using preset fusion weights can flexibly generate a comprehensive basic confidence level according to the contribution of each indicator to the reliability of the association. Furthermore, the stage calibration confidence level of the technical entities is multiplicatively adjusted as a global scaling factor, ensuring that the cross-domain association quantification confidence level not only reflects the evidentiary strength and semantic relevance of the association itself but also incorporates the reliability assessment of the evolutionary state of the technical entities at both ends of the association. This multiplicative weighting mechanism ensures that the association confidence level can only be highly evaluated when both technical entities at both ends are in a relatively reliable evolutionary stage, thus solving the inaccuracies that may arise from simple weighting, improving the accuracy and credibility of the cross-domain association quantification confidence level, and providing a more solid data foundation for subsequent graph adjudication and visualization.

[0109] It should be noted that the first, second, and third normalization operations described above can employ the same or different normalization methods. As an optional implementation, the evidence co-occurrence strength (crossover union ratio, ranging from [0,1]) and semantic affinity (cosine similarity, ranging from [-1,1]) can be mapped to the [0,1] interval using the Min-Max normalization method: x_norm=(x-x_min) / (x_max-x_min), where x_min and x_max are the minimum and maximum values ​​of the corresponding indicators among all candidates, respectively. The number of independent source supports (discrete integer counts) can be normalized using a logarithmic normalization method: x_norm=log(1+x) / log(1+x_max) to suppress the influence of long-tail distributions and prevent a few highly cited sources from excessively dominating the weighted results. The choice of the above normalization method can be flexibly adjusted according to the data distribution characteristics in the actual application scenario; this application does not impose specific restrictions on this.

[0110] In some embodiments described above in this application, a stage calibration confidence level is proposed as a weighting factor, which is then fused with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs in the cross-entity association set to generate a quantitative confidence level for cross-domain association. However, calculating this quantitative confidence level requires accurately obtaining the number of independent source supports. If the method for obtaining the number of independent source supports is not rigorous enough or contains redundancy, it may lead to a deviation in the calculation of the quantitative confidence level, thereby affecting the accuracy and reliability of the industry technology map.

[0111] To address this, this application further proposes a method for obtaining the number of independent source supports, specifically including: extracting all association records involving both technical entities from the association records of the cross-entity association set corresponding to cross-domain associations; tracing the source identifier in the source tracing annotation bound to each association record involving both technical entities, and performing deduplication processing on all the traced source identifiers; and using the number of source identifiers remaining after deduplication as the number of independent source supports.

[0112] For example, when obtaining independent source support data, it's necessary to extract all association records involving both technical entities from the cross-entity association set corresponding to cross-domain associations. This step aims to precisely filter out the raw data that directly supports a specific cross-domain association. For instance, for a cross-domain association connecting technical entity A and technical entity B, the system will traverse all association records in the cross-entity association set, identifying and extracting records whose head entity identifier and tail entity identifier match those of A and B. This precise filtering ensures that the data foundation for subsequent analysis is highly relevant.

[0113] For each related record involving both technical entities at both ends, the source identifiers in the source tracing annotations are traced, and deduplication is performed on all traced source identifiers. During structured extraction, each related record is assigned a source tracing annotation, which includes evidence identifiers supporting the record, pointing to the original document, patent, or other data source. Since different related records may cite the same original source, deduplication is required for all collected source identifiers to avoid duplicate counting. For example, if multiple related records cite the same patent document as evidence, the source identifier of that patent document will only be counted once after deduplication.

[0114] The number of source identifiers remaining after deduplication is taken as the number of independent source supports. This number directly quantifies how many different, independent sources of original evidence collectively support the existence of this cross-domain association.

[0115] Through the above technical solution, this application provides a rigorous and systematic method for obtaining the number of independent source supports. It accurately extracts associated records directly related to specific cross-domain associations, ensuring the accuracy of the data foundation for subsequent analysis. Strict deduplication is performed on the source identifiers in the source tracing annotations bound to these associated records, effectively avoiding duplicate counting caused by multiple citations of the same source, thereby eliminating the interference of redundant information on confidence calculation. The number of deduplicated source identifiers is used as the number of independent source supports, which more realistically and objectively reflects the strength of the evidentiary basis of cross-domain associations. This precise number of independent source supports, when integrated with evidence co-occurrence strength and semantic affinity, can improve the accuracy and reliability of the quantitative confidence of cross-domain associations, thus making the generated industry technology map more accurate and providing stronger data support for industry technology analysis.

[0116] In generating an industry technology map, it is necessary to admissibility decisions for each object in the set of technology entities and the set of cross-entity relationships. However, a key challenge is how to comprehensively consider multi-dimensional information to ensure the accuracy and reliability of the decisions, and avoid including objects lacking sufficient evidence or of insufficient importance in the map. If the decision-making mechanism is imperfect, the map may contain a large amount of noise or low-value information, affecting its practicality and credibility.

[0117] To address this, this application proposes a method for jointly adjudicating the admissibility of objects within a set of technical entities and a set of cross-entity associations. For example, see [link to relevant documentation]. Figure 4 The method includes: 401. For candidate associations in the cross-entity association set, generate source document support signals based on the number of different sources in the association traceability list bound to the candidate association.

[0118] 402. For candidate entities in the set of technical entities, generate source document support signals based on the number of different sources in the source tracing annotations bound to the candidate entity.

[0119] 403. Multiply the co-occurrence intensity of evidence corresponding to the candidate entity or candidate association by the stage calibration confidence level, and use the multiplication result as the evidence co-occurrence signal after stage calibration confidence level weighting.

[0120] 404. Load the pre-set authoritative frontier dictionary, perform fuzzy matching on the names of candidate entities, and generate authoritative frontier dictionary hit signals based on the matching results.

[0121] 405. The co-occurrence signals of evidence after source document support, stage calibration confidence weighting, and authoritative frontier dictionary hit signals are weighted and summed according to preset adjudication weights. The weighted summation result is compared with the preset adoption threshold to determine the admissibility of the corresponding object.

[0122] It should be noted that the aforementioned authoritative cutting-edge dictionary can be constructed and maintained in the following ways: (1) Extracting standardized technical names from the technical terminology databases of authoritative standards organizations (such as ISO, IEC, ITU, etc.). (2) Extracting emerging technology names from the technology maturity curves and cutting-edge technology reports published by well-known research institutions (such as Gartner, MIT Technology Review, etc.). (3) Extracting technical field terms from the classification definitions of patent classification systems (such as IPC, CPC, etc.). The dictionary can be stored using a Bloom filter or a Trie tree structure to support efficient fuzzy matching operations. Fuzzy matching can be achieved using various methods such as edit distance (Levenshtein Distance), Jaccard similarity, or semantic similarity based on the above semantic encoding model. When the matching similarity exceeds a preset threshold (e.g., 0.85), a hit signal is generated. In addition, the authoritative cutting-edge dictionary can be updated regularly (e.g., quarterly or semi-annually) to promptly include emerging technical terms and eliminate outdated terms, maintaining the timeliness and cutting-edge nature of the dictionary.

[0123] Specifically, for candidate associations within a cross-entity association set, a source document support signal is generated based on the number of different sources in the association tracing list bound to the candidate association. For candidate entities within a technical entity set, a source document support signal is generated based on the number of different sources in the source tracing annotations bound to the candidate entity. The source document support signal aims to quantify the number of independent sources of evidence possessed by a candidate association or candidate entity; the higher the number, the more reliable and persuasive the object. For example, for a candidate association, all source identifiers (such as document IDs, patent IDs, report IDs, etc.) can be extracted from its association tracing list, and then these identifiers are deduplicated and counted to obtain the number of different sources as its source document support signal. For a candidate entity, all source identifiers are extracted from its bound source tracing annotations and deduplicated counts are performed.

[0124] The co-occurrence strength of evidence corresponding to a candidate entity or association is multiplied by the stage-calibrated confidence level, and the result is used as the stage-calibrated confidence level-weighted evidence co-occurrence signal. The evidence co-occurrence signal reflects the frequency or intensity of the co-occurrence of technical entities or associations in the evidence, while the stage-calibrated confidence level reflects the reliability of the entity's temporal state. By multiplying the two, the aim is to weight the strength of evidence co-occurrence based on the reliability of the entity's temporal state, so that the evidence co-occurrence signals of entities or associations with more certain temporal states have a greater weight in the adjudication. For example, if the evidence co-occurrence strength of a candidate association is 0.8 and the stage-calibrated confidence level of its associated entity is 0.9, then the weighted evidence co-occurrence signal is 0.72.

[0125] The system loads a pre-defined authoritative frontier dictionary, performs fuzzy matching on the names of candidate entities, and generates an authoritative frontier dictionary hit signal based on the matching results. The authoritative frontier dictionary hit signal is used to determine whether a technical entity has been mentioned or included in industry-recognized authoritative sources (such as standards organizations, well-known research institutions, government work reports, etc.), or whether it belongs to currently recognized cutting-edge technologies. This authoritative frontier dictionary can be a database containing industry standard terms, lists of emerging technologies, keywords published by authoritative institutions, etc. Fuzzy matching can employ various text matching algorithms, such as edit distance, Jaccard similarity, or semantic similarity. When the matching similarity exceeds a preset threshold, a hit signal is generated.

[0126] The adjudication mechanism employs a weighted sum of source document supporting signals, evidence co-occurrence signals weighted by stage-calibrated confidence levels, and authoritative frontier dictionary hit signals, all weighted according to preset adjudication weights. This weighted sum is then compared to a preset adoption threshold to determine the admissibility of the corresponding object. This adjudication mechanism assigns weights to each signal (e.g., based on expert experience, historical data, or machine learning models for training and optimization), multiplies each signal's value by its corresponding weight, and sums the results to obtain a comprehensive score. If the comprehensive score is greater than or equal to the preset adoption threshold, the object (entity or related entity) is deemed admissible. Otherwise, it is deemed inadmissible.

[0127] Through the above technical solutions, this application provides a multi-dimensional, quantitative admissibility adjudication mechanism. The source document support signal ensures that each candidate object has sufficient independent evidence, effectively avoiding misjudgments based on single or weak evidence. The evidence co-occurrence signal, weighted by stage-calibrated confidence, incorporates the temporal reliability of entity evolution into the evaluation of evidence strength, allowing entities or associations with clearer and more reliable temporal states to receive more reasonable weight in the adjudication. The authoritative cutting-edge dictionary hit signal introduces verification from external authoritative knowledge, enabling the identification of industry-recognized or forward-looking technical entities. Weighting and summing these three types of signals according to preset weights and comparing them with the adoption threshold allows for a comprehensive and objective assessment of the reliability and importance of technical entities and cross-domain associations. This improves the quality and credibility of each object in the industry technology map, effectively filtering noise and low-value information, ensuring the generated map has higher accuracy and practicality, thus providing a more solid foundation for industry analysis and decision-making.

[0128] In the generation of industry technology maps, adjudicating the admissibility of objects within the set of technology entities and cross-entity associations is crucial. Existing methods typically determine admissibility by weighted summation of source document support signals, weighted evidence co-occurrence signals, and authoritative frontier dictionary hit signals, and comparing this summation to a preset admissibility threshold. However, this purely weighted summation approach can lead to a problem: even if a technology entity or association lacks basic, direct source document support, it may still be admissible if it scores highly on other signals (e.g., authoritative frontier dictionary hit signals). This can introduce speculative information lacking actual evidence support, thus affecting the reliability and traceability of the industry technology map.

[0129] To address this, this application further proposes setting a hard threshold for source document support when jointly adjudicating the admissibility of each object in the set of technical entities and the set of cross-entity associations. This hard threshold serves as a mandatory pre-screening mechanism, ensuring that all technical entities or cross-domain associations to be adjudicated must have at least one direct source evidence. For example, before performing subsequent weighted summation and threshold comparison on candidate objects, the system checks their source document support signal. If the signal indicates that the object's associated source tracing label (for technical entities) or association tracing list (for cross-domain associations) contains no source identifier, i.e., lacks any direct evidence support, the object will be immediately rejected and not allowed to proceed to the subsequent refined evaluation stage. This effectively prevents objects lacking basic evidence support from being adopted, thus ensuring the reliability of the graph data from the source.

[0130] For objects that have successfully passed the hard threshold of source document support, this method further quantifies their evaluation. Specifically, the system calculates a comprehensive confidence score by weighting the source document support signal, the evidence co-occurrence signal after stage-calibrated confidence level, and the authoritative frontier dictionary hit signal according to preset decision weights. This comprehensive confidence score fully reflects the object's overall credibility across multiple dimensions. The calculated comprehensive confidence score is then compared with a pre-set adoption threshold to determine whether the object is adopted.

[0131] This is the decision-making logic used to clearly determine whether a technical entity or cross-domain association should be adopted. When the overall confidence score is not lower than the preset adoption threshold, the object is considered reliable and supported by sufficient evidence, and therefore will be deemed adoptable and included in the generated industry technology map. Conversely, when the overall confidence score is lower than the preset adoption threshold, it indicates that the object's credibility is insufficient for adoption, and it will be deemed unadoptable and rejected, not included in the map. This decision-making mechanism ensures that only objects that have undergone rigorous multi-dimensional evaluation and meet the preset credibility standards can enter the map.

[0132] By introducing a hard threshold supported by source documentation, this method mandates that all candidates must have at least one direct source piece of evidence to support their adoption when adjudicating the admissibility of technical entities and cross-domain associations. This mechanism fundamentally solves the problem that a simple weighted summation might lead to the adoption of objects lacking basic supporting evidence, improving the traceability and reliability of all adopted information in the industry technology map. Building on this, further multi-signal weighted summation is applied to objects that pass the hard threshold and compared with the adoption threshold. This allows for a more refined and comprehensive evaluation of objects with basic evidence, ensuring that entities and associations included in the map are not only verifiable but also meet high credibility standards across multiple dimensions. This makes the generated industry technology map more robust and reliable, avoiding the introduction of speculative or unverified information, thereby providing users with high-quality, highly credible industry technology insights.

[0133] In some of the embodiments described above in this application, the admissibility of a technology entity or related entity is determined by comparing a comprehensive confidence score with a preset adoption threshold. However, in its implementation, relying solely on a single adoption or rejection logic may not be sufficient to distinguish the inherent characteristics of different types of technology objects. For example, for emerging technologies or technologies that are supported by authoritative dictionaries but lack sufficient evidence strength, a simple binary judgment may lead to the neglect or misjudgment of valuable information, thereby affecting the comprehensiveness and accuracy of the industry technology map.

[0134] In response, this application further proposes that for objects that meet the hard threshold of support from the aforementioned source documents and are not indicated by the hit signal of the aforementioned authoritative frontier dictionary, if the overall confidence score is not lower than the aforementioned preset adoption threshold, then the corresponding object is adopted and marked as an emerging technology. For objects that meet the hard threshold of support from the aforementioned source documents and are indicated by the hit signal of the aforementioned authoritative frontier dictionary, if the overall confidence score is lower than the aforementioned preset adoption threshold, then the corresponding object is downgraded to a dictionary candidate state and is temporarily not included in the generated industry technology map. For objects that meet the hard threshold of support from the aforementioned source documents and are indicated by the hit signal of the aforementioned authoritative frontier dictionary, if the overall confidence score is not lower than the aforementioned preset adoption threshold, then the corresponding object is adopted and included in the generated industry technology map.

[0135] For example, this technical solution aims to identify technology entities or related fields that have sufficient supporting evidence and a high overall confidence score, but are not yet included in authoritative frontier dictionaries. The system checks whether the technology entity or related field meets the hard threshold of source document support, i.e., it must be supported by at least one source piece of evidence. It then determines whether its authoritative frontier dictionary hit signal indicates a miss, meaning that the name of the technology entity or related field cannot be found in a pre-defined authoritative frontier dictionary. Based on this, if its overall confidence score reaches or exceeds a preset adoption threshold, the object is considered an emerging technology with high credibility but not yet widely recognized. At this point, the system adopts the object and specially marks it, setting its status to "emerging technology status" for subsequent analysis and display. This approach can effectively capture nascent technologies at the forefront of industry trends, providing early signals for technology trend prediction.

[0136] This application also addresses technical entities or related fields whose names are already included in authoritative frontier dictionaries, but whose overall confidence score fails to reach the adoption threshold. After checking for hard thresholds supported by source documents, if an authoritative frontier dictionary hit signal indicates that the object is included in the dictionary, but its overall confidence score is below the preset adoption threshold, it indicates that although the technical concept is mentioned at the authoritative level, its support in multi-source heterogeneous industrial technology evidence is insufficient. In this case, the system does not directly reject the object, but downgrades it to a "dictionary candidate state." Objects in this state will not be included in the generated industrial technology map for the time being, but will be retained in the candidate set for subsequent iterations or manual verification. This mechanism avoids simply discarding key concepts already appearing in authoritative dictionaries due to insufficient evidence, leaving room for possible future data supplementation or manual intervention.

[0137] Furthermore, this application describes a standard adoption process for technical entities or related technologies that are supported by authoritative frontier dictionaries and possess sufficient evidentiary strength and a comprehensive confidence score. When a technical entity or related technology passes a hard threshold check supported by source documents, and an authoritative frontier dictionary hit signal indicates that it has been included in the dictionary, while its comprehensive confidence score reaches or exceeds a preset adoption threshold, the system determines that the object is a highly credible and widely recognized technology. At this point, the object will be directly adopted and unconditionally included in the generated industry technology map. This ensures that the technical entities and related technologies included in the map are core industry technology elements with high confidence that have undergone multiple verifications.

[0138] Through the aforementioned technical solutions, this application introduces a more refined decision-making logic when adjudicating the admissibility of technical entities and their associations, addressing the limitations that may arise from relying solely on a single adoption threshold. For example, for technical objects supported by source documents with a hard threshold and a comprehensive confidence score not lower than the preset adoption threshold, but not included in authoritative frontier dictionaries, this application explicitly marks them as emerging technologies. This proactively identifies and captures early innovations and potential trends in industrial development, providing crucial information for forward-looking technology planning. For technical objects included in authoritative frontier dictionaries but with a comprehensive confidence score lower than the preset adoption threshold, this application downgrades them to dictionary candidate status rather than directly rejecting them. This avoids omitting potentially important technical concepts due to insufficient current evidence, providing a buffer mechanism for subsequent data supplementation, manual verification, or iterative optimization, thus improving the robustness and completeness of the graph. For technical objects supported by authoritative dictionaries and possessing high comprehensive confidence, they are directly adopted, ensuring the reliability of the core content of the graph. This multi-state, hierarchical adjudication mechanism ensures that the generated industry technology map not only includes mature and stable technologies, but also effectively reveals emerging technology dynamics and potential areas of interest, thereby enhancing the depth, breadth, and practical value of the map.

[0139] In some of the embodiments described above in this application, a structured extraction, evolutionary stage calibration, and joint adjudication of multi-source heterogeneous industrial technology evidence are proposed to identify and confirm technological entities and their cross-domain relationships. However, after these technological entities and relationships are adopted, how to effectively organize, visualize, and present this complex, multi-dimensional information to users so that they can intuitively understand the development trend of industrial technology, the strategic value of key technologies, and the reliability of relationships, while ensuring the traceability of information, is a problem that needs to be solved.

[0140] To address this, this application proposes a method for generating structured industry technology maps while requiring supporting source documentation; see [link to relevant documentation]. Figure 5 The method includes: 501. For each technical entity adopted after adjudication, determine its radial layout position according to its respective circle category, its angular layout position according to its respective quadrant category, and its timing display level according to its respective calibrated timing status.

[0141] 502. The strategic value of each technology entity adopted after adjudication is mapped to the node display scale. The strategic value is obtained by weighted fusion of the technology entity's technical influence reference value and the average quantitative confidence of the adjacent edges of the technology entity. The average quantitative confidence of the adjacent edges is the arithmetic mean of the quantitative confidence of all adjacent edges with the technology entity as the endpoint.

[0142] 503. Map the quantified confidence of each cross-domain related edge that has been adopted after adjudication to the edge rendering intensity, and store the related traceability list bound to each cross-domain related edge as a traceable label and associated with the corresponding edge to generate the structured industry technology map.

[0143] For example, when generating a structured industry technology map, the technology entities within the map need to be laid out. Each technology entity, after being adopted, is positioned radially according to its respective concentric circle category, angularly according to its respective quadrant category, and at a temporal display level according to its calibrated temporal state. The radial and angular positions together determine the spatial location of the technology entity nodes on the two-dimensional plane, typically used to represent the classification attributes of the technology entities. For instance, technology entities can be placed at different radial distances from the map center based on their concentric circle category (e.g., basic technology, core technology, applied technology), forming a concentric circle layout. They can also be distributed across different sector areas based on their quadrant category (e.g., technology maturity, market potential), visually distinguishing technology groups with different attributes. The temporal display level represents the evolutionary state of technology entities over time. For example, technology entities at different stages of evolution (e.g., nascent, growth, maturity) can be mapped to different display levels, or their development process can be visually displayed through color gradients, animation effects, etc.

[0144] To highlight the importance of key technological entities in the graph, the strategic value of each adopted technological entity after adjudication is mapped to the node display size. Strategic value is a comprehensive indicator measuring the importance and influence of a technological entity within the industry. It is calculated by weighting and fusing the entity's technological influence reference value with the average quantitative confidence score of its adjacent edges. The technological influence reference value reflects the entity's authority and cutting-edge nature, while the average quantitative confidence score of its adjacent edges reflects the reliability and strength of its connections with other technologies. By mapping strategic value to visual attributes such as node size, thickness, or color saturation, technological entities with higher strategic value occupy a more prominent visual space in the graph, thus guiding users to quickly identify technologies that play a crucial role in industrial development.

[0145] To visually demonstrate the reliability of relationships between technical entities, the quantified confidence scores of each cross-domain relationship edge adopted after adjudication are mapped to edge rendering intensity. Quantified confidence score is an indicator of the reliability of cross-domain relationships; a higher value indicates a more credible relationship. By mapping the quantified confidence score to rendering attributes such as edge thickness, color depth, or transparency, edges with high confidence scores are displayed more prominently in the graph, while edges with low confidence scores are relatively weakened, thus helping users quickly determine the strength and credibility of different relationships.

[0146] To ensure the transparency and verifiability of the graph information, a traceable list of each cross-domain related edge is stored as a traceable annotation associated with its corresponding edge. This traceable list records all original evidence identifiers and text fragments supporting the association, serving as direct proof of its reliability. By storing this list as traceable annotations associated with edges, users can click or query any related edge at any time while viewing the graph to obtain the original evidence information behind it, thereby enhancing the graph's credibility and supporting in-depth exploration and verification.

[0147] The aforementioned technical solutions transform complex industrial technology information into an intuitive, multi-dimensional, and traceable structured graph. By arranging technical entities radially, angularly, and chronologically, users can quickly understand the classification and evolution stages of these entities. Mapping strategic value to node display scale effectively highlights key technical entities, guiding users to focus on core innovations. Mapping the quantified confidence level of related edges to rendering intensity makes the reliability of associations readily apparent. The binding of the association traceability list provides transparent evidence, enhancing the graph's credibility and usability. Overall, this visualization approach not only improves users' efficiency in understanding the industrial technology graph but also provides a more solid and verifiable data foundation for strategic decision-making.

[0148] When generating a structured industry technology map, it is necessary to accurately assess the strategic value of each technology entity in order to reasonably map its display scale. However, assessments based on a single dimension (such as considering only technological influence or only the strength of association) often fail to fully reflect the true importance of technology entities, which may lead to an imbalance in the map's presentation of key or emerging technologies, affecting the map's decision support effectiveness.

[0149] To address this, this application further proposes a method for calculating strategic value. This strategic value is obtained by weighted fusion of the technical influence reference value of the technical entity and the average quantitative confidence score of its adjacent edges. Specifically, this includes: extracting an authority score from the source tracing annotation bound to the technical entity; determining the percentile ranking of this authority score among all adopted technical entities; normalizing this percentile ranking to obtain the technical influence reference value; identifying all adjacent edges with the technical entity as the endpoint; extracting the quantitative confidence score of each adjacent edge; calculating the arithmetic mean of the quantitative confidence scores of all adjacent edges to obtain the average confidence score of the adjacent edges; and weighting and summing the technical influence reference value and the average confidence score of the adjacent edges according to preset strategic weights, using the weighted sum as the strategic value.

[0150] For example, the system extracts an authority score from the source citations associated with the technology entity, determines its percentile ranking among all adopted technology entities, and normalizes this percentile ranking to obtain a reference value for the technology's influence, aiming to quantify the technology entity's own technological importance or authority. This process assesses the technology entity by analyzing the quality and influence of the original evidence it relies on. For each technology entity, the system traces all its associated source citations. These citations contain detailed information about the original evidence (such as literature, patents), which may include authority scores provided by external authorities or databases, such as journal impact factors, patent citation counts, expert review scores, etc. The system extracts these authority scores. To ensure a fair comparison among all adopted technology entities, these authority scores are converted into percentile rankings. For example, if a technology entity's authority score is higher than 90% of other technology entities, its percentile ranking is 90%. This percentile ranking is then normalized, typically by scaling it to between 0 and 1, to obtain a reference value for technological influence. Normalization can use linear scaling, the sigmoid function, or other suitable mapping functions.

[0151] The system identifies all adjacent edges with the technical entity as an endpoint, extracts the quantized confidence score of each edge, and calculates the arithmetic mean of all quantized confidence scores to obtain the mean edge confidence score. This mean score reflects the connectivity and centrality of the technical entity within the entire technical graph, i.e., the tightness and reliability of its association with other technical entities. For each technical entity in the graph, the system identifies all associated edges with that technical entity as a head or tail entity. These edges represent the relationships between that technical entity and other technical entities. For each such adjacent edge, the system extracts its pre-calculated quantized confidence score. The arithmetic mean of all these adjacent edge quantized confidence scores is then calculated to obtain the mean edge confidence score for that technical entity. This mean score reflects the overall connectivity quality of the technical entity in the network.

[0152] The strategic value is determined by weighting the reference value of the technological influence with the mean confidence score of adjacent edges according to preset strategic weights. This weighted sum is used to comprehensively consider the technological influence of the entity and its correlation strength within the technological network, resulting in a more comprehensive and accurate strategic value assessment. The calculated reference value of technological influence and the mean confidence score of adjacent edges are weighted and summed according to preset strategic weights. For example, a weight α can be assigned to the reference value of technological influence, and a weight (1-α) can be assigned to the mean confidence score of adjacent edges, where α is a value between 0 and 1. These strategic weights can be trained and optimized based on practical application scenarios, expert experience, or through machine learning models to reflect the relative importance of different dimensions in the strategic value assessment. The result of the weighted sum is the strategic value of the technological entity.

[0153] The aforementioned technical solution comprehensively considers both the technological authority of a technical entity and the strength of its connections within the technological network, avoiding the limitations of single-dimensional assessment. The technological influence reference value ensures an accurate measurement of the intrinsic quality and importance of the technical entity, while the average confidence score of the adjacent edges reflects the breadth and depth of its connections within the technological ecosystem. By weighted and fused together, a more comprehensive and robust strategic value assessment result can be generated. This multi-dimensional strategic value integration allows for a more precise highlighting of technical entities that possess both high technological content and play a crucial connecting role in the industrial ecosystem when mapping nodes to display scale in the industry technology map. This enhances the map's insight into technological development trends and its decision-making support capabilities.

[0154] It should be noted that the preset weights mentioned multiple times in this application (including but not limited to the preset strategic weights, preset fusion weights, preset adjudication weights, preset sub-indicator weights, etc.) can all be determined in one or more of the following ways: (1) Manual setting based on the experience of domain experts, for example, senior experts in the field of industrial technology analysis directly specify the weight values ​​based on their professional knowledge of the relative importance of each evaluation dimension. (2) Objective weighting based on historical data statistical analysis, for example, by performing regression analysis or correlation analysis on the labeled industrial technology map samples, the relative weights of each indicator are automatically determined. (3) Automatic learning based on machine learning models, for example, by using the weights as trainable parameters and automatically adjusting the weights by minimizing the prediction error on the training data through optimization algorithms such as gradient descent. In practical applications, one or a combination of the above methods can be selected. This application does not impose specific restrictions on the specific values ​​and determination methods of each preset weight, as long as they can reasonably reflect the relative importance of each evaluation dimension in the corresponding fusion calculation.

[0155] In the process of structurally extracting multi-source heterogeneous industrial technology evidence and generating industrial technology maps based on large models, even after preliminary entity and association adjudication, there may still be technical entities with different textual expressions or variations that are highly similar in semantics but are identified as independent entities by the system. This phenomenon may lead to redundant information in the generated industrial technology map, making the map structure less refined and unable to accurately reveal the internal connections and evolutionary patterns of industrial technologies, thus affecting the map's analytical efficiency and decision support capabilities.

[0156] To address this issue, this application proposes an entity normalization process performed after the aforementioned joint ruling. This process utilizes a pre-trained semantic encoding model to perform word segmentation and lexical encoding on the names of each entity in the technical entity set, obtaining semantic encoding representations for each entity and calculating the similarity of semantic encoding representations between pairs of entities. Pre-trained semantic encoding models typically refer to deep learning models trained on large-scale corpora, such as BERT and RoBERTa. These models, trained on massive amounts of text data, can learn deep semantic information of words and phrases and convert text information into semantic representations in a high-dimensional vector space. Word segmentation divides the entity name text into meaningful words or lexical units, for example, decomposing "artificial intelligence chip" into "artificial intelligence" and "chip." Lexical encoding maps these lexical units into numerical vectors. Through these processes, each technical entity name is converted into a vector that captures its semantic information. By calculating the similarity between these semantic encoding representations, for example using cosine similarity, the semantic closeness of different technical entity names can be quantified, thereby identifying potential synonyms or near-synonyms.

[0157] When the similarity between two entities is not lower than a preset similarity threshold, a connection edge is established between them in the entity similarity network. The preset similarity threshold is a configurable parameter used to define whether two entities are semantically close enough to be considered the same concept; for example, it can be set to 0.8 or 0.9. The entity similarity network is a graph structure where nodes represent technical entities and edges represent semantic similarity between entities. Based on this, a community partitioning based on modularity optimization is performed on the entity similarity network. Modularity optimization is a commonly used community detection algorithm that aims to divide the network into several communities, making the connections within each community dense and the connections between communities sparse. Through this partitioning, semantically highly related entities are clustered in the same community. Entities partitioned into the same community are merged into a canonical entity, with the name of the entity within the community that matches the authoritative frontier dictionary being selected as the name of the canonical entity. The canonical entity represents a unified concept for all semantically similar entities within the community. Prioritizing entity names existing in the authoritative frontier dictionary as canonical entity names ensures the accuracy, authority, and standardization of the canonical entity names. If no entity in the community matches an authoritative and cutting-edge dictionary, other strategies can be adopted, such as selecting the most frequently occurring entity name or having it confirmed by human experts, to ensure the rationality of the standardized entity name.

[0158] Before merging, cross-domain association edges with the original entities as endpoints are remapped according to the canonical entity. Duplicate association edges pointing to the same canonical entity pair after remapping are then aggregated. Remapping means redirecting all association edges originally connected to the merged entity to their corresponding canonical entity. For example, if entity A and entity B are merged into canonical entity C, then all association edges connected to entity A or entity B will now connect to canonical entity C. Since multiple original entities may be merged into a single canonical entity, this can lead to multiple association edges connecting to the same canonical entity pair (i.e., the head and tail entities are the same) after remapping. Therefore, these duplicate association edges need to be aggregated to eliminate redundancy and integrate information, ensuring the simplicity of the graph.

[0159] The above technical solutions effectively identify and integrate semantically repetitive or highly similar technical entities in the industry technology graph, significantly reducing graph redundancy. Pre-trained semantic encoding models are used for accurate semantic similarity calculations, combined with modularity-optimized community partitioning, ensuring the accuracy and rationality of the merging process. This allows each canonical entity in the graph to more comprehensively and accurately represent a unique industry technology concept. By remapping existing related edges to canonical entities and aggregating repetitive related edges, the integrity of the association information between technical entities in the graph is maintained, and the graph structure is further simplified, improving clarity and readability. This entity normalization process makes the generated industry technology graph more refined and accurate, more effectively supporting industry technology trend analysis, competitive intelligence mining, and strategic decision-making. It avoids information dispersion and analytical bias caused by entity redundancy, thereby enhancing the graph's practical value and insight.

[0160] It should be noted that the pre-trained semantic encoding models mentioned above can be pre-trained language models based on the Transformer architecture, such as BERT (Bidirectional Encoder Representations from Transformers), RoBERTa, and Sentence-BERT, which are encoder models pre-trained on large-scale general or domain-specific corpora. These models can capture long-distance dependencies and contextual semantic information between words in entity name text through self-attention mechanisms. In practical applications, domain-adapted models that have been fine-tuned for academic literature or patent texts can be preferred to improve the semantic encoding accuracy of industrial technical terms. Word segmentation divides entity names into subwords; for example, "solid electrolyte" can be segmented into "solid" and "electrolyte" or even finer-grained subword units. Lexical encoding maps each subword to an integer ID in the pre-trained vocabulary. Contextual pooling typically employs [CLS] labeling of the corresponding output vector or performs average pooling or max pooling on all word output vectors to obtain a fixed-dimensional semantic feature representation vector (e.g., 768-dimensional or 1024-dimensional). This semantic feature representation vector can learn the deep semantics of entity names from co-occurrence patterns in large-scale corpora, enabling semantically similar but differently expressed entity names such as "solid electrolyte" and "solid electrolyte" to have high cosine similarity in the vector space, providing a reliable semantic basis for subsequent entity normalization and merging.

[0161] The community partitioning based on modularity optimization described above can be implemented using the Louvain algorithm or the Leiden algorithm. Taking the Louvain algorithm as an example, this algorithm consists of two alternating phases: The first phase is a local optimization phase, where each node is initialized as an independent community. Then, each node is traversed, and the modularity increment caused by moving it from its current community to a neighboring community is calculated. The node is then moved to the neighboring community that maximizes the modularity increment, and this process is repeated until the modularity no longer increases. The second phase is a network aggregation phase, where each community obtained in the first phase is shrunk into a "supernode." The edge weights between supernodes are the sum of the edge weights between all the original communities, forming a new aggregated network. These two phases are iterated alternately until the modularity converges, yielding the community partitioning result of the network. The modularity is calculated as follows: Q = (1 / 2m) × Σ_ij(A_ij - k_i × k_j / 2m) × δ(c_i, c_j), where A_ij is the weight of the edge connecting node i and node j (semantic similarity), k_i and k_j are the weighted degrees of node i and node j respectively, m is the sum of the weights of all edges in the network, c_i and c_j are the community numbers to which node i and node j belong, and δ is the Kronecker delta function (1 when c_i = c_j, otherwise uniform). The modularity Q typically ranges from -0.5 to 1. A larger Q value indicates a denser connection within a community and a sparser connection between communities, meaning a higher quality community partitioning. Through this community partitioning, semantically highly similar technical entities are aggregated into the same community and then merged into a single canonical entity, effectively eliminating entity redundancy in the graph.

[0162] After normalizing technical entities and remapping their associated edges, multiple remapped cross-domain associated edges may point to the same pair of canonical entities. Without handling this, these duplicated edges not only cause data redundancy but may also increase the complexity of the graph, affecting subsequent analysis and visualization, and making it impossible to effectively integrate the support strength for the same association from different sources.

[0163] To address this, this application further proposes a method for aggregating duplicate association edges that point to the same canonical entity pair after remapping. Specifically, this includes: traversing all cross-domain association edges after remapping, identifying multiple edges with identical head and tail entity identifiers as duplicate association edges; comparing the quantization confidence scores of each of these duplicate association edges, selecting the highest quantization confidence score as the quantization confidence score of the aggregated edge; merging the association source lists bound to each of these duplicate association edges and removing duplicate source identifiers, using the merged and deduplicated result as the association source list of the aggregated edge.

[0164] For example, after entity normalization, the system needs to systematically check all remapped cross-domain association edges. This process is achieved by traversing all remapped cross-domain association edges. For each edge, its head entity identifier and tail entity identifier are the key to its unique identification. When multiple association edges are found to have the exact same head entity identifier and tail entity identifier, these edges are determined to be duplicate association edges. For instance, if "artificial intelligence" and "AI" are normalized to the canonical entity "artificial intelligence", and there is an association between "artificial intelligence" and "machine learning" in the original data, and "AI" and "machine learning" are also related, then after remapping, there will be two association edges pointing from the canonical entity "artificial intelligence" to "machine learning". These two edges are then identified as duplicate association edges.

[0165] After identifying duplicate association edges, this application compares the quantified confidence scores of each of these duplicate association edges to ensure that the aggregated association edges accurately reflect their intrinsic strength. Quantified confidence score is an important indicator of association reliability. Through comparison, the system selects the value with the highest quantified confidence score among these duplicate association edges as the quantified confidence score of the aggregated single association edge. This method of selecting the maximum value rather than the average or cumulative value aims to retain the strongest evidence supporting the association, avoiding inflated confidence scores due to the simple superposition of multiple weak pieces of evidence, or the dilution of the strength of key evidence due to the averaging of strong and weak evidence.

[0166] Furthermore, to ensure that the aggregated related edges still possess complete traceability capabilities, this application merges the traceability lists bound to each of the duplicate related edges. Each traceability list records the original evidence identifiers and original text fragments supporting the association. During the merging process, the system deduplicates the merged lists to ensure that each unique source identifier is recorded only once. The merged and deduplicated result serves as the traceability list for a single aggregated related edge. This guarantees that even when related edges are aggregated, users can still trace back to all original, independent sources of evidence, thereby enhancing the transparency and credibility of the graph.

[0167] Through the above technical solution, this application solves the problem of redundant related edges that may occur after entity normalization. Selecting the highest quantized confidence level as the confidence level of the aggregated edge ensures that the strongest evidence of the association is preserved, avoiding undue dilution of confidence levels, thereby improving the accuracy and reliability of the relationships in the graph. Merging and deduplicating the association traceability list allows the aggregated related edges to fully trace all their original supporting evidence, enhancing the credibility and interpretability of the graph. This not only optimizes the graph structure and reduces data redundancy but also improves the overall quality and analytical efficiency of the industry technology graph, making the generated graph more concise, accurate, and easier to understand.

[0168] In the process of generating industry technology maps based on large models, one-off decisions and map construction may not be able to fully capture the dynamics of technological evolution, especially for technological entities with limited initial evidence or those in a rapid development phase, where their temporal status and correlation confidence may be uncertain. Furthermore, the overall stability and accuracy of the map may be affected by data updates or biases in model understanding, resulting in a map that fails to fully reflect the latest trends and intrinsic connections within the industry.

[0169] To address this, this application further proposes an internal iterative processing method. This method includes: using the adopted technical entities and cross-domain associations after joint adjudication as input for the next iteration; reverting to the calculation process of the evolution stage calibration index; and supplementing the evolution stage labels for technical entities judged as emerging technologies by joint adjudication. The method then checks whether the current iteration meets preset convergence conditions. These preset convergence conditions include no new adopted technical entities or cross-domain associations in two consecutive iterations, and the change in the adoption score generated by each adopted object in the joint adjudication step being lower than a preset stability threshold between adjacent iterations. When the preset convergence conditions are met, the iteration terminates and the industry technology map generated in the current round is output; when the preset convergence conditions are not met, the next iteration continues.

[0170] For example, after completing the construction and adjudication of an industry technology map, the system does not immediately output results. Instead, it uses the technology entities that have been jointly adjudicated and adopted in this round, along with their cross-domain associations, as input for the next iteration. These adopted entities and associations are considered more reliable and stable information, providing a more solid foundation for subsequent analysis. This information is fed back into the calculation process of evolution stage calibration indicators. This means that the system will re-evaluate these technology entities, especially those that may have been initially identified as emerging technologies due to insufficient evidence or non-compliance with authoritative frontier dictionaries. For these emerging technology entities, the system uses the fed-back information, such as source tracing annotations that may be added during the iteration process, to recalculate their evolution stage calibration indicators and supplement or correct their evolution stage labels based on the updated indicators. For example, the system can re-traverse all source traceability annotations bound to the technical entity, re-aggregate the document count and patent count by year for annotations not included in the previous iteration, thereby updating the corresponding annual document output sequence and annual patent output sequence, and recalculating the evolution stage calibration index based on this, thus completing the evolution stage label for the technical entity.

[0171] To ensure the effectiveness and termination of the iterative process, this application also proposes a pre-defined convergence condition detection mechanism. This convergence condition aims to determine whether the graph has reached a stable and reliable state. Specifically, the pre-defined convergence condition includes two aspects: First, in two consecutive iterations, the number and content of adopted technical entities or cross-domain associations remain unchanged, meaning the graph structure no longer changes significantly. Second, the change in the adoption score generated by each adopted object in the joint adjudication step between adjacent iterations is lower than a pre-defined stability threshold, indicating that the confidence level of each element in the graph has stabilized. For example, the system can obtain the adoption score of each adopted object in the current iteration and the adoption score of the corresponding adopted object in the previous iteration, and for each adopted object, determine the absolute value of the difference between the current iteration's adoption score and the previous iteration's adoption score as its change range. When the maximum value of the change ranges of all adopted objects is lower than the pre-defined stability threshold, the condition is deemed satisfied.

[0172] Based on the detection results of the aforementioned convergence conditions, the system will determine the direction of the iteration. When the preset convergence conditions are met, it indicates that the industry technology map has reached the expected stability and accuracy. At this point, the system will terminate the iteration process and output the industry technology map generated in the current round as the result. Conversely, when the preset convergence conditions are not met, the system will continue to execute the next round of iteration, repeating the above backflow, calculation, and decision-making process until the map converges.

[0173] Through the aforementioned iterative processing, this application addresses the issues of insufficient dynamic capture, incomplete information on emerging technologies, and poor overall stability of the technology map that may arise from single-cycle generation. This mechanism allows the system to utilize richer and more stable information accumulated during the iteration process, particularly for more accurate completion of evolutionary stage labels for emerging technologies, thereby improving the map's ability and accuracy in capturing technological development dynamics. The introduction of convergence conditions ensures the stability of the map in terms of structure and confidence. When there are no significant changes after two consecutive iterations, it indicates that the map has stabilized, reducing uncertainties caused by data fluctuations or model fine-tuning, making the output industry technology map more robust and reliable. Furthermore, by setting explicit convergence conditions, the system terminates calculations promptly after reaching the expected stable state, effectively saving computational resources and time costs, and optimizing resource utilization. This iterative optimization method enables the generated industry technology map to better adapt to the dynamic updates and evolution of industry technology evidence. Even if the evidence is insufficient in the initial stage, the map can be gradually improved and optimized through multiple iterations, providing a comprehensive, accurate, and stable panoramic view of industry technology.

[0174] It should be further explained that the rationale for the aforementioned iterative processing mechanism lies in the fact that, in the reality of continuously updated multi-source heterogeneous industrial technology evidence, a single-time constructed industrial technology map often only reflects the technological situation at a certain time point. By reverting the adjudicated adopted objects back to the calculation process of the evolution stage calibration indicators, this application achieves the map's adaptive capability to evidence updates—each iteration is based on the reliable information confirmed in the previous round, and the remaining uncertain information is recalibrated. This progressive information confirmation strategy is analogous to the "draft-revision-final draft" working mode of human experts when writing technical review reports: first, a preliminary judgment is formed based on existing evidence, and then supplementary searches and verifications are conducted for uncertain parts until the overall conclusion tends to stabilize. In addition, the dual setting of convergence conditions (no new structures + convergence of score fluctuations) ensures that the iteration process does not fall into an infinite loop, but automatically terminates when the map stability reaches the expected level. This design enables this application to achieve a reasonable balance between computational efficiency and map quality.

[0175] In the iterative process of generating an industry technology map based on a large model, to ensure the stability and convergence of the map, it is necessary to determine whether the change in the adoption score generated by each adopted object in the joint decision-making step is lower than a preset stability threshold between adjacent iterations. However, accurately and comprehensively evaluating the score changes of all adopted objects to avoid misjudging the convergence state due to local fluctuations is crucial to ensuring the effective termination of the iteration process.

[0176] To address this, this application further proposes a method for determining that the change in the adoption score generated by each adopted object in the joint adjudication step is below a preset stability threshold between two adjacent iterations. For example, this method includes: obtaining the adoption score of each adopted object in the current iteration round, and the adoption score of the corresponding adopted object in the previous iteration round. The adoption score is a quantitative confidence level or comprehensive score calculated for each technical entity or cross-domain association by the joint adjudication step (such as jointly adjudicating the adoptability of each object in the set of technical entities and the set of cross-entity associations in the method described above), used to measure its suitability for inclusion in the industry technology map. After each iteration, the system stores the adoption scores of all adopted technical entities and cross-domain associations in the current round. This step aims to retrieve the adoption score of the current iteration round and the adoption score of the corresponding object in the immediately preceding iteration round from the stored historical data, providing the necessary data foundation for subsequent calculation of the change range. For example, the system can maintain an iteration history record, which includes the identifiers of all adopted objects and their corresponding adoption scores at the end of each iteration, thereby facilitating the acquisition of data from two adjacent iterations.

[0177] For each adopted entity, the absolute value of the difference between the adoption score in the current round and the adoption score in the previous round is determined as the magnitude of change for that entity. The magnitude of change is an indicator that measures the degree of fluctuation in the adoption score of an individual adopted entity between two adjacent iterations. By calculating the difference between the adoption score in the current round and the adoption score in the previous round, and taking its absolute value, the magnitude of the change in the entity's score can be quantified, regardless of whether it is an increase or a decrease. This step ensures an accurate assessment of the stability of each adopted entity. For example, if a technical entity has an adoption score of 0.85 in the current round and an adoption score of 0.82 in the previous round, then its magnitude of change is |0.85 - 0.82| = 0.03. This calculation process will traverse all technical entities and cross-domain associations that have been adopted in two consecutive iterations.

[0178] The condition that all changes in the adopted scores below the preset stability threshold are met when the maximum value among all adopted scores is lower than the preset stability threshold. The preset stability threshold is a pre-defined value representing the maximum acceptable fluctuation range of the adoption scores, used to determine whether the entire iteration process has reached a stable state. This step identifies the value with the largest change among all adopted scores and compares it to the preset stability threshold to assess overall convergence. The entire system is considered stable only when the score fluctuations of all adopted scores are less than or equal to this threshold. For example, if the preset stability threshold is 0.01, and the largest change among all adopted scores is 0.008, the condition is met. If the largest change is 0.015, the condition is not met. This maximum-value-based judgment method ensures that even a few highly fluctuating objects are captured, thus avoiding misjudgments caused by average values ​​masking local instability.

[0179] Through the above technical solution, this application provides a rigorous and comprehensive mechanism for determining iterative convergence conditions. This mechanism calculates the change in adoption score for each adopted object between adjacent iterations and uses the maximum value among all changes as a measure of overall stability, comparing it with a preset stability threshold. This method avoids the limitations of relying solely on average changes or changes in only a few objects to determine convergence, ensuring that even a few highly volatile technical entities or cross-domain relationships must reach a stable state before iterative convergence can be determined. This makes the generated industry technology map more robust in structure and content, avoiding inaccuracies or incompleteness caused by local fluctuations, thereby improving the quality and usability of the map and providing a reliable foundation for subsequent map analysis and applications.

[0180] To facilitate understanding, the above judgment process is illustrated below with specific numerical examples. Assume the preset stability threshold is 0.01. In the current iteration round, there are three adopted entities (Technical Entity A, Technical Entity B, and Cross-Domain Association C). Their adoption scores in the previous round were 0.82, 0.65, and 0.78, respectively, and their adoption scores in the current round are 0.83, 0.66, and 0.77, respectively. The magnitude of change for each entity is calculated as follows: Entity A's magnitude of change = |0.83 - 0.82| = 0.01. Entity B's magnitude of change = |0.66 - 0.65| = 0.01. Cross-Domain Association C's magnitude of change = |0.77 - 0.78| = 0.01. The maximum magnitude of change among all adopted entities is max(0.01, 0.01, 0.01) = 0.01, which is not lower than the preset stability threshold of 0.01 (here, "lower than" means strictly less than). Therefore, the condition is not met, and the next iteration should continue. If all changes in the next iteration decrease to below 0.008, the maximum value is below 0.01, the convergence condition is met, and the iteration terminates. It should be emphasized that the preset stability threshold should be set according to the specific application scenario's requirements for the stability accuracy of the graph. It is typically set to a value between 0.005 and 0.05, and this application does not impose specific restrictions on it.

[0181] In some of the embodiments described above in this application, a method is proposed to adjudicate, quantify, and map technical entities and cross-domain relationships into visual attributes. However, when generating an operational and interactive structured industrial technology map, how to encapsulate and assemble these discrete technical entities and relationships with rich attributes in a unified and standardized manner to form a complete map data structure for subsequent storage, display, and analysis is a problem that needs to be solved.

[0182] In response, this application further proposes the following steps for generating the structured industry technology map: Encapsulating each adopted technology entity after adjudication into a map node, with node attribute fields including at least entity identifier, technology name, circle category, quadrant category, post-calibration time series status, strategic value, and stage calibration confidence level. Encapsulating each adopted cross-domain related edge into a map edge, with edge attribute fields including at least head entity identifier, tail entity identifier, relationship type, quantified confidence level, and a list of related origins. Topologically assembling the map nodes and edges according to the radial layout position, the angular layout position, and the time series display level to generate the structured industry technology map.

[0183] For example, when generating this structured industry technology map, each technology entity that has been adjudicated and adopted is encapsulated as a map node. Here, "encapsulation" refers to integrating the various attribute information of a technology entity into a standardized data structure, making it an independent unit within the map. Each map node contains at least the following attribute fields: entity identifier, technology name, circle category, quadrant category, post-calibration time-series status, strategic value, and stage calibration confidence level. The entity identifier is the unique identifier of the technology entity. The technology name is the textual description of the technology entity. The circle category and quadrant category determine the radial and angular layout position of the technology entity in the map. The post-calibration time-series status indicates the evolutionary stage of the technology entity. The strategic value quantifies the importance of the technology entity. The stage calibration confidence level reflects the reliability of its time-series status determination. Through this encapsulation, all key information about the technology entities can be centrally managed and accessed.

[0184] Each cross-domain association edge that has been adjudicated and adopted is encapsulated as a graph edge. Similar to graph nodes, "encapsulating" a graph edge involves integrating the various attribute information of cross-domain associations into another standardized data structure, making it an independent unit connecting graph nodes. Each graph edge contains at least the following attribute fields: head entity identifier, tail entity identifier, relationship type, quantified confidence level, and a traceability list for the association. The head entity identifier and tail entity identifier point to the starting and ending technical entities of the association, respectively. The relationship type describes the specific nature of the association between the two technical entities. The quantified confidence level indicates the reliability of the association. The traceability list provides original evidence supporting the association, ensuring the traceability of the graph. This encapsulation allows for the clear representation and processing of association relationships and their metadata.

[0185] Building upon this foundation, the encapsulated graph nodes and edges are topologically assembled according to predetermined radial, angular, and temporal display hierarchies. The radial and angular layout positions collectively determine the spatial distribution of technical entities on the two-dimensional plane; for example, the concentric circle category might determine the distance of a node from the graph center, and the quadrant category might determine the node's position within a specific sector. The temporal display hierarchy can be used to hierarchically display technical entities over time; for instance, technical entities at different stages of evolution might be placed on different timeline hierarchies. Through this topological assembly, discrete nodes and edges are organized into a cohesive whole with a clear structure and layout, thereby generating a structured industrial technology graph.

[0186] The aforementioned technical solution encapsulates complexly processed and adjudicated technical entities and cross-domain relationships into unified and standardized graph nodes and edges. Based on their inherent attributes, these are then topologically assembled, solving the problem of integrating discrete information into an operable and interactive structured graph. This structured output not only facilitates the storage, transmission, and management of graph data but, more importantly, provides a solid data foundation for subsequent graph visualization, interactive exploration, and advanced analysis (such as path discovery and community detection). The rich attributes contained in the graph nodes and edges, such as strategic value, confidence level, and traceability lists, ensure that the generated industry technology graph is not merely a simple display of connections but a semantically rich, traceable, and quantifiable knowledge graph, enhancing its practical value and decision support capabilities.

[0187] In the process of structured extraction of entities and associations from multi-source heterogeneous industrial technology evidence based on large-scale models, directly processing unprocessed raw evidence may face challenges such as inconsistent data formats, information redundancy, and difficulties in efficiently locating the required evidence fragments in subsequent steps (such as entity identification, association discovery, and calculation of evolution stage calibration indicators). This not only increases the complexity and computational cost of large-scale language model processing but may also affect the accuracy of the extraction results and the efficiency of source tracing.

[0188] To address this, this application further proposes that, prior to the structured extraction of entities and associations from multi-source heterogeneous industrial technology evidence, the method further includes: standardizing evidence records from different data sources within the multi-source heterogeneous industrial technology evidence according to a preset unified evidence pattern, assigning each standardized evidence record an evidence identifier, the year of origin, and a source citation field. Based on the standardized evidence records, a forward index from the technology name to the evidence identifier list and an inverted index from the year to the evidence identifier list are constructed. During this structured extraction process, the forward index and the inverted index are used to quickly locate the corresponding evidence set, document annual output sequence, and patent annual output sequence for any candidate entity.

[0189] For example, in the case of multi-source heterogeneous industrial technology evidence from different data sources, evidence records are standardized according to a pre-defined unified evidence model, assigning each standardized evidence record an evidence identifier, the year of origin, and a source citation field. This step aims to preprocess original industrial technology evidence from different sources and in varying formats, ensuring it conforms to a unified data structure and standard. Standardization processing may include data cleaning, format conversion, and field mapping to ensure consistency across all evidence records in subsequent processing. For instance, data from different databases, document platforms, or patent offices can be unified into structured records containing fields such as "title," "abstract," "author / applicant," "publication / application date," "source institution," and "original link." Each standardized evidence record is assigned a unique evidence identifier, which can be used to track and cite the evidence throughout the processing flow. Extracting or parsing the year of origin from the evidence record is crucial for subsequent time series analysis and evolutionary stage calibration. Furthermore, binding source citation fields, such as the URL or storage path of the original document, ensures that all extracted information can be traced back to its original source.

[0190] It should be noted that the construction of the above-mentioned forward index and inverted index can be implemented in the following specific ways: The forward index adopts a hash table or inverted index file structure, with the technology name as the key and the list of evidence identifiers of all evidence records in which the technology name appears as the value, supporting approximate search with O(1) time complexity. The inverted index adopts an ordered mapping or B+ tree structure, with the year as the key and the list of all evidence identifiers in that year as the value, supporting range queries (such as querying all evidence records within a certain year interval). Both types of indexes can be built and persisted in the data preprocessing stage at once, and repeatedly used in the subsequent structured extraction and evolution calibration stages, thereby avoiding scanning the full amount of data each time and improving processing efficiency.

[0191] Based on this, and using the standardized evidence records, a forward index from technology name to a list of evidence identifiers, and an inverted index from year to a list of evidence identifiers are constructed. After the evidence records are standardized, to improve the efficiency of subsequent information retrieval, this application further constructs two types of indexes. A forward index is a data structure that maps a technology name (or keywords extracted from a technology name) to a list of evidence identifiers for all evidence records containing that technology name. For example, when searching for all evidence related to a specific technology, this index can be used for quick location. An inverted index is a data structure that maps a year to a list of evidence identifiers for all evidence records published or applied for in that year. This accelerates the data retrieval process when data needs to be aggregated by time, such as calculating annual output sequences for documents and patents. These indexes can be constructed using hash tables, B-trees, or other efficient data structures.

[0192] In this structured extraction process, the forward index and inverted index are used to quickly locate the corresponding evidence set, document annual output sequence, and patent annual output sequence for any candidate entity. When performing structured extraction of entities and associations, if the large language model identifies a potential technical entity or requires further analysis of a candidate entity, the pre-built forward index can be used to quickly retrieve all evidence records mentioning that entity based on its name. Combined with the inverted index, the distribution of these evidence records across different years can be efficiently analyzed, thereby quickly generating the document annual output sequence and patent annual output sequence corresponding to the candidate entity. This indexing mechanism significantly reduces the time required for full scanning or complex queries on large-scale, multi-source heterogeneous data, ensuring the efficiency and accuracy of structured extraction and subsequent evolutionary calibration stages.

[0193] The above technical solution standardizes the original evidence before performing structured extraction of entities and relationships from multi-source heterogeneous industrial technology evidence. Each record is assigned a unique evidence identifier, year, and source citation field. This solves the problems of inconsistent original data formats and difficulty in traceability, providing high-quality, standardized input for subsequent processing. Based on this, by constructing a forward index from technology name to the evidence identifier list and an inverted index from year to the evidence identifier list, the efficiency of quickly locating specific technology entity-related evidence sets and their annual output sequences within massive amounts of evidence is improved. This enables the large language model to obtain the necessary contextual information more accurately and efficiently when performing entity recognition and relationship discovery, and provides fast and reliable data support for the calculation of calibration indicators in subsequent evolution stages. This improves the processing efficiency, data accuracy, and traceability of the entire industrial technology map generation method.

[0194] Before performing structured extraction of entities and associations from multi-source heterogeneous industrial technology evidence, these evidence records need to be standardized and assigned evidence identifiers, year references, and source citation fields to facilitate subsequent indexing and rapid evidence location. However, efficiently, accurately, and traceably generating these key metadata records for massive amounts of heterogeneous evidence is a crucial prerequisite for ensuring the quality of subsequent graph construction and traceability. If the generated metadata is not standardized or accurate, it will directly affect the efficiency and accuracy of subsequent evidence index construction, thereby impacting the calibration of technological entity evolution stages and the calculation of association confidence.

[0195] To address this, this application further proposes a method for assigning an evidence identifier, a year, and a source citation field to each standardized evidence record. Specifically, this includes: combining the source identifier of the evidence record with the document number corresponding to the evidence record and performing a hash operation to generate the evidence identifier; extracting year information from the publication date of the document or the patent application date, and using the extracted year information as the year; setting the original document storage address or accessible link as the source citation field, and appending the source citation field to the corresponding standardized evidence record.

[0196] The evidence identifier is a unique identifier for each standardized evidence record, designed to provide a concise, stable, and traceable way to quickly locate and reference specific evidence records in subsequent processing. To generate this identifier, the source identifier to which the evidence record belongs is combined with the document number corresponding to the evidence record. The source identifier distinguishes the original source of the evidence record, such as a patent database or journal article repository, while the document number is a unique identifier for the document within that source. Combining the two, for example through string concatenation, ensures that even if different sources have the same document number, they can be distinguished by the source identifier, thus guaranteeing the uniqueness of the combined identifier. The combined information is then hashed to convert it into a fixed-length hash value with an extremely low collision probability, serving as the evidence identifier. This hashing process simplifies the storage and comparison of identifiers and improves the efficiency of index building and querying.

[0197] The year in question refers to the point in time when the technical content or event represented by the evidentiary record occurred. Its purpose is to provide a temporal basis for calibrating subsequent stages of technological evolution, such as for statistically analyzing annual literature output sequences and annual patent output sequences. This application extracts year information from the publication date of the document or the patent application date as the year in question. The publication date of the document typically refers to the date of public publication of journal articles, conference papers, etc., while the patent application date is the date the patent document was filed. This date information is usually clearly provided in the original evidentiary record, possessing high accuracy and authority, and is an important indicator reflecting the timeline of technological development.

[0198] The source citation field provides information about the access path or storage location of the original evidence record. Its purpose is to ensure that any technical entity or association in the evidence map can be traced back to its original evidence, thereby enhancing the credibility and transparency of the map. This application sets the original storage address or accessible link as the source citation field. The original storage address typically refers to the physical path of the evidence record in the local storage system, suitable for internal data sources. The accessible link refers to the URL of the evidence record on the Internet, suitable for publicly accessible online data sources. By providing this information, users can directly access the original evidence for content verification and in-depth analysis. These generated evidence identifiers, the year they belong, and the source citation field, as metadata, are attached to and integrated into the data structure of each normalized evidence record, ensuring that each evidence record carries complete traceability and time information, facilitating subsequent unified management and processing.

[0199] The above technical solutions enable efficient, accurate, and traceable generation of key metadata for massive amounts of multi-source heterogeneous industrial technology evidence records. For example, combining the source identifier with the document number and performing a hash operation generates a unique evidence identifier, resolving the identifier conflict problem caused by the same document number from different sources, ensuring the uniqueness and traceability of each evidence record. Extracting the year from the publication date or patent application date provides an accurate time dimension for calibration in subsequent technology entity evolution stages, making the statistics of annual document output sequences and annual patent output sequences more accurate. Setting the original document storage address or accessible link as a source citation field and attaching it to the evidence record enhances the credibility and transparency of all technology entities and associations in the graph, allowing users to easily trace back to the original evidence for verification, thereby improving the quality and usability of the entire industrial technology graph. This standardized metadata generation mechanism lays a solid foundation for the subsequent construction of efficient forward and inverted indexes, improves the efficiency of quickly locating the evidence set corresponding to candidate entities during structured extraction, and thus ensures the accuracy of subsequent evolution stage calibration and association confidence calculation.

[0200] In some embodiments described above, this application proposes a technical solution that uses stage calibration confidence as a weighting factor, and fuses it with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs in the cross-entity association set to generate a quantitative confidence score for cross-domain associations, and then binds it to an association tracing list. However, after generating the quantitative confidence score, if the binding operation of the association tracing list is not performed indiscriminately on all associated edges, it may lead to a large number of low-confidence associated edges also being traced and bound. This not only increases unnecessary computation and storage overhead, but may also introduce noise, affecting the quality and credibility of the industry technology graph.

[0201] To address this, this application further proposes that, after obtaining the quantified confidence level of the cross-domain association but before binding the association tracing list, the method further includes: setting a preset confidence threshold; removing cross-domain association edges with a quantified confidence level lower than the preset confidence threshold from the cross-entity association set; and performing the binding operation of the association tracing list only on cross-domain association edges with a quantified confidence level not lower than the preset confidence threshold, while retaining the corresponding cross-domain association edges in the cross-entity association set.

[0202] For example, setting a pre-defined confidence threshold refers to pre-determining a numerical value as a standard for quantifying the confidence level of cross-domain associations. This threshold can be set based on various factors. For instance, it can be determined based on historical data analysis to establish an empirical value that effectively distinguishes between high-quality and low-quality associations. It can also be manually set using the knowledge of domain experts, or dynamically adjusted using statistical methods, such as percentiles based on confidence distributions. The purpose of this threshold is to initially filter out associations with low confidence, potential noise, or unimportant implications before subsequent resource-intensive source tracing and binding operations, thereby improving the efficiency of subsequent processing and the quality of the graph.

[0203] After setting a threshold, cross-domain association edges with a quantified confidence level below the preset threshold are removed from the cross-entity association set. This step aims to perform preliminary quality control and data simplification. After calculating the quantified confidence level of each cross-domain association edge, the system iterates through all association edges and compares their quantified confidence levels with the preset confidence threshold. Any association edge with a quantified confidence level below the threshold is considered an unreliable or unimportant association and is removed from the current cross-entity association set. This removal operation can be a logical flag deletion or a physical data cleanup to ensure that subsequent processing only targets high-quality associations. For example, a boolean flag can be maintained to mark associations that do not meet the threshold as "to be deleted," and these marked associations can be skipped in subsequent processing. Alternatively, these association records can be directly deleted from memory or the database.

[0204] Based on this, only cross-domain association edges with a quantified confidence level not lower than the preset confidence threshold are bound to the association tracing list, and the corresponding cross-domain association edges are retained in the cross-entity association set. After the above screening, only those cross-domain association edges with a quantified confidence level reaching or exceeding the preset confidence threshold are retained and further bound to the association tracing list. This means that for these associations deemed sufficiently credible, the system will aggregate all source tracing annotations supporting their formation and the corresponding original text fragments to form an association tracing list, and bind it to the corresponding association edge. This ensures the effectiveness and relevance of tracing information, avoids investing unnecessary resources in low-quality associations, and thus optimizes the overall processing flow.

[0205] Through the aforementioned technical solution, after generating the quantified confidence level of cross-domain associations, this application introduces a screening mechanism based on a pre-set confidence threshold. This mechanism can effectively identify and remove cross-domain association edges with low quantified confidence levels, thereby avoiding unnecessary computational and storage overhead caused by binding a link traceability list to these low-quality associations. By performing link traceability binding only on high-confidence association edges, the accuracy and effectiveness of traceability information are ensured, making the generated industry technology map more refined and reliable. This pre-screening mechanism improves the efficiency of the entire map generation process, reduces noise interference, and enhances the credibility and practical value of the map, allowing analysts to focus more on high-quality, evidence-supported industry technology associations.

[0206] In some of the above implementations, a method was proposed to map the quantified confidence level of each cross-domain related edge adopted after adjudication to the edge rendering intensity, and to store the related traceability list bound to each cross-domain related edge as a traceable label associated with the corresponding edge, in order to generate a structured industry technology graph. However, in practical applications, if all adopted related edges are rendered directly, even if their quantified confidence level is low, it may lead to an overly complex and redundant graph interface, making it difficult for users to quickly identify key, high-confidence technology connections, thereby affecting the readability and analysis efficiency of the graph.

[0207] To address this, this application further proposes a technical solution for optimizing the visualization of the industry technology graph, specifically including: setting an edge display threshold; when the quantized confidence level of a cross-domain related edge is lower than the edge display threshold, hiding the corresponding edge in the generated industry technology graph; when the quantized confidence level of a cross-domain related edge is not lower than the edge display threshold, converting the quantized confidence level into the corresponding edge rendering intensity according to a preset mapping relationship between the quantized confidence level and the edge rendering intensity; and driving the visualization of the corresponding edge in the industry technology graph according to the edge rendering intensity, wherein the value of the edge rendering intensity is positively correlated with the quantized confidence level.

[0208] For example, the edge display threshold refers to a pre-defined value as the minimum quantitative confidence level required to determine whether cross-domain related edges should be displayed in the industry technology map. This threshold serves to initially screen related edges, filtering out those with low confidence levels that may not be of much value to the map analysis, thus preventing the map from appearing cluttered due to excessive information. This threshold can be set based on the needs of the actual application scenario, through expert experience or statistical analysis. For example, it can be set to 0.3, 0.5, etc., indicating that only related edges with a quantitative confidence level reaching or exceeding this value will be considered for display. Furthermore, this threshold can also be designed as a configurable parameter, allowing users to dynamically adjust it according to their desired confidence level to achieve map views of different granularities.

[0209] When the quantified confidence level of a cross-domain association edge falls below its display threshold, the corresponding edge is hidden in the generated industry technology map. This step is a specific application of the edge display threshold, designed to directly remove or make invisible association edges that fail to meet the preset confidence standard. When the quantified confidence level of a cross-domain association edge fails to reach the preset edge display threshold, the system will not draw it in the industry technology map. The hiding operation can be implemented in several ways, such as skipping the drawing instruction for that edge during the map rendering stage, or setting its transparency to completely transparent, making it visually invisible. In this way, the map can automatically eliminate associations deemed unreliable or insufficiently important, thereby improving the clarity of the map and the prominence of core information.

[0210] For related edges filtered by an edge display threshold—that is, when the quantized confidence of a cross-domain related edge is not lower than the edge display threshold—its quantized confidence will be further converted into a specific visual rendering attribute, namely, the edge rendering intensity. This needs to be done according to a preset mapping relationship between the quantized confidence and the edge rendering intensity. The preset mapping relationship can be linear, non-linear (such as logarithmic or exponential), piecewise, or stepwise. For example, a linear function can be defined to map the quantized confidence range of 0.5 to 1.0 to a line width range of 1 pixel to 5 pixels. Or a color gradient mapping can be defined to map the confidence from low to high to a color range from light to dark. This transformation allows the confidence of related edges to be intuitively expressed through visual elements (such as line width, color saturation, transparency, etc.).

[0211] The rendering intensity of an edge drives its visualization within the industry technology map, and this rendering intensity is positively correlated with the quantified confidence level. This is the process of applying the calculated edge rendering intensity to the map visualization. The system adjusts the visual representation of corresponding edges in the map based on the converted rendering intensity. For example, if edge rendering intensity represents line width, edges with higher confidence levels will be displayed as thicker lines. If it represents color saturation, edges with higher confidence levels will have more vibrant colors. This positive correlation ensures that edges with higher quantified confidence levels are visually more prominent and eye-catching, guiding users to prioritize these important technological connections and improving the map's information delivery efficiency and user experience.

[0212] Through the above technical solution, this application solves the problem of information redundancy and visual clutter in the generation of industry technology maps due to an excessive number of low-confidence related edges. By setting an edge display threshold, edges with low confidence and little analytical value can be intelligently filtered out, thereby significantly reducing the visual burden of the map and making it more concise and clear. For related edges that pass the threshold filtering, their quantified confidence is mapped to edge rendering intensity, ensuring a positive correlation between the two. This visually strengthens high-confidence related edges, for example, by using thicker lines or darker colors, thus intuitively highlighting the core technology connections and critical paths in the map. This not only improves the readability and aesthetics of the map, but more importantly, it guides users to quickly focus on important technological development trends and potential technological cooperation opportunities, improving the analytical efficiency and decision support capabilities of the industry technology map.

[0213] In some of the embodiments described above in this application, although visualization attributes such as the radial layout position, angular layout position, temporal display level, node display scale, and edge rendering intensity of the graph have been proposed to map technical entities and cross-domain associations to the graph, in the actual process of generating structured industrial technology graphs, if there is a lack of standardized data encapsulation and attribute definition for the core components of the graph—namely, technical entities and associated edges—it may lead to incomplete graph information expression, difficulty in subsequent query and analysis, and affect the traceability and credibility of the graph.

[0214] To this end, this application further proposes the following steps for generating the structured industry technology map: Encapsulating each adopted technology entity as a map node, with node attribute fields including at least entity identifier, technology name, circle category, quadrant category, post-calibration time series status, strategic value, and stage calibration confidence level. Encapsulating each adopted cross-domain related edge as a map edge, with edge attribute fields including at least head entity identifier, tail entity identifier, relationship type, quantified confidence level, and the related traceability list. Assembling the map nodes and edges according to the radial layout position, the angular layout position, and the time series display level to generate the structured industry technology map.

[0215] For example, encapsulating each adopted technology entity as a graph node means converting each adjudicated and adopted technology entity into an independent graph node object with a specific data structure. This encapsulation process ensures that each technology entity has a clear and operable representation in the graph, facilitating subsequent storage, retrieval, and visualization. The node attribute fields include at least the entity identifier, technology name, circle category, quadrant category, post-calibration time-series status, strategic value, and stage calibration confidence level. The entity identifier is a unique identifier assigned to each graph node, typically a string or number, used to uniquely identify the technology entity in the graph, ensuring its accurate referencing and management. The technology name is the textual name of the technology entity, providing key information for users to intuitively understand and identify it. The circle category refers to the classification of the technology entity in the radial layout of the graph, such as basic technology, core technology, and cutting-edge technology, indicating its macro-level positioning within the industrial technology system. Quadrant categories refer to the classification of a technology entity in the angular layout of the industry map. For example, it can be divided into different quadrants under dimensions such as technology maturity and market potential, used to further refine its position in specific dimensions. The calibrated time series status refers to the time series development stage determined after the technology entity has undergone evolutionary stage calibration. For example, it can be divided into the nascent stage, growth stage, maturity stage, and decline stage, reflecting the current life cycle stage of the technology entity. Strategic value refers to the assessment of the importance and influence of the technology entity in industry development. Its value can comprehensively consider factors such as technological influence and market prospects, used to measure the potential contribution of the technology entity to the industry. Stage calibration confidence level refers to the reliability of the judgment on the time series status of the technology entity, reflecting the accuracy of the time series status assessment result; a higher value indicates a more reliable judgment.

[0216] Encapsulating each adopted cross-domain association edge into a graph edge refers to converting each adjudicated and adopted cross-domain association into an independent graph edge object with a specific data structure. This encapsulation process allows the associations in the graph to be clearly represented and managed, serving as a bridge connecting technical entities. The edge attribute fields at least include a head entity identifier, a tail entity identifier, a relationship type, a quantified confidence level, and a traceability list for the association. The head entity identifier and tail entity identifier are unique identifiers of the two technical entities connected by the association edge, clearly defining the start and end points of the association. The relationship type refers to the specific semantics of the relationship represented by the association edge, such as "contains," "applies to," or "depends on," used to describe the specific nature of the connection between the two technical entities. The quantified confidence level is a quantitative assessment of the reliability of the association; a higher value indicates a more credible association. The traceability list for the association is a collection of identifiers and original text fragments supporting all original evidence for the association, providing traceability and facilitating user verification.

[0217] Based on this, the graph nodes and edges are topologically assembled according to the radial layout position, angular layout position, and temporal display level to generate the structured industry technology graph. Topological assembly refers to connecting the graph nodes and edges according to their definitions to form a complete graph structure. This process is not simply about putting nodes and edges into a set, but about establishing connections between them to ensure the connectivity and logical integrity of the graph. Radial layout position, angular layout position, and temporal display level are key parameters guiding the layout of graph nodes in two-dimensional or three-dimensional space. The radial layout position determines the distance of the node from the graph center point, the angular layout position determines the angle of the node on the circumference, and the temporal display level may determine the depth or level of the node in the time dimension. Through these parameters, the graph can be displayed in an orderly and intuitive manner. The product is a graph data structure with clearly defined nodes, edges, and their attributes. This structure not only contains technical entities and relationships but also carries rich metadata, such as classification, temporal sequence, value, and confidence level, providing a solid data foundation for subsequent visualization, analysis, and application.

[0218] Through the aforementioned technical solution, adjudicated and adopted technical entities and cross-domain relationships can be encapsulated into standardized graph nodes and edges. This encapsulation not only clarifies the core attributes of each node and edge, such as entity identifier, technology name, relationship type, and quantitative confidence level, but more importantly, it systematically integrates rich metadata obtained in previous steps, including circle categories, quadrant categories, calibrated time-series status, strategic value, stage calibration confidence level, and related traceability lists. By topologically assembling these graph nodes and edges with rich attributes, this application generates a highly structured, information-complete, and traceable industry technology graph. This improves the usability, analyzability, and credibility of the graph, enabling users to perform efficient querying, filtering, analysis, and visualization operations based on a clearly defined graph structure, thereby gaining a deeper understanding of the development trajectory and internal connections of industry technologies and providing strong support for decision-making.

[0219] In the generation of industry technology maps, iterative processing mechanisms are typically employed to ensure accuracy and timeliness. However, completing the evolution stage labels for technology entities initially identified as emerging technologies by joint adjudication presents challenges. Because emerging technologies may lack sufficient evidence accumulation in the early stages of iteration, the calculation of their evolution stage calibration indicators may be incomplete or inaccurate, thus affecting the precise determination of their temporal status.

[0220] To address this, this application further proposes that, after the joint ruling is completed, the method also includes internal iterative processing. Specifically, this involves reverting to the calculation process of the evolution stage calibration index and supplementing the evolution stage label for technical entities deemed emerging technologies by the joint ruling. This includes: for technical entities adopted after the joint ruling and deemed emerging technologies, re-tracing all source traceability annotations bound to that technical entity. For source traceability annotations discovered during the tracing process that were not included in the previous iteration, re-aggregating document counts and patent counts by year, and updating the corresponding annual document output sequence and annual patent output sequence. Based on the updated annual document output sequence and annual patent output sequence, the evolution stage calibration index is recalculated, and the evolution stage label is supplemented for the technical entity based on the recalculated evolution stage calibration index.

[0221] For example, for a technology entity that has been adopted and determined to be an emerging technology after a joint ruling, the system will re-traverse all its associated source tracing annotations. This step aims to ensure that all evidence related to the emerging technology entity, whether discovered in the initial data processing stage or in subsequent iterations, is fully considered. The re-traversal process accesses stored evidence identifiers and original text fragments to obtain the most complete set of evidence.

[0222] During the re-traversal, the system identifies source traceability annotations that were not included in the previous iteration. These uncounted annotations may represent newly generated documents or patents during the iteration, or newly discovered old evidence due to data updates, expanded search scope, or other reasons. For these newly discovered source traceability annotations, the system re-aggregates document counts and patent counts according to their respective years, and updates the annual document output sequence and annual patent output sequence corresponding to the technology entity. For example, if a new document published in 2023 is found, the document count for 2023 will increase accordingly, thus making the annual output sequence more accurately reflect the latest developments of the technology entity.

[0223] Based on the updated annual document output sequence and annual patent output sequence, the system will recalculate the evolution stage calibration indicators for this emerging technology entity. This recalculation process will utilize the latest output sequence data to reassess key indicators such as the annual document output growth rate, the annual patent output growth rate, and citation acceleration. Based on the recalculated evolution stage calibration indicators, the system will complete or correct the evolution stage label for this technology entity to ensure the accuracy of its temporal status.

[0224] The aforementioned technical solution addresses the issue of inaccurate or incomplete evolution stage labels for emerging technology entities in the early stages of iteration due to insufficient evidence accumulation. By conducting targeted evidence re-examination and dynamically updating the annual output sequence for emerging technology entities, the calculation of evolution stage calibration indicators is ensured to always be based on the latest and most comprehensive data. This enables the system to more accurately capture the development dynamics of emerging technologies and assign them more realistic evolution stage labels, thereby improving the accuracy and timeliness of the industry technology map in identifying and tracking cutting-edge technologies, and providing decision-makers with more reliable insights into technological development trends.

[0225] In some of the embodiments described above in this application, it is proposed to determine whether the evolution trend of each technical entity is growth-oriented or slowing-down-decline-oriented based on the changing trends of the annual output sequence of literature and the annual output sequence of patents corresponding to each technical entity relative to their respective base period windows. However, relying solely on the general description of "changing trend" may lead to ambiguity or subjectivity in the determination of the evolution trend category of technical entities in actual operation, thereby affecting the accuracy of subsequent time series status determination.

[0226] In response, this application further proposes a method for determining whether the evolution trend of each technical entity is growth-oriented or slowing-downward-declining, based on the changing trends of the annual literature output sequence and annual patent output sequence corresponding to each technical entity relative to their respective base period windows. This method includes: For the annual output sequence of literature, if the average annual output of literature in the recent output window is not lower than the average annual output of literature in the base period window, the literature dimension is judged to be growing; otherwise, the literature dimension is judged to be declining.

[0227] For the annual patent output sequence, if the average annual patent output in the recent output window is not lower than the average annual patent output in the base period window, the patent dimension is judged to be growing; otherwise, the patent dimension is judged to be declining.

[0228] When both the document dimension and the patent dimension show growth, the evolution trend of the technological entity is classified as growth-oriented. When at least one of the document dimension and the patent dimension shows decline, the evolution trend of the technological entity is classified as slowing decline-oriented.

[0229] For example, when determining the literature dimension, the system analyzes the annual literature output sequence for each technological entity. This sequence reflects the entity's literature output in different years. By comparing the average annual literature output within a recent output window (e.g., the last three years) with the average annual literature output within a base period output window (e.g., the three years prior), the system can quantitatively assess changes in the technological entity's activity in academic research. If the recent average annual output is not lower than the base period average annual output, it indicates that the technological entity's research attention or output in the academic community is showing an upward or stable trend, thus its literature dimension is classified as "growth." Conversely, if the recent average annual output is lower than the base period average annual output, it indicates that the technological entity's activity in academic research has declined, and its literature dimension is classified as "decline." This comparison based on quantitative indicators provides an objective basis for subsequent determination of evolution trend categories.

[0230] Similarly, when determining the patent dimension, the system analyzes the annual patent output sequence of the technology entity. This sequence reflects the patent output of the technology entity in different years. By comparing the average annual patent output within the recent output window with the average annual patent output within the base period output window, the changes in the activity of the technology entity in industrial application and technological innovation can be quantitatively assessed. If the recent average annual patent output is not lower than the base period average annual patent output, it indicates that the technology entity's input and output in technological innovation and industrialization are showing an increasing or stable trend, and its patent dimension is judged as "growth". Conversely, if the recent average annual output is lower than the base period average annual output, it indicates that the technology entity's activity in patent layout and technology transfer has weakened, and its patent dimension is judged as "decline". The combination of literature and patent dimensions can comprehensively reflect the life cycle status of the technology entity from different perspectives.

[0231] After obtaining the results for the literature and patent dimensions, the system further determines the evolution trend category of the technological entity. When both the literature and patent dimensions are determined to be "growth," it indicates that the technological entity is showing strong growth in both academic research and industrial application, and its evolution trend category is explicitly determined to be "growth-oriented." This dual growth condition ensures the rigor of the "growth-oriented" determination, avoiding misjudgments based solely on growth in a single dimension, thus more accurately reflecting the overall development momentum of the technological entity. Conversely, if at least one of the literature and patent dimensions is determined to be "decline," it indicates that the development momentum of the technological entity is no longer comprehensive growth, and it may have entered a maturity, slowdown, or decline phase, at which point its evolution trend category is determined to be "slowdown and decline-oriented." This "one-vote veto" determination mechanism can promptly identify technological entities whose development momentum is weakening, providing more accurate stage information for subsequent map generation and analysis.

[0232] Through the aforementioned technical solution, this application provides a clear and quantitative standard for determining the evolution trend category of technical entities. By examining the annual output sequence of literature and patents respectively, comparing their average annual output in the recent period with the base period window, the development trend of technical entities in the two core dimensions of academic research and industrial application can be objectively assessed. This dual-dimensional, quantitative comparison method effectively avoids the arbitrariness and inconsistency that may result from judging solely based on vague "trends," thus improving the accuracy and reliability of evolution trend category determination. When both the literature and patent dimensions show growth, it is clearly identified as a growth-type entity, ensuring accurate identification of technical entities truly in an upward phase. Conversely, if either dimension shows decline, it is identified as a slowing-down decline-type entity, enabling timely detection of signals of weakening momentum in the development of technical entities. This clear determination logic allows for more accurate and reliable input when subsequently determining the temporal status of each entity based on evolution stage calibration indicators, thereby improving the scientific rigor and practicality of the entire industrial technology map generation method.

[0233] The following example will provide a more detailed explanation of the above technical solution: To more intuitively demonstrate the effectiveness of the technical solution in this application, specific numerical examples of the key quantitative indicators in the above examples are provided below. In the above example of "new energy battery technology", for the technical entity "solid-state battery", assuming that its annual literature output sequence covers data from 2015 to 2025, the average annual literature output in the recent window (2023-2025) is 45 articles, and the average annual literature output in the base window (2018-2020) is 28 articles, then the annual literature output growth rate is (45-28) / 28, which is approximately 0.607. In the annual patent output sequence, the average annual patent output in the recent window is 12, and the average annual patent output in the base window is 6, then the annual patent output growth rate is (12-6) / 6 = 1.0. In the cumulative citation sequence, the cumulative citations for the current year (2025) are 520, the cumulative citations for the previous year (2024) are 380, and the cumulative citations for the year before that (2023) are 290. Therefore, the first difference is 520 - 380 = 140, the second difference is 380 - 290 = 90, and the citation acceleration is 140 - 90 = 50. Assuming that the authority score of "solid-state batteries" ranks in the 85th percentile among all technology entities (i.e., higher than 85% of technology entities), the normalized score is 0.85. The four sub-indicators (annual literature output growth rate of 0.607 normalized to 0.61, annual patent output growth rate of 1.0 normalized to 0.95, citation acceleration of 50 normalized to 0.72, and percentile ranking of 0.85) are weighted and summed according to the preset sub-indicator weights (e.g., 0.25 each) to obtain the evolution stage calibration index = 0.25×0.61+0.25×0.95+0.25×0.72+0.25×0.85=0.7825. For the cross-domain association "solid-state battery - silicon-carbon anode", assuming the source traceability annotation sets of the two entities are {src_1, src_2, src_3, src_5} and {src_1, src_3, src_4, src_6} respectively, the intersection is {src_1, src_3} (2 entities), and the union is {src_1, src_2, src_3, src_4, src_5, src_6} (6 entities). The evidence co-occurrence strength (cross-union ratio) is 2 / 6, approximately 0.333. Using a pre-trained semantic encoding model, the cosine similarity between the semantic feature representations of the entity names at both ends is calculated to be 0.78. The number of independent source supports is the number of deduplicated source identifiers, assumed to be 4. The evidence co-occurrence strength (0.333), semantic affinity (0.78), and independent source support number (4, assumed to be 0.85 after log normalization) are weighted and summed according to preset fusion weights (e.g., 0.3, 0.3, 0.4) to obtain the basic confidence score = 0.3 × 0.333 + 0.3 × 0.78 + 0.4 × 0.85 = 0.674.Then, using the stage calibration confidence level of "solid-state battery" (assumed to be 0.85) as a weighting factor, a multiplicative weighting is performed to obtain the quantitative confidence level of the cross-domain association = 0.674 × 0.85, approximately equal to 0.573. This value indicates that the association between "solid-state battery and silicon-carbon anode" has a moderately high confidence level, and its inclusion in the map will be determined based on subsequent admissibility decision steps. The above values ​​are for illustrative purposes only; in actual applications, the parameters and weights can be adjusted according to the data characteristics of the specific domain.

[0234] Suppose a research institution needs to construct an industry technology map in the field of "new energy battery technology" to support its technology strategic planning. This institution possesses a large amount of multi-source, heterogeneous industry technology evidence, including research papers, patent documents, industry reports, project documents, etc.

[0235] Before performing structured extraction of entities and associations, the system standardizes these multi-source, heterogeneous industrial technology evidences. For example, literature and patent records from different databases are formatted uniformly, and each record is assigned a unique evidence identifier (generated through hashing), the year (extracted from the publication or application date), and a source citation field (such as the original link). The system constructs forward and inverted indexes to quickly locate the evidence set corresponding to a specific technology name and its annual output sequence.

[0236] The system initiates a structured extraction process. It inputs carefully crafted prompts into a large language model. These prompts define the large language model as an "industry technology analyst," then specify format constraints, requiring the model to categorize the extracted "technology names" as "layer category labels" (e.g., "core technology," "supporting technology") and "quadrant category labels" (e.g., "basic research," "application development"), and output "relationships" as triples of "head entity identifier," "tail entity identifier," and "relationship type." Crucially, the prompts also embed binding instructions, forcing the large language model to simultaneously refer back to the "evidence identifier" corresponding to the original text block that generated the output when outputting each technology entity name or relationship, and to extract the "original text fragment" supporting the output. These "evidence identifiers" and "original text fragments" together constitute the "binding field."

[0237] After the large language model processes the evidence, the system parses its output to extract "entity records" and "related records." For each record, the system rigorously verifies whether it carries a valid "binding field," that is, checks whether the "evidence identifier" can be traced back to the original evidence. Any record without a valid "binding field" is discarded, effectively suppressing the "illusionary" information without factual basis that the large language model might generate. Through this step, the system obtains a preliminary "technical entity set" and "cross-entity association set," where each entity and association is bound to a clear "source traceability label." This solves the problem of large model output lacking objective evidence binding and traceability information in related technologies.

[0238] The system performs evolution stage calibration for each technical entity (e.g., "solid-state battery") in the "technical entity set". It aggregates document counts, patent counts, and cumulative citation counts by year from all "source traceability annotations" linked to "solid-state battery", forming an "annual document output sequence", a "annual patent output sequence", and a "cumulative citation sequence". Based on these sequences, the system calculates the "annual document output growth rate", the "annual patent output growth rate", and the "citation acceleration". For example, it calculates the annual document output growth rate by comparing the average annual document output of the most recent three years with the base period three years (e.g., three to six years ago). The system also extracts authority scores (such as journal impact factor and patent citation counts) from the "source traceability annotations" and determines their percentile ranking among all technical entities. These indicators are normalized and weighted, then summed to obtain the "evolution stage calibration indicators".

[0239] The system determines the "time series state" and "stage calibration confidence level" of "solid-state batteries" based on the "evolution stage calibration index." It analyzes the changing trends of the "annual literature output sequence" and "annual patent output sequence" relative to the base period window to determine whether the "evolution trend category" of "solid-state batteries" is "growth-oriented" or "gradual decline-oriented." By comparing the "evolution stage calibration index" with a preset threshold and combining it with the "evolution trend category," the system maps "solid-state batteries" to discrete evolution stages such as "emergence stage" and "growth stage," which are then used as their "time series state." To further improve the time series confidence level, the system independently calls a large language model to output an "evolution stage label" based on the semantic content of the evidence. If the label output by the large language model matches the "time series state" obtained based on the objective index mapping, the "stage calibration confidence level" will be incremented by an upward floating increment on the preset base confidence level. This increment increases with the number of "source tracing annotations." If inconsistencies are found, the system still adopts the "time-series state" based on objective indicators, but applies a consistency discount factor to reduce the "stage calibration confidence" and marks the "solid-state battery" as a "state to be strengthened for verification," so that evidence can be supplemented for verification in subsequent iterations. This process solves the problem of lack of objective time-series evidence binding and cross-validation for evolution stage labels in related technologies, and improves the calibrability and reliability of the time-series dimension of the graph.

[0240] After determining the confidence level of the evolution stage of the technical entities, the system begins to process the "cross-domain technical entity pairs" (e.g., the association between "solid-state battery" and "silicon-carbon anode") in the "cross-entity association set". It calculates the "cross-intersection over union ratio" (CUIR) at the evidentiary level, based on the "source traceability annotations" bound to each of the two entities, as the "evidence co-occurrence strength". It then uses a pre-trained semantic encoding model to calculate the "semantic affinity" (e.g., cosine similarity) of the entity names at both ends. Furthermore, the system counts the number of deduplicated "source identifiers" bound to association records involving both entities, as the "number of independent source supports". The "evidence co-occurrence strength", "semantic affinity", and "number of independent source supports" are normalized and weighted to obtain a "baseline confidence level". The system uses the "stage calibration confidence levels" of "solid-state battery" and "silicon-carbon anode" as weighting factors to perform a multiplicative weighting on the "baseline confidence level", generating the "quantified confidence level" of the cross-domain association. Association edges with a quantified confidence level below a preset threshold are removed. The system compiles all "source traceability annotations" and corresponding "original text fragments" supporting the association, forming a "relationship traceability list," which is then bound to the cross-domain association. This solves the problem in related technologies where the confidence score of cross-domain associations is not linked to the credibility of source evidence and entity evolution stages, and lacks traceable information. This allows the associations in the graph to be quantified and accepted, and their evidentiary basis to be verified afterward.

[0241] The system makes "admissibility" decisions for each object in the "technical entity set" and the "cross-entity association set." For each candidate entity or association, the system generates a "source document support signal" (i.e., the number of different sources), a "stage-calibrated confidence-weighted evidence co-occurrence signal" (e.g., the quantitative confidence of the association), and an "authoritative frontier dictionary hit signal" (through fuzzy matching of entity names with pre-set authoritative dictionaries). The decision-making process sets a "source document support hard threshold," meaning that an object is only allowed to proceed to the subsequent weighted summation and threshold comparison if the "source document support signal" indicates the existence of at least one source piece of evidence; otherwise, it is directly rejected. This hard threshold effectively prevents "illusory nodes" without factual support from entering the graph. For objects that pass the hard threshold, the system weights and sums the three signals according to preset weights to obtain a "comprehensive confidence score," and compares it with the "admissibility threshold." If the "comprehensive confidence score" is not lower than the "admissibility threshold," the object is deemed admissible. It is worth noting that if an adoptable object does not match the "authoritative frontier dictionary", it will be marked as an "emerging technology", thus solving the problem that the hard constraints of the authoritative dictionary in related technologies cannot cover emerging technologies.

[0242] Following the joint ruling, the system performs entity normalization. It uses a semantic encoding model to calculate the semantic similarity between adopted entity names. When the similarity between two entities reaches a preset threshold, a connection edge is established in the entity similarity network, and the network is divided into communities. Entities assigned to the same community are merged into a single "canonical entity," with priority given to entity names within the community that match authoritative leading-edge dictionaries as the "canonical entity" name. Cross-domain connections with the original entity endpoints before merging are remapped according to the "canonical entity," and duplicate connections pointing to the same "canonical entity pair" after remapping are aggregated, selecting the highest quantized confidence level and merging the source list.

[0243] The entire process is not completed in one go, but involves internal iterative processing. The adopted technical entities, along with their cross-domain associations, serve as input for the next iteration, flowing back into the calculation of evolution stage calibration indicators. Specifically, for technical entities identified as "emerging technologies," their evolution stage labels are supplemented. The system checks whether the current iteration meets preset convergence conditions, such as no new adopted entities for two consecutive iterations, and the change in adoption scores for each adopted entity being below a preset stability threshold. If the conditions are met, the iteration terminates; otherwise, the next round continues. This iterative mechanism ensures the continuous optimization and stability of the graph.

[0244] The system generates a structured industry technology graph. Adopted technology entities are positioned radially according to their "layer category," angularly according to their "quadrant category," and their temporal display level according to their calibrated "temporal status." Each technology entity's "strategic value" (obtained by a weighted fusion of its technological influence reference value and the average quantified confidence score of its adjacent edges) is mapped to the node display scale. The "quantified confidence score" of adopted cross-domain related edges is mapped to "edge rendering intensity," and its associated "related traceability list" is stored as a traceable label associated with the corresponding edge. All this information is encapsulated into graph nodes and edges, and topologically assembled according to layout rules to form an intuitive, reliable, traceable industry technology graph that reflects the laws of technological evolution.

[0245] All of the above-mentioned optional technical solutions can be combined in any way to form the optional embodiments of this application, and will not be described in detail here.

[0246] The above are merely optional embodiments of this application and are not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for generating an industry technology map based on a large model, characterized in that, The method includes: The entity and association structures of multi-source heterogeneous industrial technology evidence are extracted and bound with source tracing labels to obtain a set of technology entities and a set of cross-entity associations; Based on the annual output sequence of documents and the annual output sequence of patents in the evidence fragments bound to each entity in the set of technical entities, the evolution stage calibration index is calculated, and the temporal state and stage calibration confidence of each entity are determined according to the evolution stage calibration index. The stage calibration confidence is used as a weighting factor and fused with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs in the cross-entity association set to generate a quantitative confidence of cross-domain association, and then bound to the association tracing list. The evidence co-occurrence signal weighted by the stage calibration confidence level, the source document support signal composed of the associated tracing list, and the hit signal of the authoritative frontier dictionary are used to jointly determine the admissibility of each object in the set of technical entities and the set of cross-entity associations, and generate a structured industry technology map under the premise of requiring source document support.

2. The method according to claim 1, characterized in that, The process involves structured extraction of entities and associations from multi-source heterogeneous industrial technology evidence, along with binding source tracing annotations, to obtain a set of technical entities and a set of cross-entity associations, including: Construct a prompt message carrying entity category constraints and output format specifications, and input the prompt message and the multi-source heterogeneous industrial technology evidence into a large language model to drive the large language model to perform entity recognition and relationship discovery; Entity records and associated records are parsed from the output of the large language model. For each entity record and each associated record, it is verified whether they carry a binding field. The binding field is used to record the evidence identifier and original text fragment supporting the corresponding record. The entity records carrying the binding field are grouped into the technical entity set, and the associated records carrying the binding field are grouped into the cross-entity association set. Records that do not carry the binding field are discarded.

3. The method according to claim 2, characterized in that, The construction includes prompts carrying entity category constraints and output format specifications. These prompts, along with the multi-source heterogeneous industrial technology evidence, are input into a large language model to drive the model to perform entity recognition and association discovery, including: Input role setting prompts into the large language model to constrain the large language model to interpret the multi-source heterogeneous industrial technology evidence in the role of an industry technology analyst; After the role setting prompt, a format constraint prompt is input to the large language model, requiring the large language model to classify and label the extracted technical names according to the circle category label and quadrant category label, and output the association relationship in the format of a triple of head entity identifier, tail entity identifier and relationship type; The binding instruction is embedded in the format constraint prompt, requiring the large language model to simultaneously extract the evidence identifier and the original text fragment on which it is based for each output field, so as to form the binding field.

4. The method according to claim 1, characterized in that, The calculation of evolution stage calibration indicators based on the annual document output sequence and annual patent output sequence in the evidence fragments bound to each entity in the technology entity set includes: For each technical entity in the set of technical entities, the document count, patent count, and cumulative citation count are aggregated by year from all source traceability annotations bound to the corresponding technical entity, forming the document annual output sequence, the patent annual output sequence, and the cumulative citation sequence, respectively. The annual output growth rate of literature is calculated based on the comparison of output between the recent window and the base window in the annual output sequence of literature; the annual output growth rate of patents is calculated based on the comparison of output between the recent window and the base window in the annual output sequence of patents; and the citation acceleration is calculated based on the cumulative citation sequence. The authoritative score is extracted from the source tracing annotation and the percentile ranking of the authoritative score among all technical entities is determined. The annual output growth rate of the literature, the annual output growth rate of the patent, the citation acceleration and the percentile ranking are normalized respectively and then weighted and summed according to the preset sub-index weights. The weighted summation result is used as the calibration index of the evolution stage.

5. The method according to claim 1, characterized in that, The determination of the temporal state and stage calibration confidence of each entity based on the evolution stage calibration index includes: Based on the changing trends of the annual output sequence of literature and the annual output sequence of patents corresponding to each technical entity relative to their respective base period windows, the evolution trend of each technical entity is determined to be either growth-oriented or slowing-downward-declining. The evolution stage calibration index is compared with a preset first threshold, and combined with the evolution trend category, each technical entity is mapped to a discrete evolution stage in the full life cycle timeline, and the mapping result is used as the time sequence state. In the structured extraction process, the large language model is independently invoked, and evolution stage labels are output based on the semantic content of the multi-source heterogeneous industrial technology evidence. The temporal state obtained by jointly mapping the evolution stage calibration index and the evolution trend category is compared with the evolution stage labels. When the comparison result is consistent, the stage calibration confidence is obtained by adding an upward increment to a preset base confidence. The upward increment increases monotonically with the increase of the number of source traceability labels and does not exceed a preset upper limit. When the comparison result is inconsistent, the temporal state is adopted as the final determined evolution stage, and a consistency discount factor is applied to the preset base confidence to reduce the stage calibration confidence. At the same time, the technical entity is marked as a state to be strengthened for verification. The consistency discount factor is configured so that the stage calibration confidence calculated in the case of inconsistency is not lower than a preset minimum confidence, and the preset minimum confidence is higher than the fallback confidence given to entities without any source traceability labels.

6. The method according to claim 1, characterized in that, The step of using the stage calibration confidence as a weighting factor, and fusing it with the evidence co-occurrence strength and semantic affinity of cross-domain technical entity pairs in the cross-entity association set, generates a quantitative confidence score for cross-domain association, and binds it to the association tracing list, including: For cross-domain technical entity pairs in the cross-entity association set, the source tracing labels bound to each of the two technical entities are used as statistical basis to calculate the cross-union ratio of the two technical entities at the evidence level, and the cross-union ratio is used as the evidence co-occurrence strength. Using a pre-trained semantic coding model, word segmentation, lexical encoding, and context pooling are performed on the name texts of the technical entities at both ends to obtain the semantic feature representations of the technical entities at both ends. The cosine similarity between the two semantic feature representations is calculated as the semantic affinity. After normalizing the evidence co-occurrence strength and the semantic affinity, the basic confidence level is obtained by weighting and summing them with the number of independent source support according to a preset fusion weight. Then, the basic confidence level is multiplicatively weighted by the stage calibration confidence level, and the multiplicative weighted result is used as the quantitative confidence level of the cross-domain association. The number of independent source support is the number of deduplicated source identifiers in the source tracing annotations bound to the association records that involve both technical entities at the same time. All source traceability annotations and corresponding original text fragments that support the cross-domain association are compiled to form the association traceability list and bound to the cross-domain association.

7. The method according to claim 1, characterized in that, The evidence co-occurrence signal weighted by the stage-calibrated confidence level, the source document support signal constituted by the associated tracing list, and the authoritative frontier dictionary hit signal jointly determine the admissibility of each object in the technical entity set and the cross-entity association set, including: For candidate associations in the cross-entity association set, the source document support signal is generated based on the number of different sources in the association traceability list bound to the candidate association; for candidate entities in the technical entity set, the source document support signal is generated based on the number of different sources in the source traceability annotation bound to the candidate entity. Multiply the co-occurrence strength of the evidence corresponding to the candidate entity or the candidate association by the stage calibration confidence level, and use the multiplication result as the evidence co-occurrence signal weighted by the stage calibration confidence level; load the preset authoritative frontier dictionary, perform fuzzy matching on the name of the candidate entity, and generate the authoritative frontier dictionary hit signal based on the matching result; The source document support signal, the evidence co-occurrence signal after stage calibration confidence weighting, and the authoritative frontier dictionary hit signal are weighted and summed according to preset adjudication weights. The weighted summation result is compared with a preset adoption threshold to determine the admissibility of the corresponding object.

8. The method according to claim 1, characterized in that, The generation of a structured industry technology map, provided that source documentation is required, includes: The radial layout positions of each technical entity that is adopted after the ruling will be determined according to its respective circle category, the angular layout positions will be determined according to its respective quadrant category, and the timing display levels will be determined according to its respective calibrated timing status. The strategic value of each technology entity adopted after adjudication is mapped to the node display scale. The strategic value is obtained by weighted fusion of the technology entity's technical influence reference value and the average quantitative confidence of the adjacent edges of the technology entity. The technical influence reference value is the percentile ranking normalized value of the authoritative score of the source traceability annotation bound to the technology entity among all adopted technology entities. The average quantitative confidence of the adjacent edges is the arithmetic mean of the quantitative confidence of all adjacent edges with the technology entity as the endpoint. The quantified confidence of each cross-domain related edge that has been adopted after adjudication is mapped to the edge rendering intensity, and the related traceability list bound to each cross-domain related edge is stored as a traceable label associated with the corresponding edge to generate the structured industry technology map.

9. The method according to claim 1, characterized in that, The method also includes entity normalization processing performed after the joint ruling: The names of each entity in the technical entity set are segmented and encoded using a pre-trained semantic coding model to obtain the semantic coding representation of each entity, and the similarity of the semantic coding representations between pairs of entities is calculated. When the similarity between two entities is not lower than a preset similarity threshold, a connection edge is established between the two entities in the entity similarity network. Community division based on modularity optimization is performed on the entity similarity network. Entities assigned to the same community are merged into a single normalized entity. The name of the entity that matches the authoritative frontier dictionary within the community is selected as the name of the normalized entity. Before merging, cross-domain related edges with the original entity as the endpoint are remapped according to the specified entity, and duplicate related edges pointing to the same specified entity pair after remapping are aggregated.

10. The method according to claim 1, characterized in that, After the joint ruling is completed, the method further includes: The technology entities adopted after the joint adjudication are associated with cross-domain relationships and used as input for the next iteration. This process is then fed back into the calculation of the evolution stage calibration index to complete the evolution stage labels for technology entities that were determined to be emerging technologies by the joint adjudication. The current iteration is checked to see if it meets the preset convergence conditions. The preset convergence conditions include that there are no new adopted technical entities or cross-domain associations in two consecutive iterations, and the change in the adoption score generated by each adopted object in the joint adjudication step is lower than the preset stability threshold between two adjacent iterations. When the preset convergence condition is met, the iteration terminates and the industry technology map generated in the current round is output. When the preset convergence condition is not met, the next round of iteration continues.

Images