A technology development situation awareness system and method

Abstract

The present invention discloses a technology development situation awareness system and method, which belongs to the field of information processing. The system includes a data input module, a field overview module, an intelligence task question and answer module, and a data output module. The method includes: step one: receiving user questions and performing deep semantic analysis; step two: building a thinking chain and prompt word engineering system based on metadata; step three: executing thinking chain reasoning driven by a large language model and generating a draft report; step four: executing metadata consistency check of the draft report, performing report content update, presenting the research report and evidence; step five: archiving high-quality scientific data in the research report. The present invention can effectively mine the key characteristic laws in technology development, realize comprehensive, convenient, efficient and intelligent perception of technology development trends, provide automated reasoning of technology situation awareness results, and provide efficient and reliable intelligence reference for decision-making.

Description

Technical Field

[0001] The present invention relates to the field of information processing, and more specifically, to a technology development situation awareness system and method. Background Art

[0002] How to grasp the trends of technological development with high quality is a key technical issue of concern to frontline scientific research teams and research management institutions. Current mainstream technology analysis systems face three major bottlenecks: First, insufficient data integration capabilities. While technology carriers include literature data such as patents and papers, patent analysis platforms like DerwentInnovation lack the ability to integrate and analyze paper data; second, the analysis dimension is limited. Existing tools often use simple statistical methods (such as patent application trend analysis) that fail to reveal the underlying patterns of technological evolution; and third, insufficient intelligence. Tools like VOSviewer require manual interpretation of analysis results and lack the ability for automated artificial intelligence reasoning.

[0003] Existing technologies only analyze general technology indicators, lacking a rich set of analytical dimensions to deeply explain the underlying laws of technological evolution, and also lack intelligent automated reasoning capabilities. Therefore, there is an urgent need to develop comprehensive, convenient, efficient, and intelligent solutions for sensing technological development trends. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology and provide a technology development situation awareness system and method, which can effectively explore the key characteristic laws in technology development, realize comprehensive, convenient, efficient and intelligent perception of technology development trends, provide automated reasoning of technology situation awareness results, and provide efficient and reliable intelligence reference for decision-making.

[0005] The object of the present invention is achieved through the following solutions:

[0006] A technology development situation awareness system includes: a data input module, a domain overview module, an intelligence task question and answer module, and a data output module;

[0007] The data input module is used to import technical document carriers and search formulas to obtain domain data sets;

[0008] The domain overview module is used to identify the technical subject of the input data set, obtain descriptive statistical analysis results, and implement the domain overview function based on the descriptive statistical analysis results;

[0009] The intelligence task question-and-answer module is used to integrate existing technology analysis results, build a thinking chain for interpreting the development trend of field technology, and use a large language model combined with the thinking chain to perform technology supply chain deduction, technology evolution path question-and-answer, technology opportunity identification and risk point question-and-answer, technology transformation law prediction, and technology integration trend analysis.

[0010] The data output module is used to output the results of the intelligence task question and answer module.

[0011] Furthermore, the technical document carrier includes a patent dataset and a paper dataset.

[0012] Furthermore, the technical subject identification of the input data set specifically includes: using a name disambiguation method to identify the subject of the country, institution, and secondary institution.

[0013] Furthermore, the descriptive statistical analysis results specifically include: country technology distribution, institution technology distribution, technology category distribution and technology life cycle curve.

[0014] A method for sensing technology development trends, based on the technology development trends sensing system as described in any one of the above items, comprises the following steps:

[0015] Step 1: Receive user questions and perform deep semantic analysis;

[0016] Step 2: Build a metadata-based thinking chain and prompt word engineering system;

[0017] Step 3: Execute the thought chain reasoning driven by the large language model and generate the first draft of the report;

[0018] Step 4: Perform metadata consistency checks on the draft report, update the report content, and present the research report and evidence;

[0019] Step 5: Archive high-quality scientific data in your research report.

[0020] Furthermore, in step 1, receiving the user question and performing deep semantic analysis specifically includes the following sub-steps:

[0021] S1. Obtain user input from the user requirements interface: Utilize the interactive interface to obtain the user's input, representing their technology development situation awareness needs (Q_{raw}). This Q_{raw} is captured and used as the topic of the technology development situation awareness research report and the starting point for subsequent processing.

[0022] S2. Preprocessing and normalization of the raw query: After receiving Q_{raw}, perform preprocessing operations to obtain the normalized representation Q_{norm} of the technology development situation awareness requirements;

[0023] S3. Perform intent analysis on Q_{norm}: Convert Q_{norm} into a vector representation using a text embedding model. This vector representation is fed into a pre-trained intent classification model, which learns and outputs the pre-defined intelligence task category to which the query belongs.

[0024] S4. Extract key parameters related to intent: Extract key parameters based on the identified intelligence task categories and standardized representations;

[0025] S5. Extracting explicit metadata requirements in queries: Leveraging a large language model to identify and link entities from Q_{norm} to the knowledge base, combined with key parameters, to identify explicit metadata requirements.

[0026] Furthermore, in step 2, the construction of the metadata-based thought chain and prompt word engineering system specifically includes the following sub-steps:

[0027] S6. Retrieving metadata related to the query intent from the metadata knowledge base: Based on the structured query object parsed in step 1, query the metadata database that stores the structured analysis results; retrieve a subset of metadata related to the entities and key parameters in the structured query object to form a candidate metadata set;

[0028] S7. Filter metadata that matches user explicit / implicit requirements: Based on the structured query object and association rules, infer explicit and implicit metadata requirements from the intelligence task, and filter the candidate metadata sets to obtain metadata sets that are relevant to the requirements.

[0029] S8. Select or generate a thought chain template for the selected task: Maintain a library of intelligence task thought chain templates, where each template corresponds to a selected intelligence task category and defines logical reasoning steps to guide the large language model in structured thinking. Based on the intelligence task category in the structured query object, select the most matching thought chain template. If there is no template in the library that fully matches the current query intent, adaptive adjustment is performed based on the template with the highest semantic similarity.

[0030] S9. Construct metadata-enhanced prompt word sequence: For the jth node of the selected thought chain, construct a set prompt word content for it, denoted as Prompt_j. Prompt_j is generated by combining the following five elements (a), (b), (c), (d), and (e):

[0031] (a) Based on the specific instructions of node j, guide the large language model to perform the analysis task;

[0032] (b) The subtask defined by node j that requires a large language model to solve;

[0033] (c) The core conclusion summary output by the large language model of the previous j-1 nodes, which serves as the context of the current node step;

[0034] (d) A subset of metadata directly related to the analysis task of the current node j, extracted from the metadata set filtered in step S7. These metadata are embedded in the prompt words in a structured form and serve as known facts and core evidence for the large language model reasoning;

[0035] (e) a description or requirement of the expected output format of the large language model at this node;

[0036] S10. Construct a complete input sequence for the large language model: organize the Prompt_j generated by each node in the instantiated thinking chain in sequence to form an ordered, complete input sequence ready to be submitted to the large language model for step-by-step reasoning.

[0037] Furthermore, in step 3, the execution of thought chain reasoning and generation of a draft report driven by a large language model specifically includes the following sub-steps:

[0038] S11. Initialize Large Language Model Dialog: Select and initialize a large language model, pre-set inference parameters, load the model on the server, and verify its normal operation through standard test cases.

[0039] S12. Reasoning about the First Node of the Thought Chain: First, the first node of the thought chain is processed, prompt content is constructed, and input into the initialized large language model. Based on the instructions and sub-questions in prompt Prompt_1, as well as the provided metadata context, the model generates a preliminary analysis or answer to the first node of the thought chain.

[0040] S13. Saving the first node output and extracting context: The system saves the output content of the first node; at the same time, the output content is processed to obtain dynamic context information for subsequent node reasoning;

[0041] S14. Step-by-step reasoning in thought chains: For subsequent nodes in a thought chain, the corresponding components are first updated, integrating the dynamic context information output by all previous nodes. This updated prompt, which includes the complete historical context and the metadata required for the current node, is then input into the large language model to obtain the model's output for the current node.

[0042] S15. Loop through step S14 until all thought chain nodes are processed: follow the thought chain logic path, and the output of each step is generated based on the conclusions of the previous steps in the thought chain and the metadata context reasoning of the current node, thereby forming a draft report.

[0043] Furthermore, in step 4, the metadata consistency check of the draft report is carried out, the content of the report is updated, and the research report and evidence are presented, which specifically includes the following sub-steps:

[0044] S16. Calculate and store the confidence level of each node in the thought chain: As the large language model generates the output content for each node, an internal consistency check is performed to determine whether there are factual inconsistencies between the output content and the metadata in the prompt word. A confidence level is then obtained to assess the output quality. If the confidence level is found to be below a set threshold, the node and its output are marked as low confidence.

[0045] S17. Metadata alignment and output retry for low-confidence nodes: A maximum number of retries is preset. For nodes marked as low-confidence and their output content, output corrections are attempted. The prompt word and low-confidence output content are re-input into the large language model to obtain the updated output content, which is then compared again with the metadata in the prompt word for factuality, until the confidence level exceeds the threshold or the maximum number of cycles is reached.

[0046] S18. Preliminary summary of the structured research report: Organize and summarize the output of all nodes in the thinking chain in a logical order to form a research report that includes the reasoning process for the original user needs;

[0047] S19. Fact-checking and evidence link visualization: Citation links are established for metadata cited in research reports. In the final presentation, users can click to view the source of evidence. Key numerical or factual assertions are re-verified with the original data source.

[0048] S20. Calculate the overall confidence of the research report: Aggregate the confidence of the output content of each node and calculate the overall confidence score of the research report;

[0049] S21. Research Report Presentation and Confidence Grading Strategies: Different presentation strategies are employed based on the confidence score. High-confidence answers are presented directly; medium-confidence answers are accompanied by uncertainty explanations or prompts to the user regarding key assumptions; and low-confidence answers are recommended to the user for consultation with other information sources or experts.

[0050] S22. Language Style Fine-tuning: Based on a locally stored intelligence research report knowledge base, a large language model is used to fine-tune the scientific and technological intelligence language style of the research report, ensuring that its expression meets the standard, objective, and rigorous requirements of professional intelligence analysis reports while also improving readability.

[0051] S23. Merge data analysis images and text into a document: Based on a predetermined layout template, integrate the optimized research report text, images corresponding to each analysis module metadata, metadata evidence links, and comprehensive confidence levels to form a complete final document research report, which is then output to the specified path location. Key metadata and images are structured and output to the specified path location.

[0052] Furthermore, in step 5, the high-quality scientific data in the archiving research report specifically includes the following sub-steps:

[0053] S24. Positive Feedback Collection and Knowledge Base Archiving: High-confidence research reports that have received positive user feedback, along with their corresponding standardized representations, the instantiated thought processes used, and key metadata, will be structured and stored on the system's server platform for continuous learning and knowledge accumulation to enhance the quality of the platform's automated analysis.

[0054] S25. Negative Feedback Collection: When metadata entries frequently cause large language models to generate low-confidence conclusions and negative user feedback, the metadata entries will be flagged to prompt platform developers and data managers to review or update them.

[0055] S26. Dynamic maintenance and expansion of the thought chain template library: Based on analysis of actual reasoning paths in high-frequency, high-quality Q&A, optimize existing thought chain templates and incorporate them into the technical intelligence task thought chain template library, enabling dynamic maintenance and expansion of the template library.

[0056] S27. User Preference Learning: Over the long term, the system collects information about the query habits, technical areas of interest, and preferences for answer detail from selected users or user groups. These preferences are then used as a guide to adjust subsequent query parsing, thought chain selection, and answer presentation.

[0057] S28. Similar Question Retrieval and Utilization of Cached Answers: When a new user query is received, after fully parsing in Step 1, the updated knowledge base is first queried and the semantic similarity with historical research reports is calculated. If a historical question-answer pair with similarity and answer quality above a set threshold is found, the cached answer is recommended to the user or quickly adjusted based on the cached answer.

[0058] S29. Scheduling mechanism for concurrent user queries: Design a request queue and resource scheduling mechanism to handle query requests initiated by multiple users simultaneously.

[0059] S30. Interaction with API interfaces of other modules of the platform: Ensure that there is a data interface between the Q&A module and the data input module and the domain overview module to obtain the latest analysis results as metadata.

[0060] The beneficial effects of the present invention include:

[0061] This invention improves the accuracy, controllability and consistency with the overall cognition of the platform of large language models in professional question-answering tasks, thereby effectively alleviating the problem that general large models are prone to "hallucinations" or irrelevant outputs due to the lack of deep domain knowledge in scenario-based applications.

[0062] This invention combines instantiated thought chains with metadata knowledge base retrieval and enhanced generation depth. Based on metadata matching, it constructs a clear, followable logical reasoning framework and confidence evaluation indicators for intelligence analysis tasks. This enables large language models to conduct methodical and progressively deeper analysis, rather than divergent and unpredictable text generation. Each step is based on the analysis of verifiable data, rather than the divergent, unpredictable, black-box text generation used in the past. This better meets the requirements of technological development situational awareness for rigorous analysis processes and accurate conclusions.

[0063] The present invention not only records the textual reasoning steps, but more importantly, it records the key metadata entries that each step of the reasoning relies on. Ultimately, this reasoning path (textual steps, associated metadata content, and confidence) can be visualized for users, significantly improving the data traceability of technology development situation awareness and greatly enhancing the interpretability and credibility of the final technology development situation research report. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0065] Figure 1 This is a schematic diagram of the overall framework of the technology development situation awareness system proposed in the present invention;

[0066] Figure 2 Flowchart of the steps of the method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] All features disclosed in all embodiments in this specification, or steps in all methods or processes implicitly disclosed, except for mutually exclusive features and / or steps, can be combined and / or expanded or replaced in any manner.

[0068] This invention aims to address the difficulty existing systems meet the demands of modern technological development, such as real-time monitoring and intelligent early warning, for situational awareness and management. In particular, it is unable to cope with the challenges of rapid iteration in emerging technologies. The inventors believe that existing solutions suffer from systemic flaws in three areas: data processing architecture, algorithmic models, and interaction methods.

[0069] Specifically, as a first aspect of the present invention, a technology development situation awareness system is provided, including a data input module, a field overview module, an intelligence task question and answer module and a data output module.

[0070] (1) Data input module, which imports technical literature carriers such as patent data sets and paper data sets and their search formulas to realize field data collection.

[0071] More specifically, the data input module is the basic component of the technology development situation awareness system of the present invention. It is mainly responsible for collecting various types of technical literature data and providing raw data support for the subsequent analysis of the platform. It specifically implements the following functions: the data input module supports the access of multi-source heterogeneous data, and is compatible with local file uploads and online database retrieval of formatted technical literature data. Users can select local files through the graphical interface, and the system will automatically identify the file type and parse it. For online database retrieval, the module supports docking with common patent databases, academic paper libraries, etc. Users can set search formulas in the system to collect technical literature data in specific fields.

[0072] To improve the efficiency and accuracy of data collection, the data input module also provides search query configuration. Users can flexibly customize search queries based on different technical fields and research needs. Search queries can be combined based on keywords, classification codes, applicants, inventors, and other criteria. Based on the user-configured search query, the system will periodically or in real time retrieve the latest technical literature data from the specified data source.

[0073] In terms of data preprocessing, the data input module performs preliminary parsing and format conversion on the collected data. It performs operations such as encoding conversion and character cleaning to ensure text accuracy and consistency. For structured data files such as CSV and Excel, the system parses the table structure, extracts the data content, and converts it into a unified data format for processing by subsequent modules.

[0074] The data input module can be implemented using multi-threading technology and a distributed architecture to improve data collection and processing efficiency. Multi-threading technology enables the module to simultaneously handle the access and data preprocessing tasks of multiple data sources, significantly reducing data preparation time. A distributed architecture allows the module to collect and process data across multiple servers or computing nodes, further improving system scalability and performance.

[0075] In summary, the data input module is an indispensable component of the present invention's technology development situational awareness system. By supporting the access of multi-source heterogeneous data, providing flexible search-based configuration capabilities, and efficient preprocessing mechanisms, it ensures the platform's ability to acquire comprehensive and accurate technical literature data, laying a solid foundation for subsequent analysis and mining.

[0076] (2) The field overview module identifies the technical subjects of the input data set, including name disambiguation of countries, institutions, and secondary institutions, and realizes descriptive statistical analysis of country technology distribution, institution technology distribution, technology category distribution, technology life cycle curve, etc., thereby realizing the field overview function.

[0077] More specifically, the domain overview module is an important component of the technology development situational awareness system. Its main function is to identify technology entities and conduct descriptive statistical analysis on the input data set, providing users with a macro overview of the technology field. The following is the detailed implementation of this module:

[0078] The Domain Overview module first receives technical literature data from the Data Input Module. After preliminary preprocessing, this data is stored in a unified format in the module's temporary database. One of the module's core functions is to identify and disambiguate technology entities, including countries, institutions, and secondary organizations. To achieve this, the Domain Overview module utilizes a machine learning-based disambiguation algorithm. This algorithm comprehensively considers information such as the name, address, and R&D field of the technology entity, constructing a feature vector to represent each technology entity. By calculating similarity and matching metrics, it accurately distinguishes different technology entities.

[0079] Based on the identification of technology entities, the field overview module further generates various statistical analysis results. These include country-specific technology distribution maps, institutional technology distribution maps, technology category distribution maps, and technology lifecycle curves. The country-specific technology distribution map displays indicators such as the number of patent applications and papers published in specific technology fields in different countries, reflecting each country's technological strength and R&D investment in these fields. The institutional technology distribution map focuses on specific institutions, such as enterprises, universities, and research institutes, and shows their distribution and competitive situation in the technology field. The technology category distribution map classifies and compiles technical literature based on the International Patent Classification (IPC) or other technology classification systems, revealing the development status and trends of different technology categories. The technology lifecycle curve analyzes the temporal distribution of technical literature to depict the entire process of technology from inception, growth, maturity, to decline, providing users with the stage characteristics of technological development.

[0080] The Domain Overview module utilizes a variety of data visualization techniques and interactive operations. Data visualization presents complex statistical analysis results in intuitive charts, such as bar charts, line charts, and pie charts. By exporting image metadata as prompts for a large language model, explanatory text is output, enabling users to quickly understand the underlying meaning of the data. Interactive operations allow users to obtain and view statistical information of varying dimensions and granularity by setting various parameters, enhancing the user experience and analytical flexibility.

[0081] The Domain Overview module supports basic customized report generation. Based on user settings, the system regularly retrieves the latest technical literature from the Data Input Module and automatically updates statistical analysis results. Users can select different statistical indicators and chart types to generate personalized domain overview reports based on their needs. These reports can be exported in formats such as Word, PDF, and Excel for easy sharing and archiving.

[0082] In summary, the Domain Overview module provides users with a comprehensive and intuitive macro view of the technology landscape through advanced technology subject identification algorithms and rich statistical analysis capabilities. It helps users quickly understand the overall landscape of the technology landscape, laying the foundation for subsequent in-depth analysis and decision support.

[0083] (3) The intelligence task question-and-answer module integrates existing technical analysis results to build a thinking chain for interpreting the development trend of field technology. It uses a large language model combined with thinking chain technology to realize functions such as technology supply chain deduction, technology evolution path question-and-answer, technology opportunity identification and risk point question-and-answer, and technology integration trend analysis.

[0084] More specifically, the intelligence task question-and-answer module is a key component of the technology development situation awareness system. It aims to achieve intelligent interpretation of the technology development situation in a field through the deep integration of existing technology analysis results. This module utilizes a large language model combined with thought chain technology to construct a multi-level technology reasoning system that supports multi-dimensional intelligence deduction capabilities. These dimensions include the aforementioned technology supply chain, technology evolution path analysis results, identification of technology opportunities and potential risk points, construction of a technology value assessment system, and prediction of technology integration trends. The resulting technology deduction results provide targeted research and judgment results, help grasp the future direction of technology development, and provide efficient and reliable intelligence support for decision-making.

[0085] In actual operation, the intelligence task question-answering module receives user questions and then goes through a series of steps, including question analysis, thought chain construction, large language model processing, question-answer generation, and statistical verification of the question-answer results, ultimately providing higher-quality technology situation question-answering services. The results of this question-answer generation can provide systematic research reports supporting technology forecasting, innovation decision-making, and risk management.

[0086] (4) Data output module, which integrates the process data, intermediate results and final results of the aforementioned data analysis, and automatically generates word, pdf and other versions of the research report on the field technology development situation awareness, as well as metadata of key analysis results.

[0087] More specifically, the data output module is responsible for systematically integrating the process data, intermediate results, and final results of the aforementioned data analysis, and automatically generating standardized research reports. This module supports output in multiple formats, including Word and PDF, ensuring the readability and portability of research reports. Figure 2 The system platform's homepage interface is displayed. The module also extracts metadata from key analysis results, forming structured data sets that facilitate subsequent data mining and reuse. Through automated report generation, the module significantly improves data analysis efficiency and reduces the time and cost of manual operations. Furthermore, the module's design fully considers actual user needs, supporting flexible customization of report content and format, ensuring that output results meet the application requirements of diverse scenarios and providing reliable, evidence-based support for comprehensive awareness of technological development trends and efficient decision-making.

[0088] like Figure 1 As shown, based on the above system embodiment, a technical law mining module and a technical value analysis module can be further integrated and set up. Both modules can be implemented using existing technologies. Here, a simple example is given below:

[0089] The Technical Pattern Mining module uses technical content to identify named entities, their classifications, and relationships, enabling the identification of topic and entity evolution paths. More specifically, its primary function is to uncover the inherent connections and development patterns between technologies through in-depth analysis of technical content. The following is a detailed implementation of this module: The Technical Pattern Mining module receives preprocessed data from the Data Input Module and the Technical Subject Identification Module. This data contains detailed information about technical documents and related information about the technical subject. The module first uses a large language model to deeply analyze the content of the technical documents and extract key information, including algorithms such as named entity recognition, classification, and relationship identification. These entities include technical terms, inventors, organizations, device names, and technical subject names. By identifying these entities, the module accurately extracts core information from the technical documents, providing a foundation for further analysis and mining. Based on named entity identification, the module further performs entity classification and relationship identification. The entity classification function categorizes the identified entities into predefined categories, such as technical fields, technical methods, and technical applications. The relationship identification function analyzes the relationships between entities, such as whether technology A cites technology B or whether technology C and technology D share semantic similarities in technical effects. The module uses large language models and thought chain technology to perform semantic understanding and relationship extraction on technical documents. These models are capable of capturing complex semantic information within technical documents and accurately identifying relationships between entities. Based on the results of named entity recognition, classification, and relationship identification, existing technologies can be used to construct topic and entity evolution paths. For example, the constructed topic evolution paths can demonstrate the development and evolution of technical topics over time, revealing the origins, development, and possible future directions of technology. The constructed entity evolution paths can track the emergence and changes of specific technical entities at different points in time, reflecting the life cycle of technical entities and their technological inheritance relationships. The module analyzes and optimizes the constructed evolution paths using a main path analysis algorithm, extracting representative and valuable technology evolution paths. The technology pattern mining module provides a variety of visualization and interactive features to help users better understand and utilize the discovered technical patterns. The module visualizes the mined entities, relationships, and evolution paths in the form of a knowledge graph. The knowledge graph, through its node and edge structure, intuitively presents the complex relationships and evolution paths between technical entities. The module also supports filtering, sorting, and clustering of mining results, helping users quickly locate technical information of interest and gain a deeper understanding of the inherent connections between technologies. In summary, the Technology Pattern Mining module mines key information and inherent connections in technical literature, providing a rich set of technology evolution paths and knowledge graphs. This reveals the inherent laws of technological development and provides strong support for technology value evaluation and technology forecasting.

[0090] The Technology Value Analysis module leverages existing technologies to explore and construct a technology value analysis system encompassing technological innovation, disruptiveness, stability, readiness, technology chain leadership, and industrial chain leadership. This enables high-value technology forecasting, technology convergence forecasting, technology transformation forecasting, and technology risk early warning. More specifically, the Technology Value Analysis module is primarily used to construct a situational awareness technology value assessment system. Its core lies in its multi-dimensional assessment system, encompassing key dimensions such as technological innovation, disruptiveness, stability, readiness, technology chain leadership, and industrial chain leadership. Each dimension is equipped with a series of specific evaluation indicators and calculation methods to ensure the scientific and accurate evaluation results. The following is a specific implementation of this module: Technological innovation measures the uniqueness and novelty of a technology. Using the semantic embedding vectors of technical texts from a large model, the semantic cosine similarity between the target technology and existing technologies can be calculated in terms of function, structure, and methodology. For patented technologies, the breadth of technology coverage is assessed by the number of IPC categories, while for technologies in papers, the breadth of technology coverage is assessed by the number of disciplines covered. This comprehensive assessment results in a comprehensive evaluation of technological innovation. The assessment of technological disruptiveness focuses on the potential impact of a technology on the existing market and its ability to transform the landscape. The module assesses a technology by analyzing its market penetration, its substitution effect on traditional technologies, and its impact on industry standards. For example, the system collects and analyzes market application cases, market share changes, and industry expert opinions. Using market analysis and expert evaluation models, the system predicts the disruptive impact of a technology on the market. The technology disruption assessment aims to quantify the potential impact of a technology on the existing market and technology landscape. The module analyzes data such as the number of countries in the technology family, the scope of claim coverage, and the degree of mutation of technical keywords. The evaluation formula includes the ratio of the number of countries covered by the family to the number of major global market countries, the growth rate of citations, and the claim expansion rate (the ratio of the current number of claims to the number of basic patents in the family) to comprehensively determine the technology disruption assessment. The stability assessment examines the reliability and consistency of the technology under different environments and conditions. This is achieved by analyzing multiple dimensions, including historical performance data, patent maintenance years, annual fee payment status, and invalidation declarations. The system develops a technology stability assessment model that comprehensively considers the performance fluctuations of the technology over time to derive a comprehensive stability score. The readiness assessment considers the maturity of the technology from laboratory research and development to practical application. This module assesses a technology's development stage, test results, degree of standardization, and readiness for commercial application based on available data such as the technology's legal status, application date, and standardization association identifier. A technology readiness assessment model is established to determine the technology's readiness level. The technology chain leadership assessment focuses on the technology's core position and leading role in the overall technological system.This module quantifies a technology's hubness and cross-domain connectivity within the technology chain by analyzing patent and paper citation networks and technology keyword co-occurrence matrix data. The module uses the PageRank algorithm to calculate a technology's centrality within the citation network and assesses its cross-domain connectivity by combining citation shares of non-related technology and subject categories. The industry chain leadership assessment focuses on the technology's influence and leadership within the industry chain. The module quantifies the technology's driving effect on upstream and downstream industries by analyzing technology applicant type, national economic industry labels, industry chain link keywords, and licensing filing data. Based on the analysis results of these technology value indicators, a technology forecasting function can be further developed. The high-value technology forecasting function aims to identify potentially high-value technology directions through multi-dimensional data analysis. This function is based on data columns (including patent citation frequency, IPC classification number, number of claims, independent claim ratio, and IPC coverage breadth) and combines multiple assessment dimensions such as technological innovation, disruptiveness, and market adaptability. For patent technologies, we use the five high-value patent screening methods provided by the State Intellectual Property Office (SIPO) (invention patents in strategic emerging industries, invention patents with similar patent families overseas, invention patents with a tenure exceeding 10 years, invention patents with high pledge financing amounts, and invention patents that have won National Science and Technology Awards or China Patent Awards) to identify current high-value technology seed nodes. For paper technologies, we use Price's Law to select highly cited papers as valuable technology seed nodes. We then use the Analytic Hierarchy Process (AHP) to determine the weights of each dimension, and finally combine it with the Random Forest algorithm for comprehensive scoring to identify potential high-value technologies. The technology convergence prediction function analyzes the correlations and synergies between technologies to identify areas and directions likely to lead to technological convergence. This function utilizes data from the co-occurrence matrix of technology keywords, technology citation networks, patent technology classification numbers, and the co-occurrence of technology discipline classifications in papers to construct a technology convergence potential assessment model. The technology transformation prediction function monitors the risks and obstacles that may arise during the transition of technologies from the R&D stage to practical application. This function generates a technology transformation risk index based on data such as the legal status of patent technologies, tenure, annual fee payment status, and number of invalidation declarations. The technology risk warning function is used to identify potential risks during technology development. This function constructs a technology risk assessment model based on the number of countries in the technology family, the scope of claim coverage, citation frequency mutation data, technology life cycle curves, and invalidation records. In addition, the module also provides a variety of visualization and interactive functions to help users better understand and utilize the assessment results. The system presents the results of technology value assessment in the form of intuitive charts, such as radar charts, heat maps, bar charts, etc., enabling fast and efficient multi-dimensional analysis, prediction, and visualization of the technology value system. It supports the display of detailed assessment results and the viewing of technology performance and changing trends in different dimensions.In summary, the technology value analysis module realizes multi-dimensional evaluation and prediction of the potential value of technology by constructing a scientific multi-dimensional evaluation system and applicable machine learning algorithms, provides comprehensive technology value evaluation results and intuitive visualization, and provides strong support for technology investment, R&D decision-making and strategic planning.

[0091] As a second aspect of the present invention, a method for sensing technology development situation is provided, such as Figure 2 The figure shows a more comprehensive implementation of the key technical module in the system of the present invention, namely the intelligence task question-answering module. As the core interactive and intelligent analysis component of the technology development situational awareness system platform, the intelligence task question-answering module of the system of the present invention is unique in that it deeply integrates the structured analysis results of the platform's previous modules as high-quality metadata, driving a large language model for controllable and reliable reasoning and question-answering for specific intelligence tasks. The module includes the following steps:

[0092] Step 1: Receive user questions and perform deep semantic analysis. Specifically, step 1 includes the following sub-steps:

[0093] S1. User Requirement Interface Activation and Input Capture: The system can be used to activate the user interface for intelligence task question-answering, allowing users to enter a raw expression (Q_{raw}) of their technology development situation awareness needs, or to use a customized technology development situation question template. For example, "Please analyze the development trends and key influencing factors of technology field T1 over the next three years." The system then captures this Q_{raw} and uses it as the topic of the technology development situation awareness research report. This raw string also serves as the starting point for subsequent processing.

[0094] S2. Raw query preprocessing and normalization: After receiving the raw representation of the technology development situation awareness requirement, Q_{raw}, the system performs a series of preprocessing operations to improve the accuracy of subsequent parsing. These operations include removing leading and trailing whitespace, converting full-width characters to half-width characters, performing case normalization (e.g., converting to lowercase or performing intelligent case conversion based on a domain dictionary), and verifying that the character encoding conforms to the platform's unified standard (e.g., UTF-8). This results in the normalized representation of the technology development situation awareness requirement, Q_{norm}.

[0095] S3. Perform intent analysis on the normalized representation Q_{norm}: The normalized query Q_{norm} is converted into its vector representation using the platform's text embedding model (including multiple large language models). This vector representation is fed into a pre-trained intent classification model. The model learns and outputs the predefined intelligence task category to which the query belongs (e.g., domain overview analysis, technology supply chain deduction, technology evolution path, technology opportunities and risks, and technology convergence trend risk assessment). This determination is made by calculating the semantic similarity (e.g., cosine similarity) between the vector representation and the prototype vectors of each predefined intelligence task category, ultimately resulting in the determined output category.

[0096] S4. Extract key parameters related to the intent: Based on the identified intelligence task categories and standardized expressions, extract key parameters such as time span and country comparison dimensions of specific intelligence tasks.

[0097] S5. Extracting explicit metadata requirements from queries: Leveraging a large language model to identify and link entities from Q_{norm} to the knowledge base, combined with key parameters, we identify explicit metadata requirements. For example, if the query is "How do I interpret the patent growth rate for technology T1 over the past three years?", then "patent growth rate" is an explicit metadata requirement, pointing to a specific field in the metadata database that represents patent growth rates. These explicit requirements will guide the subsequent construction of metadata prompts.

[0098] Step 2: Build a metadata-based thought chain and prompt word engineering system. Specifically, Step 2 includes the following sub-steps:

[0099] S6. Retrieve metadata relevant to the query intent from the metadata knowledge base: Based on the structured query object parsed in Step 1 (including information such as system-defined intelligence task categories and entities), the system queries the "Technology Development Situation Awareness Platform Analysis Results Metadata Database." This database stores the structured analysis results of various technology entities and concepts from previous modules (data input, domain overview, technology law mining, and technology value analysis). For example, the metadata set for technology entity X includes information on the distribution of technology entities, annual trends in patent numbers, descriptions of technology theme evolution paths, comprehensive value assessment scores, and a list of core related literature. The system retrieves a subset of metadata related to the entities and key parameters in the structured query object to form a candidate metadata set.

[0100] S7. Filter metadata that matches user explicit / implicit requirements: Infer explicit and implicit metadata requirements from intelligence tasks based on structured query objects and association rules, and filter metadata sets related to the requirements from candidate metadata sets.

[0101] S8. Select or generate task-specific thought chain templates: The system maintains a library of intelligence task thought chain templates. Each template corresponds to a specific intelligence task category and defines a series of logical reasoning steps (also known as subtask nodes) that guide the large language model in structured thinking. Based on the intelligence task category in the structured query object, the system selects the most matching thought chain template. If no template in the library fully matches the current query intent, adaptive adjustments are made based on the template with the highest semantic similarity.

[0102] S9. Constructing metadata-enhanced prompt word sequences: For the jth node in the selected thought chain, the system constructs a specific prompt word content for it, denoted as Prompt_j. Prompt_j is generated by combining the following five elements: (a), (b), (c), (d), and (e):

[0103] (a) Based on the specific instructions of node j, guide the large language model to perform the analysis task;

[0104] (b) The subtask defined by node j that requires a large language model to solve;

[0105] (c) The core conclusion summary output by the large language model of the previous j-1 nodes, which serves as the context of the current node step;

[0106] (d) A subset of metadata directly related to the analysis task of the current node j, extracted from the metadata set filtered in step S7. These metadata are embedded in the prompt word in a structured form (e.g., a JSON fragment, a list of key-value pairs) and serve as known facts and core evidence for the large language model reasoning;

[0107] (e) A description or requirement of the expected output format of the large language model at this node (e.g., output bullet points, tables, specific JSON structure).

[0108] The design of this Prompt_j is one of the core technical improvement details of this invention. It ensures that the reasoning process of the large language model is closely centered around the platform's existing, high-quality, structured metadata that has been analyzed by previous modules, thereby improving the accuracy and controllability of reasoning.

[0109] S10. Construct a complete input sequence for the large language model: Organize the Prompt_j generated for each node in the instantiated thought chain (which already contains all the context information required by the node, including metadata and the conclusion summary of the previous node) in sequence to form an ordered complete input sequence ready to be submitted to the large language model for step-by-step reasoning.

[0110] Step 3: Execute the thought chain reasoning driven by the large language model and generate the first draft of the report. Specifically, step 3 includes the following sub-steps:

[0111] S11. Initialize Large Language Model Dialogue: Select and initialize a large language model, presetting inference parameters such as temperature, internal parameters for evaluating the confidence level of thought chain node results, the maximum number of output tokens per node, and the maximum memory length for the entire dialogue. The model is loaded on the server and verified to be operating normally through standard test cases.

[0112] S12. Reasoning about the First Node of the Thought Chain: First, the first node of the thought chain is processed, prompt content is constructed, and input into the initialized large language model. Based on the instructions in Prompt_1, the sub-question, and the provided metadata context, the model generates a preliminary analysis or answer to the first node of the thought chain.

[0113] S13. Save the first node output and extract context: The system saves the output of the first node. Simultaneously, it processes the output (including extracting key information and conclusion summaries through rules or small models) to obtain dynamic context information for subsequent node reasoning.

[0114] S14. Step-by-step reasoning (subsequent nodes): For subsequent nodes in a thought chain, the system first updates the corresponding component, integrating and populating the dynamic context information output by all previous nodes. This updated prompt, which includes the complete historical context and the metadata required for the current node, is then input into the large language model to obtain the model's output for the current node.

[0115] S15. Loop through step S14 until all thought chain nodes are processed: follow the thought chain logic path, and the output of each step is generated based on the conclusions of the previous steps in the thought chain and the metadata context reasoning of the current node.

[0116] Step 4: Perform metadata consistency checks on the draft report, update the report content, and present the research report and evidence. Specifically, Step 4 includes the following sub-steps:

[0117] S16. Calculate and store the confidence level of each node in the thought chain: As the large language model generates the output for each node, it performs an internal consistency check to determine whether there are any factual inconsistencies between the output and the metadata in the prompt word. This confidence level is then used to assess the output quality. If the confidence level is found to be below a set threshold, the node and its output are marked as low confidence.

[0118] S17. Metadata alignment and output retry for low-confidence nodes: To avoid infinite loops, the system has a preset maximum number of retries. For nodes marked as low-confidence and their output, the system will attempt to correct the output. The prompt word and the low-confidence output are re-input into the large language model, resulting in the updated output, which is then compared again with the metadata in the prompt word for factual accuracy until the confidence level exceeds a threshold or the maximum number of cycles is reached.

[0119] S18. Preliminary summary of structured research report: Organize and summarize the outputs of all nodes in the thinking chain in a logical order to form a research report on the user's original needs that includes a reasoning process.

[0120] S19. Fact-checking and evidence link visualization: Create citation links for metadata cited in research reports. In the final presentation, users can click to view the source of evidence. Key numerical or factual assertions are re-verified with the original data source.

[0121] S20. Calculate the comprehensive confidence of the research report: Calculate the comprehensive confidence score of the research report based on the confidence of the output content of each node using aggregation methods such as weighted averaging.

[0122] S21. Research Report Presentation and Confidence Grading Strategies: Different presentation strategies are employed based on the confidence score. High-confidence answers are presented directly; medium-confidence answers may include uncertainty explanations or prompts about key assumptions; and low-confidence answers may suggest that users should "consider this issue complex and provide only for reference. We recommend consulting other information sources or experts."

[0123] S22. Language style fine-tuning: Based on a locally stored intelligence research report knowledge base, a large language model is used to fine-tune and polish the scientific and technological intelligence language style of the research report, ensuring that its expression meets the standard, objective, and rigorous requirements of professional intelligence analysis reports while improving readability.

[0124] S23. Merge data analysis images and text into a Word document: Based on a pre-defined Word layout template, the optimized research report text, images corresponding to each analysis module's metadata (including bar charts, line charts, radar charts, and mulberry charts), metadata evidence links, and overall confidence levels are integrated to form a complete, final Word version of the research report, which is automatically exported to a designated path. Key metadata and images are also structured and exported to a designated path.

[0125] Step 5: Archive high-quality scientific data in the research report. Specifically, Step 5 includes the following sub-steps:

[0126] S24. Positive Feedback Collection and Knowledge Base Archiving: Research reports with high confidence and positive user feedback, along with their corresponding standardized representations, instantiated thought chains, and key metadata, are stored in a structured manner on the server platform where the system platform resides. This is used for continuous learning and knowledge accumulation to improve the quality of the platform's automated analysis.

[0127] S25. Negative Feedback Collection: When metadata entries frequently cause large language models to generate low-confidence conclusions and negative user feedback, the system will flag the metadata entries and prompt platform developers and data managers to verify or update them, thereby improving the overall data quality of the platform.

[0128] S26. Dynamic maintenance and expansion of the thought chain template library: Based on the analysis of the actual reasoning paths in high-frequency, high-quality questions and answers, the system can optimize existing thought chain templates and incorporate them into the technical intelligence task thought chain template library, realizing the dynamic maintenance and expansion of the template library.

[0129] S27. User Personalized Preference Learning: During the long-term operation of the system platform, the system collects information about the query habits, technical areas of interest, preferences for the level of detail in answers, etc. of specific users or user groups, and uses these preferences as a reference for adjusting subsequent query parsing, thought chain selection, and answer presentation.

[0130] S28. Similar Question Retrieval and Utilization of Cached Answers: When a new user query is received, after completing Step 1 parsing, the system first queries the updated "Technology Development Situation Awareness Research Report" knowledge base and calculates semantic similarity with historical research reports. If a historical question-answer pair with high similarity and an answer quality above a set threshold is found, the cached answer can be recommended to the user or quickly adjusted based on the cached answer to improve response efficiency.

[0131] S29. Scheduling mechanism for handling concurrent user queries: Design an efficient request queue and resource scheduling mechanism to handle query requests initiated by multiple users simultaneously, ensuring the reasonable allocation of computationally intensive tasks such as large language model inference and the overall responsiveness of the system.

[0132] S30. API interface interaction with other platform modules: Ensure a stable and efficient data interface between this Q&A module and the platform's early analysis modules (data input, field overview, technical law mining, and technical value analysis) to obtain the latest analysis results as metadata in real time or near real time.

[0133] The intelligence task question-answering module of the present invention, through the above series of refined design steps, achieves in-depth intelligent interpretation of technology development trends and efficient intelligence support. The specific effects are reflected in the following aspects:

[0134] (1) Based on the high-quality data analysis results calculated by the aforementioned modules of the platform, a metadata-driven mechanism for technology development trends was constructed. This invention uses this structured, highly reliable situational awareness information as a contextual knowledge base for large language model reasoning. This significantly improves the accuracy, controllability, and consistency with the overall platform cognition of the LLM in professional domain question-answering tasks, thereby effectively alleviating the problem of general large models easily generating "hallucinations" or irrelevant outputs due to the lack of deep domain knowledge in scenario-based applications.

[0135] (2) The innovative approach combines the instantiation of thought chains with metadata knowledge base retrieval and enhanced generation depth. Based on metadata matching, a clear and followable logical reasoning framework and confidence evaluation indicators are constructed for intelligence analysis tasks. This enables large language models to conduct orderly and step-by-step analysis, rather than divergent and unpredictable text generation. Each step is based on the analysis of verifiable data, rather than the previous divergent, unpredictable, black-box text generation. This better meets the requirements of technological development situation awareness for the rigor of the analysis process and the accuracy of the conclusions.

[0136] (3) The technical solution of the present invention not only records the textual reasoning steps, but more importantly, records the key metadata entries that each reasoning step relies on. Ultimately, this reasoning path (textual steps, associated metadata content, and confidence) can be visualized for the user. This significantly improves the data traceability of technology development situation awareness and greatly enhances the interpretability and credibility of the final technology development situation research report.

[0137] The units involved in the embodiments of the present invention may be implemented in software or hardware, and the units described may also be provided in a processor. In some cases, the names of these units do not limit the units themselves.

[0138] According to one aspect of an embodiment of the present invention, a computer program product or computer program is provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0139] As another aspect, embodiments of the present invention further provide a computer-readable medium, which may be included in the electronic device described in the above embodiments, or may exist independently and not incorporated into the electronic device. The computer-readable medium carries one or more programs, and when executed by the electronic device, the electronic device implements the methods described in the above embodiments.

Claims

1. A technology development situation awareness system, characterized in that: include: Data input module, domain overview module, intelligence task question and answer module, and data output module; The data input module is used to import technical document carriers and search formulas to obtain domain data sets; The domain overview module is used to identify the technical subject of the input data set, obtain descriptive statistical analysis results, and implement the domain overview function based on the descriptive statistical analysis results; The analysis result metadata database is used to store the structured analysis results of various technical entities and concepts. The system retrieves the metadata subsets related to the entities and key parameters in the structured query object to form a candidate metadata set; The intelligence task question-and-answer module is used to integrate existing technology analysis results, build a thought chain for interpreting the development trend of technology in the field, and use the large language model combined with the thought chain to perform technology supply chain deduction, technology evolution path question-and-answer, technology opportunity identification and risk point question-and-answer, technology transformation law prediction, and technology integration trend analysis. It specifically includes: Step 1: receiving user questions and performing in-depth semantic analysis; Step 2: building a metadata-based thought chain and prompt word engineering system; Step 3: executing the thought chain reasoning driven by the large language model and generating a draft report; Step 4: performing metadata consistency check on the draft report, performing report content update, and presenting the research report and evidence; It specifically includes sub-steps: S16. Calculate and store the confidence level of each node in the thought chain: As the large language model generates the output content for each node, an internal consistency check is performed to determine whether there are factual inconsistencies between the output content and the metadata in the prompt word. A confidence level is then obtained to assess the output quality. If the confidence level is found to be below a set threshold, the node and its output are marked as low confidence. S17. Metadata alignment and output retry for low-confidence nodes: A maximum number of retries is preset. For nodes marked as low-confidence and their output content, output corrections are attempted. The prompt word and low-confidence output content are re-input into the large language model to obtain the updated output content, which is then compared again with the metadata in the prompt word for factuality, until the confidence level exceeds the threshold or the maximum number of cycles is reached. S18. Preliminary summary of the structured research report: Organize and summarize the output of all nodes in the thinking chain in a logical order to form a research report that includes the reasoning process for the original user needs; S19. Fact-checking and evidence link visualization: Citation links are established for metadata cited in research reports. In the final presentation, users can click to view the source of evidence. Key numerical or factual assertions are re-verified with the original data source. S20. Calculate the overall confidence of the research report: Aggregate the confidence of the output content of each node and calculate the overall confidence score of the research report; S21. Research Report Presentation and Confidence Grading Strategies: Different presentation strategies are employed based on the confidence score. High-confidence answers are presented directly; medium-confidence answers are accompanied by uncertainty explanations or prompts to the user regarding key assumptions; and low-confidence answers are recommended to the user for consultation with other information sources or experts. S22. Language Style Fine-tuning: Based on a locally stored intelligence research report knowledge base, a large language model is used to fine-tune the scientific and technological intelligence language style of the research report, ensuring that its expression meets the standard, objective, and rigorous requirements of professional intelligence analysis reports while also improving readability. S23. Merge data analysis images and text into a document: Based on a pre-determined layout template, integrate the optimized research report text, images corresponding to each analysis module metadata, metadata evidence links, and comprehensive confidence to form a complete final document research report, which is then output to the specified path location. Among them, key metadata and images are structured and output to the specified path location. Step 5: Archive high-quality scientific data from research reports; The data output module is used to output the results of the intelligence task question and answer module.

2. The technology development situation awareness system according to claim 1, characterized in that: The technical document carrier includes patent data sets and paper data sets.

3. The technology development situation awareness system according to claim 1, characterized in that: The technical subject identification of the input data set specifically includes: using a name disambiguation method to identify the subject of the country, institution, and secondary institution.

4. The technology development situation awareness system according to claim 3, characterized in that: The descriptive statistical analysis results specifically include: country technology distribution, institution technology distribution, technology category distribution and technology life cycle curve.

5. The technology development situation awareness system according to claim 1, characterized in that: In step 1, the user question is received and deep semantic analysis is performed, which specifically includes the following sub-steps: S1. Obtain user input from the user requirements interface: Utilize the interactive interface to obtain the user's input, representing their technology development situation awareness needs (Q_{raw}). This Q_{raw} is captured and used as the topic of the technology development situation awareness research report and the starting point for subsequent processing. S2. Preprocessing and normalization of the raw query: After receiving Q_{raw}, perform preprocessing operations to obtain the normalized representation Q_{norm} of the technology development situation awareness requirements; S3. Perform intent analysis on Q_{norm}: Convert Q_{norm} into a vector representation using a text embedding model. This vector representation is fed into a pre-trained intent classification model, which learns and outputs the pre-defined intelligence task category to which the query belongs. S4. Extract key parameters related to intent: Extract key parameters based on the identified intelligence task categories and standardized representations; S5. Extracting explicit metadata requirements in queries: Leveraging a large language model to identify and link entities from Q_{norm} to the knowledge base, combined with key parameters, to identify explicit metadata requirements.

6. The technology development situation awareness system according to claim 1, characterized in that: In step 2, the construction of a metadata-based thought chain and prompt word engineering system specifically includes the following sub-steps: S6. Retrieving metadata related to the query intent in the metadata knowledge base: Based on the structured query object obtained by parsing in step 1, query the metadata database, which stores the structured analysis results; Retrieve metadata subsets related to entities and key parameters in the structured query object to form a candidate metadata set; S7. Filter metadata that matches user explicit / implicit requirements: Based on the structured query object and association rules, infer explicit and implicit metadata requirements from the intelligence task, and filter the candidate metadata sets to obtain metadata sets that are relevant to the requirements. S8. Select or generate a thought chain template for the selected task: Maintain a library of intelligence task thought chain templates, where each template corresponds to a selected intelligence task category and defines logical reasoning steps that are used to guide the large language model to perform structured thinking; Based on the intelligence task category in the structured query object, the best matching thought chain template is selected. If there is no template that fully matches the current query intent in the library, adaptive adjustment is performed based on the template with the highest semantic similarity. S9. Construct metadata-enhanced prompt word sequence: For the jth node of the selected thought chain, construct a set prompt word content for it, denoted as Prompt_j. Prompt_j is generated by combining the following five elements (a), (b), (c), (d), and (e): (a) Based on the specific instructions of node j, guide the large language model to perform the analysis task; (b) The subtask defined by node j that requires a large language model to solve; (c) The core conclusion summary output by the large language model of the previous j-1 nodes, which serves as the context of the current node step; (d) A subset of metadata directly related to the analysis task of the current node j, extracted from the metadata set filtered in step S7. These metadata are embedded in the prompt words in a structured form and serve as known facts and core evidence for the large language model reasoning; (e) a description or requirement of the expected output format of the large language model at this node; S10. Construct a complete input sequence for the large language model: organize the Prompt_j generated by each node in the instantiated thinking chain in sequence to form an ordered, complete input sequence ready to be submitted to the large language model for step-by-step reasoning.

7. The technology development situation awareness system according to claim 1, characterized in that: In step 3, the execution of thought chain reasoning and generation of a draft report driven by a large language model specifically includes the following sub-steps: S11. Initialize Large Language Model Dialog: Select and initialize a large language model, pre-set inference parameters, load the model on the server, and verify its normal operation through standard test cases. S12. Reasoning about the First Node of the Thought Chain: First, process the first node of the thought chain, construct the prompt word content, and input it into the initialized large language model; The model generates a preliminary analysis or answer to the first node of the thought chain based on the instructions, sub-questions, and metadata context in prompt word Prompt_1. S13. Saving the first node output and extracting context: The system saves the output content of the first node; at the same time, the output content is processed to obtain dynamic context information for subsequent node reasoning; S14. Step-by-step reasoning in thought chains: For subsequent nodes in a thought chain, the corresponding components are first updated, integrating the dynamic context information output by all previous nodes. This updated prompt, which includes the complete historical context and the metadata required for the current node, is then input into the large language model to obtain the model's output for the current node. S15. Loop through step S14 until all thought chain nodes are processed: follow the thought chain logic path, and the output of each step is generated based on the conclusions of the previous steps in the thought chain and the metadata context reasoning of the current node, thereby forming a draft report.

8. The technology development situation awareness system according to claim 1, characterized in that: In step 5, the high-quality scientific data in the archiving research report specifically includes the following sub-steps: S24. Positive Feedback Collection and Knowledge Base Archiving: High-confidence research reports that have received positive user feedback, along with their corresponding standardized representations, the instantiated thought processes used, and key metadata, will be structured and stored on the system's server platform for continuous learning and knowledge accumulation to enhance the quality of the platform's automated analysis. S25. Negative Feedback Collection: When metadata entries frequently cause large language models to generate low-confidence conclusions and negative user feedback, the metadata entries will be flagged to prompt platform developers and data managers to review or update them. S26. Dynamic maintenance and expansion of the thought chain template library: Based on analysis of actual reasoning paths in high-frequency, high-quality Q&A, optimize existing thought chain templates and incorporate them into the technical intelligence task thought chain template library, enabling dynamic maintenance and expansion of the template library. S27. User Preference Learning: Over the long term, the system collects information about the query habits, technical areas of interest, and preferences for answer detail from selected users or user groups. These preferences are then used as a guide to adjust subsequent query parsing, thought chain selection, and answer presentation. S28. Similar Question Retrieval and Utilization of Cached Answers: When a new user query is received, after fully parsing in Step 1, the updated knowledge base is first queried and the semantic similarity with historical research reports is calculated. If a historical question-answer pair with similarity and answer quality above a set threshold is found, the cached answer is recommended to the user or quickly adjusted based on the cached answer. S29. Scheduling mechanism for concurrent user queries: Design a request queue and resource scheduling mechanism to handle query requests initiated by multiple users simultaneously. S30. Interaction with API interfaces of other modules of the platform: Ensure that there is a data interface between the Q&A module and the data input module and the domain overview module to obtain the latest analysis results as metadata.

Images