A method, apparatus, equipment, and medium for constructing a knowledge graph of technological strategy intent.

Abstract

This application relates to a method, apparatus, device, and medium for constructing a knowledge graph of science and technology strategy intent, belonging to the field of natural language processing technology. The method includes: constructing multiple domain ontology types for science and technology strategy intent; modeling the named entity recognition task as a sequence labeling problem under the constraints of domain ontology types, and using a BIO labeling system to extract entities from science and technology strategy text using a named entity recognition model; under the domain / range constraints of the relationships between various domain ontology types, combining trigger word sets and sentence templates, using a rule extraction algorithm to extract entity relationships from the science and technology strategy text, and storing the extracted structured knowledge triples in a graph database to construct a knowledge graph of science and technology strategy intent. This method can reduce the cost of domain labeling while ensuring the robustness of entity extraction and the interpretability of relationship extraction, providing a structured knowledge foundation for science and technology strategy trend analysis and decision support.

Description

Technical Field

[0001] This application relates to the field of natural language processing technology, and in particular to a method, apparatus, device, and medium for constructing a knowledge graph of technological strategy intent. Background Technology

[0002] As the core carriers of scientific and technological development guidance, resource allocation planning, and implementation paths, science and technology strategy texts are generally characterized by their length and complex structure. Crucially, many of the core strategic intentions contained within these texts are presented implicitly and indirectly, rather than explicitly stated. This significantly limits traditional manual analysis methods: not only is the analysis inefficient and time-consuming, but it also makes it difficult to systematically sort and integrate the key information from massive amounts of science and technology strategy texts, failing to form a standardized, calculable knowledge structure, and consequently, hindering accurate assessment of the scientific and technological landscape and scientific decision-making.

[0003] Knowledge graphs organize knowledge in the form of "entity-relationship-attribute" triples, providing an effective means for the structured expression and relational analysis of unstructured text information. Introducing knowledge graphs into the field of science and technology strategy analysis helps to integrate strategy elements across documents and multiple dimensions, providing a foundation for science and technology situation awareness and decision support. However, the construction of science and technology strategy intent knowledge graphs still faces two challenges: first, the lack of a unified domain semantic framework makes it difficult to standardize the modeling of strategy intent; second, the complex entity types and diverse expressions in strategy texts make automatic extraction of entities and relations difficult, and existing technologies cannot meet the accuracy and efficiency requirements of knowledge graph construction. Summary of the Invention

[0004] Therefore, it is necessary to provide a method, apparatus, equipment, and medium for constructing a knowledge graph of technological strategy intent to address the aforementioned technical problems.

[0005] A method for constructing a knowledge graph of technology strategy intent, the method comprising:

[0006] Construct multi-domain ontology in the intent of science and technology strategy, including subject, technology, strategic intent, system and regulation, event and action, and capability assessment;

[0007] Under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete entity extraction in science and technology strategy text.

[0008] Under the domain / range constraints of the relationships between various domain ontology, and by combining the trigger word set and sentence template, a rule extraction algorithm is used to extract entity relationships from the science and technology strategy text. The structured knowledge triples of the entities and entity relationships extracted are stored in a graph database to construct a knowledge graph of science and technology strategy intent.

[0009] In one embodiment, multiple domain ontology types are constructed within the context of technological strategy intent, including:

[0010] To address the needs of technology strategy analysis, a combined top-down and bottom-up strategy is adopted to construct domain ontologies and corresponding ontology conceptual attributes for six core categories: subject, technology, strategic intent, system and regulation, event and action, and capability assessment. Subclasses corresponding to each domain ontology are abstracted and summarized based on specific examples. The top-down strategy designs the top-level conceptual structure starting from the overall analytical needs of technology strategy intent, while the bottom-up strategy summarizes fine-grained concepts from specific technology strategy text instances to improve the usability and coverage of the domain ontologies. The domain ontology construction process follows the Stanford seven-step method and a mechanism that prioritizes the reuse of existing related ontologies.

[0011] Among them, the subject category describes the implementers of the technology strategy actions, the technology category describes the key technology directions and their sub-fields that the technology strategy focuses on, the strategy intent category describes the goals of the technology strategy, the system and regulations category describes the institutional environment that supports or constrains the implementation of the technology strategy, the event and action category describes specific technology strategy behaviors, and the capability assessment category describes the impact of technology strategy behaviors on technological capabilities, competitive advantages, and risk factors.

[0012] In one embodiment, ontology concept attributes are divided into two categories: object attributes and data attributes;

[0013] The domain and range of object attributes are both classes in the domain ontology, used to describe the binary semantic relationships between various domain ontology types, including inclusion, category, instance, and attribute relationships. Among them, the inclusion relationship describes the whole-part association between the parent class and the child class; the category relationship reflects the inheritance relationship between the child class and the parent class; the instance relationship represents the correspondence between a specific instance and a class; and the attribute relationship describes the attribute association between different core classes.

[0014] Data attributes are defined by various domain ontology and have strings, numbers, and dates as value ranges. They are used to represent the inherent attribute characteristics of concept instances. Among them, the data attributes of the core class include name, establishment time, R&D start time, proposal time, system number, occurrence time and capability score. Subclasses inherit the data attributes of the corresponding core class and add exclusive data attributes according to their own characteristics.

[0015] In one embodiment, the domain ontology satisfies the following constraints:

[0016] Transitive attribute constraints are used to define relationships between object attributes, enabling logical connections between core classes and reasoning about technological strategy intentions.

[0017] Attribute and feature constraints are used to clarify the rules for data attribute values, ensuring data quality and the rationality of reasoning.

[0018] Hierarchical association constraints are used to regulate the inheritance relationship between core classes and subclasses, avoiding semantic confusion.

[0019] In one embodiment, under the constraint of the domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem, and the BIO labeling system is used for sequence labeling, including:

[0020] Under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem, that is, given a character-level sequence of science and technology strategy text as input. Learning functions through models Output the corresponding label sequence Among them, the character-level technology strategy text sequence is first decomposed by a hierarchical punctuation segmentation strategy, and then obtained by converting the text into character level. for The first in One character, For the first The label corresponding to each character , for The total number of characters in the text For a set of tags, defined as ;in, This represents the entity type determined by the domain ontology type constraints. For a collection of entity types, Indicates the starting position of the entity. Indicates the internal location of an entity. Indicates a non-physical location.

[0021] In one embodiment, the RoBERTa-wwm–IDCNN–CRF named entity recognition model is used to sequentially perform text feature encoding, feature enhancement, and decoding to complete entity extraction from science and technology strategy texts, including:

[0022] The RoBERTa-wwm–IDCNN–CRF named entity recognition model includes: RoBERTa-wwm layer, IDCNN layer and global decoding layer. The global decoding layer consists of linear layer and CRF.

[0023] The RoBERTa-wwm layer serves as a pre-trained encoder, processing the input character-level technology strategy text sequence. Encoded as a context-sensitive vector sequence , is represented as:

[0024] ;

[0025] in, For the set of real numbers, Let be the hidden layer dimension of the pre-trained encoder. The first in The context semantic vector corresponding to each character is represented as follows: The context semantic vector integrates the local and global semantic information of characters, providing a solid foundation for subsequent feature extraction;

[0026] The IDCNN layer uses two dilated convolutional layers with different dilation rates to process the input vector sequence. The feature-enhanced vector sequence is obtained after one iteration. , is represented as:

[0027] ;

[0028] in, This is a vector sequence after feature enhancement by the IDCNN layer. The feature dimension of the output vector of the IDCNN layer; The first in The character is the first The output after a dilated convolutional layer is represented as follows: ,in, It is a ReLU nonlinear function. and For the first Learnable parameters of a dilated convolutional layer This represents the half-width of the window on either side of the center of the convolution kernel. For the first The kernel size of a dilated convolutional layer For the first The expansion rate of a hollow convolutional layer The first in The convolution input window for each character is... , for The first in The character is the first The output after a dilated convolutional layer;

[0029] In the global decoding layer, the mapping from features to labels is first completed through a linear layer, as shown below:

[0030] ;

[0031] in, The output of the IDCNN layer Features corresponding to each character Representation of features With tags The degree of fit constitutes the base score for label prediction; and For the learnable parameters of the linear layer, The size of the tag set;

[0032] Then, a learnable transition matrix is ​​introduced into the CRF layer. Modeling the transition scores between adjacent labels, and defining the label sequence based on the base score and transition score, CRF is used. The overall score is:

[0033] ;

[0034] in, Indicates label With tags Transfer fractions between;

[0035] The core objective of model training is to maximize the true label sequence. The conditional probability is equivalent to minimizing the negative log-likelihood loss, expressed as:

[0036] ;

[0037] in, For the real label sequence The overall score, For possible label sequences The overall score;

[0038] Finally, based on the trained model, entity extraction is completed from the science and technology strategy text.

[0039] In one embodiment, under the domain / range constraints of relationships between various domain ontology types, a rule extraction algorithm is used to extract entity relationships from science and technology strategy texts, combining trigger word sets and sentence templates. The extracted entities and their structured knowledge triples are then stored in a graph database to construct a science and technology strategy intent knowledge graph, including:

[0040] The rule extraction algorithm uses candidate entity pairs obtained from entity extraction. As input, the relationships between various domain ontology types are first utilized. Type pruning is performed only if the entity type satisfies the domain / range constraints. and At that time, relationships are allowed Proceed to subsequent matching; among them, and Representing entities respectively and entity type Representing relations The domain, Representing relations Range; relation The domain and range of the domain are both classes in the domain ontology;

[0041] Subsequently, for each relation type after type pruning, a set of trigger words and a sentence template are constructed. When a trigger word appears in the text window and satisfies the sentence template structure, a structured knowledge triple is output. When the same candidate entity pair triggers multiple relations, the principle of prioritizing the strength of evidence is adopted, that is, relations with more explicit trigger words, closer entity distance, and stronger type constraints are retained first; relations with semantically coexistent relations are allowed to coexist on multiple sides, and evidence sentences are retained in relation attributes to support manual review and interpretable analysis;

[0042] Finally, the structured knowledge triples are written into the graph database and stored in the form of nodes-edges-attributes. Entities are mapped to nodes, and node labels are consistent with entity types. Relationships are mapped to edges, and edge types are consistent with relation types, thus constructing a knowledge graph of science and technology strategy intent.

[0043] A device for constructing a knowledge graph of technological strategy intent, the device comprising:

[0044] The domain ontology construction module is used to construct various domain ontology types in the context of science and technology strategy intent. These types include subject-based, technology-based, strategy intent-based, institutional and regulatory-based, event and action-based, and capability assessment-based.

[0045] The entity extraction module is used to model the named entity recognition task of science and technology strategy text as a sequence labeling problem under the constraint of domain ontology type. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete the entity extraction in science and technology strategy text.

[0046] The relation extraction and graph construction module is used to extract entity relations from science and technology strategy texts by combining trigger word sets and sentence templates under the domain / range constraints of relations between various domain ontology. The extracted entities and structured knowledge triples are stored in the graph database to construct a science and technology strategy intent knowledge graph.

[0047] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps:

[0048] Construct multi-domain ontology in the intent of science and technology strategy, including subject, technology, strategic intent, system and regulation, event and action, and capability assessment;

[0049] Under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete entity extraction in science and technology strategy text.

[0050] Under the domain / range constraints of the relationships between various domain ontology, and by combining the trigger word set and sentence template, a rule extraction algorithm is used to extract entity relationships from the science and technology strategy text. The structured knowledge triples of the entities and entity relationships extracted are stored in a graph database to construct a knowledge graph of science and technology strategy intent.

[0051] A computer-readable storage medium having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0052] Construct multi-domain ontology in the intent of science and technology strategy, including subject, technology, strategic intent, system and regulation, event and action, and capability assessment;

[0053] Under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete entity extraction in science and technology strategy text.

[0054] Under the domain / range constraints of the relationships between various domain ontology, and by combining the trigger word set and sentence template, a rule extraction algorithm is used to extract entity relationships from the science and technology strategy text. The structured knowledge triples of the entities and entity relationships extracted are stored in a graph database to construct a knowledge graph of science and technology strategy intent.

[0055] The aforementioned method, apparatus, equipment, and medium for constructing a knowledge graph of technological strategy intent achieve the transformation from unstructured technological strategy text to a structured knowledge graph through a process of "ontology modeling—entity recognition—relation extraction—graph construction." Compared with existing technologies, it has the following beneficial effects: First, by constructing a domain ontology of elements such as subject categories, technology categories, strategy intent categories, institutional and regulatory categories, event and action categories, and capability assessment categories in technological strategy intent, it provides a unified standard for semantic modeling of technological strategy intent; Second, it models the named entity recognition task as a sequence labeling problem, adopts the BIO labeling system, and constructs a RoBERTa-wwm–IDCNN–CRF named entity recognition model. RoBERTa-wwm provides high-quality contextual representation, and IDCNN enhances local and mid-range feature modeling. CRF ensures the legitimacy of BIOs and improves boundary accuracy through label transfer constraints and global decoding, ultimately accurately identifying entities belonging to specific ontology categories from unstructured text, providing high-quality node input for subsequent knowledge graph construction. Subsequently, considering the implicit nature of relational expressions and the scarcity of domain annotations, a rule extraction algorithm is employed, combining domain / range constraints on relations between domain ontology and trigger word sets and sentence templates, to achieve interpretable entity relation extraction and generate structured knowledge triples. Finally, the structured knowledge triples are stored in a graph database to construct a science and technology strategy intent knowledge graph. This application can reduce the cost of domain annotation while ensuring the robustness of entity extraction and the interpretability of relation extraction, providing a structured knowledge foundation for science and technology strategy trend analysis and decision support. Attached Figure Description

[0056] Figure 1 This is a flowchart illustrating a method for constructing a knowledge graph of technological strategy intent in one embodiment;

[0057] Figure 2 This is a schematic diagram of a technology strategy intent knowledge graph construction device in one embodiment;

[0058] Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0060] In one embodiment, such as Figure 1 As shown, a method for constructing a knowledge graph of technological strategy intent is provided, including the following steps:

[0061] Step 1: Construct multi-domain ontology in the science and technology strategy intent, including subject, technology, strategy intent, system and regulation, event and action, and capability assessment.

[0062] Step 2: Under the constraint of the domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete the entity extraction in the science and technology strategy text.

[0063] Step 3: Under the domain / range constraints of the relationships between various domain ontology, and combining the trigger word set and sentence template, the rule extraction algorithm is used to extract entity relationships from the science and technology strategy text. The structured knowledge triples of the entities and entity relationships extracted are stored in the graph database to construct the science and technology strategy intent knowledge graph.

[0064] In the aforementioned method for constructing a knowledge graph of technological strategy intent, firstly, by constructing multiple domain ontology types within the technological strategy intent, a unified standard is provided for the semantic modeling of technological strategy intent. Secondly, the named entity recognition task is modeled as a sequence labeling problem, employing the BIO labeling system, and a RoBERTa-wwm–IDCNN–CRF named entity recognition model is constructed. RoBERTa-wwm provides high-quality contextual representation, IDCNN enhances local and mid-range feature modeling, and CRF ensures the legality of BIO and improves boundary accuracy through label transfer constraints and global decoding. Ultimately, entities belonging to specific ontology categories can be accurately identified from unstructured text, providing high-quality node input for subsequent knowledge graph construction. Subsequently, considering the implicit nature of relational expressions and the scarcity of domain labels, a rule extraction algorithm is used, combining the domain / range constraints of relations between domain ontology types with trigger word sets and sentence templates, to achieve interpretable entity relation extraction and generate structured knowledge triples. Finally, the structured knowledge triples are stored in a graph database to construct the technological strategy intent knowledge graph. This application can reduce the cost of domain annotation while taking into account the robustness of entity extraction and the interpretability of relation extraction, and can provide a structured knowledge foundation for the analysis of science and technology strategy trends and decision support.

[0065] In one embodiment, multiple domain ontology types are constructed within the context of technological strategy intent, including:

[0066] Domain ontology aims to formally describe concepts, relationships, and constraints within a specific domain, serving as a crucial semantic foundation for knowledge graph construction. Addressing the needs of technology strategy analysis, a combined top-down and bottom-up strategy is employed to construct domain ontology and corresponding ontology conceptual attributes for six core classes: subject class, technology class, strategy intent class, system and regulation class, event and action class, and capability assessment class. Furthermore, based on concrete instances, subclasses corresponding to each type of domain ontology are abstracted and summarized.

[0067] Among them, the subject category describes the implementers of science and technology strategy actions (such as countries, departments, research institutions and enterprises, etc.), the technology category describes the key technology directions and their sub-fields of concern in science and technology strategy, the strategy intent category describes the goals of science and technology strategy (technology leadership, technology containment, cooperative development, etc.), the system and regulations category describes the institutional environment that supports or constrains the implementation of science and technology strategy, the event and action category describes specific science and technology strategy behaviors (institutional release, technological breakthroughs, sanctions), and the capability assessment category describes the impact of science and technology strategy behaviors on technological capabilities, competitive advantages and risk factors.

[0068] The top-down strategy involves designing a top-level conceptual structure based on the overall analytical needs of the technology strategy intent, while the bottom-up strategy involves summarizing fine-grained concepts from specific technology strategy text examples to improve the usability and coverage of the domain ontology.

[0069] The domain ontology construction process follows the Stanford Seven-Step Method and a mechanism of prioritizing the reuse of existing related ontologies. The Stanford Seven-Step Method includes defining the ontology's goals and scope, collecting domain knowledge, extracting concepts, constructing a concept hierarchy, defining relationships, creating instances, and evaluating and iteratively optimizing the ontology. Prioritizing the reuse of existing related ontologies ensures semantic and logical correctness and reduces workload. Specifically, given the broad scope of the science and technology strategy domain, SUMO (Suggested Upper Merged Ontology) and FOAF (Friend of a Friend) are primarily reused for top-level architecture design and some domain class and object attribute design. Subclass domain class design references multiple representative ontologies and their subclasses. For example, definitions of systems and regulations are reused from GOV (Public Governance Ontology) and POL (Normative Text Ontology); core concepts such as technology, innovation, and R&D are reused from TIO (Technology Innovation Ontology) and PAT (Patent Technology Ontology); and concepts such as regional entities and organizational entities are reused from IRO (International Organization Ontology) and SEC (Regional Entity Ontology). Meanwhile, due to the limited research in the field of technology strategy intent, existing ontologies cannot fully cover it. Therefore, by combining previously compiled terminology and relevant literature, content lacking standard definitions is supplemented through customization to form a complete technology strategy intent ontology. In the stage of sorting out important terms in the field, conceptual terms reflecting the relationships between entities and attributes in the field are extracted through analysis of reports, literature, and books. These terms are categorized by attribute into subject categories, technology categories, strategy intent categories, institutional and regulatory categories, event and action categories, and capability assessment categories. These terms provide support for ontology classes, data attributes, object attributes, and instance descriptions.

[0070] Specifically, the definition of the domain ontology hierarchy adopts a combination of top-down and bottom-up approaches. The top-down approach refines the top-level concept of technological strategy intent into six core categories. The bottom-up approach abstracts concepts such as research institutions from specific examples and further categorizes them into subcategories. After adjustments and iterations, a hierarchical system is finally formed, comprising 6 core categories (subject category, technology category, strategy intent category, system and regulation category, event and action category, and capability assessment category), 21 second-level subcategories, and multiple third-level subcategories. The subject category includes subcategories of personnel, organizations, and nations; the technology category includes subcategories of information technology, biotechnology, aerospace technology, and new energy technology; the strategy intent category includes subcategories of leadership intent, containment intent, cooperation intent, protection intent, and development intent; the system and regulation category includes subcategories of systems and regulations; the event and action category includes subcategories of system events, technology events, geopolitical events, and actions; and the capability assessment category includes subcategories of technological capabilities, competitive advantages, and risk factors.

[0071] In one embodiment, ontology concept attributes are divided into two categories: object attributes and data attributes;

[0072] The domain and range of object attributes are both classes in the domain ontology, used to describe the binary semantic relationships between various domain ontology types, including inclusion relationships, category relationships, instance relationships, and attribute relationships. Among them, the inclusion relationship is used to describe the whole-part association between the parent class and the child class; the category relationship is used to reflect the inheritance relationship between the child class and the parent class; the instance relationship is used to represent the correspondence between a specific instance and a class; and the attribute relationship is used to describe the attribute association between different core classes. Specifically, multiple attribute relationships and inverse relationships such as "formulate", "possess intent", "develop", and "restrict" are defined between the six core classes.

[0073] Data attributes are defined by various domain ontology and have strings, numbers, dates, etc. as value ranges. They are used to represent the inherent attribute characteristics of concept instances. Among them, the data attributes of the core class include name, establishment time, R&D start time, proposal time, system number, occurrence time and capability score, etc. Subclasses inherit the data attributes of the corresponding core class and add exclusive data attributes according to their own characteristics.

[0074] In one embodiment, to standardize semantic expression and avoid contradictions and ambiguities, the domain ontology satisfies the following constraints:

[0075] Transitive attribute constraints are used to define the transitive relationships between object attributes, enabling logical connections between core classes and reasoning about technological strategy intentions. For example, the core strategy intention can be deduced from the subject-system-technology path. Attribute feature constraints are used to clarify the rules for data attribute values, ensuring data quality and the rationality of reasoning. Hierarchical association constraints are used to standardize the inheritance relationship between core classes and subclasses, avoiding semantic confusion.

[0076] Furthermore, this application selects typical instances based on the sorted domain terms to fill the ontology instance layer, uses Protégé for implementation and consistency verification, and adopts OntoQA (an ontology quality analysis and evaluation framework based on metrics) to evaluate and iteratively optimize the ontology model from dimensions such as consistency, completeness, accuracy, reusability and effectiveness, thus verifying the rationality of the domain ontology construction.

[0077] In one embodiment, under the constraint of the domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem, and the BIO labeling system is used for sequence labeling, including:

[0078] Under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem, that is, given a character-level sequence of science and technology strategy text as input. Learning functions through models Output the corresponding label sequence In particular, in order to avoid the truncation of entities, the character-level technology strategy text sequence is first decomposed by a hierarchical punctuation segmentation strategy, and then obtained by converting the text into character level. for The first in One character, For the first The label corresponding to each character , for The total number of characters in the text For a set of tags, defined as ;in, This represents the entity type determined by the domain ontology type constraints. For a collection of entity types, Indicates the starting position of the entity. Indicates the internal location of an entity. Indicates a non-physical location.

[0079] Specifically, the hierarchical punctuation segmentation strategy first segments the text by the end of the sentence (。!?), then by the semicolon (;), and finally by the comma (,), breaking the long text down into several short text fragments. At the same time, the text is converted into a character-level sequence to ensure that each token is strictly aligned with the label, providing high-quality input for model training and inference.

[0080] In one embodiment, the RoBERTa-wwm–IDCNN–CRF named entity recognition model is used to sequentially perform text feature encoding, feature enhancement, and decoding to complete entity extraction from science and technology strategy texts, including:

[0081] The RoBERTa-wwm–IDCNN–CRF named entity recognition model comprises a RoBERTa-wwm layer, an IDCNN (Iterative Dilated Convolutional Neural Network) layer, and a global decoding layer. The global decoding layer consists of linear layers and a CRF (Conditional Random Field). The RoBERTa-wwm layer enhances Chinese semantic modeling capabilities through a full-word masking strategy, better capturing the semantic information of specific expressions in fields such as terminology and proper nouns, thus adapting to the characteristics of scientific and technological strategy texts. However, entities in scientific and technological strategy texts often have long lengths and ambiguous boundaries, making point-by-point classification based solely on Transformer output insufficient for capturing local boundary triggering patterns. To enhance the modeling of local and mid-range dependent features, an IDCNN layer is introduced above the RoBERTa output for feature enhancement. Two layers of dilated convolutions expand the receptive field, efficiently capturing the internal continuity and boundary patterns of entities. Furthermore, entity annotation in scientific and technological strategy texts must strictly adhere to BIO (Browser Injection) standards, such as avoiding isolated "I" labels and logical errors like cross-category label jumps, while ensuring the consistency and accuracy of entity boundaries. Simply relying on point-by-point classification makes it difficult to capture the global dependencies between labels. Therefore, this application adopts a decoding structure of "linear layer + CRF". The linear layer completes the mapping from features to labels, and the CRF is used to model the transition constraints between labels to achieve the globally optimal decoding.

[0082] The RoBERTa-wwm layer serves as a pre-trained encoder, processing the input character-level technology strategy text sequence. Encoded as a context-sensitive vector sequence , is represented as:

[0083] ;

[0084] in, For the set of real numbers, Let be the hidden layer dimension of the pre-trained encoder. The first in The context semantic vector corresponding to each character is represented as follows: The context semantic vector integrates the local and global semantic information of characters, providing a solid foundation for subsequent feature extraction.

[0085] The IDCNN layer uses two dilated convolutional layers with different dilation rates to process the input vector sequence. The feature-enhanced vector sequence is obtained after one iteration. , is represented as:

[0086] ;

[0087] in, This is a vector sequence after feature enhancement by the IDCNN layer. The feature dimension of the output vector of the IDCNN layer; The first in The character is the first The output after a dilated convolutional layer is represented as follows: ,in, It is a ReLU nonlinear function. and For the first Learnable parameters of a dilated convolutional layer This represents the half-width of the window on either side of the center of the convolution kernel. For the first The kernel size of a dilated convolutional layer For the first The expansion rate of a hollow convolutional layer The first in The convolution input window for each character is... , for The first in The character is the first The output after a dilated convolutional layer. Dilated convolution can expand the receptive field without significantly increasing the number of parameters, enabling the model to capture long-distance dependencies in shallow layers. The output of the IDCNN layer fuses pre-trained semantic features and local dependency features, providing more comprehensive feature support for subsequent label prediction.

[0088] Considering that although the feature vector output by IDCNN integrates pre-trained semantic features and local dependency features, it still needs to be further mapped to the label space to obtain the matching degree of each position for various labels, the linear layer plays this core role. In the global decoding layer, the mapping from features to labels is first completed through the linear layer, as shown below:

[0089] ;

[0090] in, The output of the IDCNN layer Features corresponding to each character Representation of features With tags The degree of fit constitutes the base score for label prediction; and For the learnable parameters of the linear layer, The size of the tag set.

[0091] Linear layers can only provide label matching scores for single locations and cannot consider logical relationships between labels (e.g., "B-ORG (start of organization name)" should be followed by "I-ORG (internal location of organization name)" instead of "O (non-entity)" or other category labels). Therefore, a learnable transition matrix is ​​then introduced into the CRF layer. Modeling the transition scores between adjacent labels, and defining the label sequence based on the base score and transition score, CRF is used. The overall score is:

[0092] ;

[0093] in, Indicates label With tags The global score considers the label matching degree at each position and the rationality of the transition between adjacent labels, effectively distinguishing between legal sequences that conform to the BIO specification and illegal sequences with logical conflicts.

[0094] The core objective of model training is to maximize the true label sequence. The conditional probability is equivalent to minimizing the negative log-likelihood loss, expressed as:

[0095] ;

[0096] in, For the real label sequence The overall score, For possible label sequences The global score is calculated efficiently using a forward algorithm, avoiding the computational explosion problem caused by exhaustively searching all sequences and ensuring the feasibility of the training process. Finally, based on the trained model, entity extraction is completed from the science and technology strategy text.

[0097] In one embodiment, under the domain / range constraints of relationships between various domain ontology types, a rule extraction algorithm is used to extract entity relationships from science and technology strategy texts, combining trigger word sets and sentence templates. The extracted entities and their structured knowledge triples are then stored in a graph database to construct a science and technology strategy intent knowledge graph, including:

[0098] The rule extraction algorithm uses candidate entity pairs obtained from entity extraction. As input, the relationships between various domain ontology types are first utilized. Type pruning is performed only if the entity type satisfies the domain / range constraints. and At that time, relationships are allowed This process proceeds to subsequent matching, effectively reducing false positives of unreasonable relationships and ensuring semantic consistency of the graph. and Representing entities respectively and entity type Representing relations The domain, Representing relations Range; relation The domain and range are both classes in the domain ontology.

[0099] Subsequently, for each relation type after type pruning, a set of trigger words and a sentence template are constructed. When a trigger word appears in the text window and satisfies the sentence template structure, a structured knowledge triple is output. When the same candidate entity pair triggers multiple relations, the principle of prioritizing the strength of evidence is adopted, that is, relations with more explicit trigger words, closer entity distance, and stronger type constraints are retained first; relations with semantically coexistent relations are allowed to coexist on multiple sides, and evidence sentences are retained in relation attributes to support manual review and interpretable analysis.

[0100] Finally, the structured knowledge triples are written into a graph database (such as Neo4j) and stored in a node-edge-attribute manner. Entities are mapped to nodes, with node labels matching entity types, and relations are mapped to edges, with edge types matching relation types, thus constructing a knowledge graph of science and technology strategy intent. Furthermore, to meet the traceability requirements of science and technology strategy intent analysis, necessary evidence information is retained when the structured knowledge triples are added to the database: node attributes include entity specification names, original text references, source documents, and locations; edge attributes include trigger words, evidence sentence fragments, rule numbers, and optional confidence scores. This storage method enables the knowledge graph to possess two capabilities simultaneously: first, it supports structured queries and association analysis; second, it supports evidence backtracking, allowing each relation to be located back to its original textual basis, thereby improving the reliability and usability of the knowledge graph in practical science and technology strategy judgment scenarios.

[0101] In one embodiment, such as Figure 2 As shown, a device for constructing a knowledge graph of technological strategy intent is provided, comprising:

[0102] The domain ontology construction module is used to construct various domain ontology types in the context of science and technology strategy intent. These types include subject-based, technology-based, strategy intent-based, institutional and regulatory-based, event and action-based, and capability assessment-based.

[0103] The entity extraction module is used to model the named entity recognition task of science and technology strategy text as a sequence labeling problem under the constraint of domain ontology type. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete the entity extraction in science and technology strategy text.

[0104] The relation extraction and graph construction module is used to extract entity relations from science and technology strategy texts by combining trigger word sets and sentence templates under the domain / range constraints of relations between various domain ontology. The extracted entities and structured knowledge triples are stored in the graph database to construct a science and technology strategy intent knowledge graph.

[0105] Specific limitations regarding the device for constructing a knowledge graph of technological strategic intent can be found in the limitations of the method for constructing a knowledge graph of technological strategic intent described above, and will not be repeated here. Each module in the aforementioned device for constructing a knowledge graph of technological strategic intent can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.

[0106] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for constructing a knowledge graph of technological strategic intent. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0107] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0108] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to perform the following steps:

[0109] Construct multi-domain ontology in the intent of science and technology strategy, including subject, technology, strategic intent, system and regulation, event and action, and capability assessment;

[0110] Under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete entity extraction in science and technology strategy text.

[0111] Under the domain / range constraints of the relationships between various domain ontology, and by combining the trigger word set and sentence template, a rule extraction algorithm is used to extract entity relationships from the science and technology strategy text. The structured knowledge triples of the entities and entity relationships extracted are stored in a graph database to construct a knowledge graph of science and technology strategy intent.

[0112] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:

[0113] Construct multi-domain ontology in the intent of science and technology strategy, including subject, technology, strategic intent, system and regulation, event and action, and capability assessment;

[0114] Under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete entity extraction in science and technology strategy text.

[0115] Under the domain / range constraints of the relationships between various domain ontology, and by combining the trigger word set and sentence template, a rule extraction algorithm is used to extract entity relationships from the science and technology strategy text. The structured knowledge triples of the entities and entity relationships extracted are stored in a graph database to construct a knowledge graph of science and technology strategy intent.

[0116] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0117] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0118] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application.

Claims

1. A method for constructing a knowledge graph of technological strategy intent, characterized in that, The method includes: This paper constructs a multi-domain ontology for science and technology strategy intent, including subject, technology, strategy intent, system and regulation, event and action, and capability assessment. The subject category describes the implementers of science and technology strategy actions; the technology category depicts the key technological directions and their sub-fields of focus; the strategy intent category represents the goals of science and technology strategy; the system and regulation category describes the institutional environment supporting or constraining the implementation of science and technology strategy; the event and action category depicts specific science and technology strategy behaviors; and the capability assessment category depicts the impact of science and technology strategy behaviors on technological capabilities, competitive advantages, and risk factors. Under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete entity extraction in science and technology strategy text. Under the domain / range constraints of the relationships between various domain ontology, and by combining the trigger word set and sentence template, a rule extraction algorithm is used to extract entity relationships from the science and technology strategy text. The structured knowledge triples of entities and entity relationships extracted are stored in a graph database to construct a science and technology strategy intent knowledge graph. Among these, under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem, and sequence labeling is performed using the BIO labeling system, including: Under the constraints of the aforementioned domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem, that is, given a character-level sequence of science and technology strategy text as input... Learning functions through models Output the corresponding label sequence The character-level technology strategy text sequence is first decomposed using a hierarchical punctuation segmentation strategy, and then obtained by converting the text into a character-level sequence. for The first in One character, For the first The label corresponding to each character , for The total number of characters in the text For a set of tags, defined as ;in, This represents the entity type determined by the domain ontology type constraints. For a collection of entity types, Indicates the starting position of the entity. Indicates the internal location of an entity. Indicates a non-physical location; Within the domain / range constraints of relationships between various domain ontology types, a rule extraction algorithm is used, combining trigger word sets and sentence templates, to extract entity relationships from science and technology strategy texts. The extracted entities and their structured knowledge triples are then stored in a graph database, constructing a science and technology strategy intent knowledge graph, including: The rule extraction algorithm uses candidate entity pairs obtained from entity extraction. As input, the relationships between various domain ontology types are first utilized. Type pruning is performed only if the entity type satisfies the domain / range constraints. and At that time, relationships are allowed Proceed to subsequent matching; among them, and Representing entities respectively and entity type Representing relations The domain, Representing relations Range; relation The domain and range of the domain are both classes in the domain ontology; Subsequently, for each relation type after type pruning, a set of trigger words and a sentence template are constructed. When a trigger word appears in the text window and satisfies the sentence template structure, a structured knowledge triple is output. When the same candidate entity pair triggers multiple relations, the principle of prioritizing the strength of evidence is adopted, that is, relations with more explicit trigger words, closer entity distance, and stronger type constraints are retained first; relations with semantically coexistent relations are allowed to coexist on multiple sides, and evidence sentences are retained in relation attributes to support manual review and interpretable analysis; Finally, the structured knowledge triples are written into the graph database and stored in the form of nodes-edges-attributes. Entities are mapped to nodes, and node labels are consistent with entity types. Relationships are mapped to edges, and edge types are consistent with relation types, thus constructing a knowledge graph of science and technology strategy intent.

2. The method for constructing a knowledge graph of technological strategy intent according to claim 1, characterized in that, Constructing multiple domain ontology types for technology strategy intent, including: To address the needs of technology strategy analysis, a combined top-down and bottom-up strategy is adopted to construct domain ontologies and corresponding ontological conceptual attributes for six core categories: subject, technology, strategic intent, system and regulation, event and action, and capability assessment. Subclasses corresponding to each domain ontology are abstracted and summarized based on specific examples. The top-down strategy designs the top-level conceptual structure starting from the overall analytical needs of technology strategy intent, while the bottom-up strategy summarizes fine-grained concepts from specific technology strategy text examples to improve the usability and coverage of the domain ontologies. The domain ontology construction process follows the Stanford seven-step method and a mechanism that prioritizes the reuse of existing related ontologies.

3. The method for constructing a knowledge graph of technological strategy intent according to claim 2, characterized in that, The ontology concept attributes are divided into two categories: object attributes and data attributes; The domain and range of the object attributes are both classes in the domain ontology, used to describe the binary semantic relationships between various domain ontology types, including inclusion relationships, category relationships, instance relationships, and attribute relationships; wherein, the inclusion relationship is used to describe the whole-part association between the parent class and the child class; the category relationship is used to reflect the inheritance relationship between the child class and the parent class; the instance relationship is used to characterize the correspondence between a specific instance and a class; and the attribute relationship is used to describe the attribute association between different core classes. The data attributes are defined by various domain ontology and have strings, numbers, and dates as value fields, and are used to represent the inherent attribute characteristics of concept instances. Among them, the data attributes of the core class include name, establishment time, R&D start time, proposal time, system number, occurrence time and capability score. Subclasses inherit the data attributes of the corresponding core class and add exclusive data attributes according to their own characteristics.

4. The method for constructing a knowledge graph of technological strategy intent according to claim 3, characterized in that, The domain ontology satisfies the following constraints: Transitive attribute constraints are used to define relationships between object attributes, enabling logical connections between core classes and reasoning about technological strategy intentions. Attribute and feature constraints are used to clarify the rules for data attribute values, ensuring data quality and the rationality of reasoning. Hierarchical association constraints are used to regulate the inheritance relationship between core classes and subclasses, avoiding semantic confusion.

5. The method for constructing a knowledge graph of technological strategy intent according to claim 1, characterized in that, Using the RoBERTa-wwm–IDCNN–CRF named entity recognition model, text feature encoding, feature enhancement, and decoding are performed sequentially to extract entities from science and technology strategy texts, including: The RoBERTa-wwm–IDCNN–CRF named entity recognition model includes: a RoBERTa-wwm layer, an IDCNN layer, and a global decoding layer, wherein the global decoding layer consists of a linear layer and a CRF. The RoBERTa-wwm layer serves as a pre-trained encoder, processing the input character-level technology strategy text sequence. Encoded as a context-sensitive vector sequence , is represented as: ； in, For the set of real numbers, Let be the hidden layer dimension of the pre-trained encoder. The first in The context semantic vector corresponding to each character is represented as follows: The context semantic vector integrates the local and global semantic information of characters, providing a solid foundation for subsequent feature extraction; The IDCNN layer uses two dilated convolutional layers with different dilation rates to process the input vector sequence. The feature-enhanced vector sequence is obtained after one iteration. , is represented as: ； in, This is a vector sequence after feature enhancement by the IDCNN layer. The feature dimension of the output vector of the IDCNN layer; The first in The character is the first The output after a dilated convolutional layer is represented as follows: ,in, It is a ReLU nonlinear function. and For the first Learnable parameters of a dilated convolutional layer This represents the half-width of the window on either side of the center of the convolution kernel. For the first The kernel size of a dilated convolutional layer For the first The expansion rate of a hollow convolutional layer The first in The convolution input window for each character is... , for The first in The character is the first The output after a dilated convolutional layer; In the global decoding layer, the mapping from features to labels is first completed through a linear layer, as shown below: ； in, The output of the IDCNN layer Features corresponding to each character Representation of features With tags The degree of fit constitutes the base score for label prediction; and For the learnable parameters of the linear layer, The size of the tag set; Then, a learnable transition matrix is ​​introduced into the CRF layer. Modeling the transition scores between adjacent labels, and defining the label sequence based on the base score and transition score, CRF is used. The overall score is: ； in, Indicates label With tags Transfer fractions between; The core objective of model training is to maximize the true label sequence. The conditional probability is equivalent to minimizing the negative log-likelihood loss, expressed as: ； in, For the real label sequence The overall score, For possible label sequences The overall score; Finally, based on the trained model, entity extraction is completed from the science and technology strategy text.

6. A device for constructing a knowledge graph of technological strategy intent, characterized in that, The device includes: The domain ontology construction module is used to construct multiple domain ontology types in the context of science and technology strategy intent. These types include subject classes, technology classes, strategy intent classes, institutional and regulatory classes, event and action classes, and capability assessment classes. Specifically, the subject class describes the executors of science and technology strategy actions, the technology class characterizes the key technological directions and their sub-fields of focus in the science and technology strategy, the strategy intent class represents the goals of the science and technology strategy, the institutional and regulatory class describes the institutional environment that supports or constrains the implementation of the science and technology strategy, the event and action class characterizes specific science and technology strategy behaviors, and the capability assessment class characterizes the impact of science and technology strategy behaviors on technological capabilities, competitive advantages, and risk factors. The entity extraction module is used to model the named entity recognition task of science and technology strategy text as a sequence labeling problem under the constraint of domain ontology type. After using the BIO labeling system for sequence labeling, the RoBERTa-wwm-IDCNN-CRF named entity recognition model is used to perform text feature encoding, feature enhancement and decoding in sequence to complete the entity extraction in science and technology strategy text. The relation extraction and graph construction module is used to extract entity relations from science and technology strategy texts by combining trigger word sets and sentence templates under the domain / range constraints of relations between various domain ontology. The extracted entities and structured knowledge triples are stored in the graph database to construct a science and technology strategy intent knowledge graph. Among these, under the constraint of domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem, and sequence labeling is performed using the BIO labeling system, including: Under the constraints of the aforementioned domain ontology type, the named entity recognition task of science and technology strategy text is modeled as a sequence labeling problem, that is, given a character-level sequence of science and technology strategy text as input... Learning functions through models Output the corresponding label sequence The character-level technology strategy text sequence is first decomposed using a hierarchical punctuation segmentation strategy, and then obtained by converting the text into a character-level sequence. for The first in One character, For the first The label corresponding to each character , for The total number of characters in the text For a set of tags, defined as ;in, This represents the entity type determined by the domain ontology type constraints. For a collection of entity types, Indicates the starting position of the entity. Indicates the internal location of an entity. Indicates a non-physical location; Within the domain / range constraints of relationships between various domain ontology types, a rule extraction algorithm is used, combining trigger word sets and sentence templates, to extract entity relationships from science and technology strategy texts. The extracted entities and their structured knowledge triples are then stored in a graph database, constructing a science and technology strategy intent knowledge graph, including: The rule extraction algorithm uses candidate entity pairs obtained from entity extraction. As input, the relationships between various domain ontology types are first utilized. Type pruning is performed only if the entity type satisfies the domain / range constraints. and At that time, relationships are allowed Proceed to subsequent matching; among them, and Representing entities respectively and entity type Representing relations The domain, Representing relations Range; relation The domain and range of the domain are both classes in the domain ontology; Subsequently, for each relation type after type pruning, a set of trigger words and a sentence template are constructed. When a trigger word appears in the text window and satisfies the sentence template structure, a structured knowledge triple is output. When the same candidate entity pair triggers multiple relations, the principle of prioritizing the strength of evidence is adopted, that is, relations with more explicit trigger words, closer entity distance, and stronger type constraints are retained first; relations with semantically coexistent relations are allowed to coexist on multiple sides, and evidence sentences are retained in relation attributes to support manual review and interpretable analysis; Finally, the structured knowledge triples are written into the graph database and stored in the form of nodes-edges-attributes. Entities are mapped to nodes, and node labels are consistent with entity types. Relationships are mapped to edges, and edge types are consistent with relation types, thus constructing a knowledge graph of science and technology strategy intent.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 5.

Images