Entity recognition method for science and technology achievement declaration and knowledge graph construction method and device thereof

Abstract

This application provides a method and apparatus for entity recognition of scientific and technological achievement application texts and for constructing a knowledge graph. The method includes: acquiring the text of the scientific and technological achievement application to be recognized; inputting the preprocessed text into a pre-trained entity recognition model; performing compound nesting boundary detection on the text and dynamically dividing the text into semantic units; extracting the bidirectional dependency graph convolution features of each semantic unit; capturing the multi-head dynamic local features of each semantic unit based on the bidirectional dependency graph convolution features and the multi-head dynamic local features; performing state transition based on a gated hierarchical state transition mechanism according to the bidirectional dependency graph convolution features and the multi-head dynamic local features of each semantic unit, generating a state vector word by word and decoding entity labels, and outputting a structured entity sequence corresponding to the scientific and technological achievement application text. This application can improve the accuracy and precision of entity recognition of scientific and technological achievement application texts, thereby ensuring the accuracy of knowledge graph construction.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and particularly to an entity recognition method for scientific and technological achievement declaration and its knowledge graph construction method and device. Background Art

[0002] Scientific and technological achievement declaration texts usually contain a large amount of highly specialized, complex-structured, and grammar-variable technical content, such as descriptions of technical principles, elaborations of innovation points, explanations of performance indicators, and analyses of application scenarios. These texts often exhibit language characteristics such as long sentences, multiple clauses, cross-sentence references, and nested compound terms, making the manual review and knowledge extraction processes time-consuming and susceptible to subjective factors. With the continuous growth of the needs for scientific and technological project management, achievement transformation, and innovation evaluation, there is an urgent need to automatically extract structured knowledge from these texts to support technical route analysis, novelty search of achievements, intelligent review, and scientific research management decision-making.

[0003] Regarding the entity recognition of scientific and technological achievement declaration texts, there are the following problems in the prior art: 1. Most of the prior art relies on sentence segmentation methods based on punctuation or fixed windows, and cannot handle nested terms and long-distance references in scientific and technological texts, resulting in entities being wrongly split or having blurred boundaries, and it is difficult to adapt to the structural characteristics of scientific and technological achievement declaration texts.

[0004] 2. Traditional graph convolutional or sequence models usually can only capture single-hop or local dependency relationships, and are insufficient in modeling long-distance dependencies such as cross-syntactic associations between technical indicators and numerical values, and hierarchical structures within technical terms, resulting in low accuracy of entity recognition in complex technical descriptions.

[0005] 3. Conventional sequence labeling models (such as BiLSTM-CRF) cannot effectively handle multi-level nested entities, and are prone to boundary conflicts or recognition omissions when encountering hierarchical structures where "sub-entities are included in entities", lacking a mechanism to handle the multi-granularity characteristics of scientific and technological entities.

[0006] 4. Existing loss functions generally ignore the confusion problems caused by the coupling of syntactic and semantic features, and at the same time cannot solve the actual situation of extremely unbalanced categories of scientific and technological entities, and have insufficient recognition ability for rare entities (such as innovation attributes, technical indicators), and the model is prone to performance degradation in complex texts. Summary of the Invention

[0007] The purpose of this application is to provide an entity recognition method for scientific and technological achievement declaration and its knowledge graph construction method and device, which can improve the entity recognition accuracy and precision for scientific and technological achievement declaration texts to ensure the accuracy of knowledge graph construction.

[0008] Firstly, this application provides an entity recognition method for technology achievement application texts. The method includes: acquiring the technology achievement application text to be recognized; preprocessing the technology achievement application text and inputting the preprocessed text into a pre-trained entity recognition model; the preprocessing includes: data cleaning and standardization; the entity recognition model includes: a composite nested boundary detection module, a bidirectional dependency graph convolution module, a multi-head dynamic convolution gating module, a gated hierarchical state transition module, and an entity recognition module; through the composite nested boundary detection module, the text is subjected to composite nested boundary detection, and the semantic units corresponding to the text are dynamically divided; through the bidirectional dependency graph convolution module, the bidirectional dependency graph convolution features of each semantic unit are extracted; through the multi-head dynamic convolution gating module, based on the bidirectional dependency graph convolution features of each semantic unit, the multi-head dynamic local features of each semantic unit are captured; through the gated hierarchical state transition and entity recognition module, based on the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit, a state transition based on the gated hierarchical state transition mechanism is performed, a state vector is generated word by word, and entity labels are decoded to output a structured entity sequence corresponding to the technology achievement application text.

[0009] Secondly, this application also provides a method for constructing a knowledge graph for scientific and technological achievement applications. The method includes: outputting a structured entity sequence corresponding to the text of the scientific and technological achievement application by using the entity recognition method for scientific and technological achievement applications as described in the first aspect; extracting relationships from the text of the scientific and technological achievement application by using dependency path enhancement and multi-hop graph convolutional aggregation to establish inter-entity association information; and constructing a knowledge graph for scientific and technological achievement applications based on the structured entity sequence and inter-entity association information.

[0010] Thirdly, this application also provides an entity recognition device for technology achievement application texts. The device includes: a text acquisition module for acquiring the technology achievement application text to be recognized; a preprocessing module for preprocessing the technology achievement application text and inputting the preprocessed text into a pre-trained entity recognition model; the preprocessing includes: data cleaning and standardization; the entity recognition model includes: a composite nested boundary detection module, a bidirectional dependency graph convolution module, a multi-head dynamic convolution gating module, a gating hierarchical state transition module, and an entity recognition module; and a model recognition module for processing the text through the composite nested boundary detection module. The system employs a composite nested boundary detection method to dynamically segment semantic units corresponding to the text. A bidirectional dependency graph convolution module extracts the bidirectional dependency graph convolution features of each semantic unit. A multi-head dynamic convolution gating module captures the multi-head dynamic local features of each semantic unit based on its bidirectional dependency graph convolution features. A gated hierarchical state transition and entity recognition module performs state transitions based on the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit, generating state vectors word-by-word and decoding entity labels to output a structured entity sequence corresponding to the scientific and technological achievement application text.

[0011] Fourthly, this application also provides a knowledge graph construction device for scientific and technological achievement applications. The device includes: an entity recognition device for scientific and technological achievement applications as described in the third aspect, which outputs a structured entity sequence corresponding to the text of the scientific and technological achievement application; a relation extraction module, which extracts relations from the text of the scientific and technological achievement application through dependency path enhancement and multi-hop graph convolutional aggregation to establish inter-entity association information; and a graph construction module, which constructs a knowledge graph for scientific and technological achievement applications based on the structured entity sequence and inter-entity association information.

[0012] Fifthly, this application also provides an electronic device, including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the methods described in the first and second aspects above.

[0013] In a sixth aspect, this application also provides a computer-readable storage medium storing computer-executable instructions, which, when invoked and executed by a processor, cause the processor to implement the methods described in the first and second aspects above.

[0014] The entity recognition method and its knowledge graph construction method and apparatus for scientific and technological achievement application texts provided in this application, through a composite nested boundary detection mechanism, fuse dependency depth mutation and term boundary probability to dynamically identify complex nested entities and long-distance referential structures in scientific and technological texts, thereby maintaining the integrity of semantic units and reducing entity fragmentation problems; by constructing a weighted bidirectional dependency graph and introducing multi-hop graph convolution feature propagation, it can simultaneously capture long-distance syntactic associations and technical attribute features, achieving fine modeling of cross-syntactic dependencies such as technical indicators-values ​​and technical composition; through a gated hierarchical state transition and multi-head hierarchical entity decoding mechanism, it can dynamically combine boundary information, prior knowledge of technical entities, and multi-level context to effectively model nested entities and multi-granular entity structures; it also designs syntactic-semantic feature decoupling loss and category balance focus loss, enabling the model to distinguish syntactic structure and semantic information during the learning process, reducing feature confusion, and alleviating entity category imbalance problems by adjusting category weights and focus coefficients; overall, it can improve the entity recognition accuracy and precision of scientific and technological achievement application texts, thus ensuring the accuracy of knowledge graph construction. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating an entity recognition method for technology achievement application provided in this embodiment of the application; Figure 2 A comparison diagram of the effects of this method and conventional methods provided in this application embodiment; Figure 3 A comparison diagram of the effects of this method and a conventional method provided in an embodiment of this application; Figure 4 A comparison diagram of the effects of this method and a conventional method provided in an embodiment of this application; Figure 5 A comparison diagram of the effects of this method and a conventional method provided in an embodiment of this application; Figure 6 A flowchart illustrating a method for constructing a knowledge graph for scientific and technological achievement application, provided as an embodiment of this application; Figure 7 A knowledge graph diagram for scientific and technological achievement application provided in this application embodiment; Figure 8A structural block diagram of a technology achievement application entity recognition device provided in this application embodiment; Figure 9 A structural block diagram of a knowledge graph construction device for technology achievement application provided in this application embodiment; Figure 10 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0017] The technical solutions of this application will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0018] In view of the above-mentioned problems in the prior art, this application provides an entity recognition method for scientific and technological achievement application and a knowledge graph construction method and apparatus thereof. To facilitate understanding of this embodiment, the entity recognition method for scientific and technological achievement application disclosed in this application will be described in detail first. Figure 1 A flowchart illustrating an entity recognition method for technology achievement applications provided in this application embodiment is included, specifically comprising the following steps: Step S101: Obtain the text of the scientific and technological achievement application to be identified; To address the need for knowledge graph construction in the field of scientific and technological achievement application, we systematically collect application materials, including scientific and technological project proposals, technical achievement appraisal reports, patent specifications, award certificates, and related technical documents.

[0019] Step S102 involves preprocessing the text for scientific and technological achievement applications and inputting the preprocessed text into a pre-trained entity recognition model. The preprocessing includes data cleaning and standardization, including removing irrelevant tables and images, standardizing number and unit formats, and correcting typos and grammatical errors. The entity recognition model includes a composite nested boundary detection module, a bidirectional dependency graph convolution module, a multi-head dynamic convolution gating module, a gating hierarchical state transition module, and an entity recognition module. The training process of the model will be detailed later.

[0020] Step S103: The text is subjected to compound nesting boundary detection by the compound nesting boundary detection module, and the semantic units corresponding to the text are dynamically divided. In scientific and technological achievement application texts, technical terms often appear in a nested form, such as the complex nested text representation of "image recognition system based on deep convolutional neural network". At the same time, there is a long-distance reference phenomenon, such as the long-distance reference text representation of "this method refers to the technology mentioned above". Conventional sentence segmentation methods rely on punctuation marks or fixed windows, which can easily destroy the integrity of technical entities and lead to fragmentation problems in the subsequent entity recognition process.

[0021] This embodiment uses a compound nesting boundary detection module to divide semantic units. Specifically, by fusing syntactic dependency depth and term boundary probability, it dynamically detects compound nesting boundary points, thereby dividing the text into semantically complete sub-units. The detailed process will be described later.

[0022] Step S104: Extract the bidirectional dependency graph convolution features of each semantic unit through the bidirectional dependency graph convolution module; In scientific and technological achievement texts, word vector representations often ignore the technical attribute associations across syntactic dependencies, such as the metric relationship between "precision" and "98.2%", while conventional graph convolutional networks only aggregate single-hop neighbor information and cannot effectively capture long-distance technical indicator descriptions.

[0023] In this embodiment, a weighted bidirectional dependency graph is constructed, and syntactic structure and technical attribute features are fused based on the semantic weights of dependency edges. The specific process will be detailed later.

[0024] Step S105: Using the multi-head dynamic convolution gating module, capture the multi-head dynamic local features of each semantic unit based on the bidirectional dependency graph convolution features of each semantic unit; Technological achievements often possess multi-granularity characteristics. For example, a "quantum computing chip," as a composite entity, exhibits multi-granularity representations, making conventional fixed-window convolution methods ill-suited for technical terms of varying lengths. This embodiment addresses this by designing a gated dynamic convolution kernel and combining it with a multi-head mechanism to capture local combined features. The specific process will be detailed later.

[0025] Step S106: Through the gated hierarchical state transition and entity recognition module, state transition based on the gated hierarchical state transition mechanism is performed according to the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit. State vectors are generated word by word and entity labels are decoded to output the structured entity sequence corresponding to the scientific and technological achievement application text.

[0026] In scientific and technological achievement texts, technical entities often exhibit nested structures and multi-granularity characteristics. For example, "quantum" and "computing chip" in "quantum computing chip" may appear as sub-entities. Conventional sequence labeling models such as CRF are difficult to effectively capture this hierarchical structure. At the same time, long-distance referencing and association with technical attributes lead to blurred entity boundaries, increasing the difficulty of identification.

[0027] In this embodiment, a gated hierarchical state transition mechanism is introduced, which dynamically adjusts the state transition path by combining boundary information and incorporates prior knowledge of technical entities, thereby achieving accurate identification of nested entities. The specific process will be detailed later.

[0028] The entity recognition method for scientific and technological achievement application texts provided in this application can improve the accuracy and precision of entity recognition in scientific and technological achievement application texts, thereby ensuring the accuracy of knowledge graph construction.

[0029] Step S103 above involves using a compound nesting boundary detection module to perform compound nesting boundary detection on the text and dynamically segment the semantic units corresponding to the text, including: (1) Obtain the dependency syntax tree depth value of each word position in the text, and calculate the weighted dependency depth change rate of each word position based on the dependency syntax tree depth value of each word position and the Gaussian weighted sliding window. In practice, the weighted dependency depth change rate is calculated based on a Gaussian weighted sliding window using the dependency syntax tree depth value at each word position, thereby capturing the degree of abrupt changes in the grammatical structure and realizing the detection of the starting boundaries of technical terms, as shown below: ; In the formula, The weighted dependency depth change rate at position t is used to quantify the degree of mutation in the grammatical structure and reflect the starting boundary of technical terms. Represents the natural constant; This is the position index of a word in the text, with a value range of 100. ,in Indicates the total number of words in the text; This represents the dependency syntax tree depth value at position t, obtained through parsing using the StanfordCoreNLP tool; The depth of the dependency syntax tree at position t+k is represented by the value obtained by parsing using the StanfordCoreNLP tool; It represents the t+kth digit. The dependency syntax tree depth value at position 1 was obtained by parsing using the StanfordCoreNLP tool; The Gaussian kernel width parameter controls the local window weight distribution, with a preferred value of 2.0; This is the offset of the sliding window, and its value range is... A context window covering 7 words.

[0030] (2) Calculate the composite boundary probability score for each word position based on the prefix word vector, the benchmark vector of the technical term category, and the Bi-gram probability trained on the corpus of scientific and technological achievements. The composite boundary probability score is used to evaluate the possibility of the position as a term boundary. In practice, based on the co-occurrence probability of Bi-gram terms and the similarity of prefix word vectors, the contributions of the two are balanced by a fusion coefficient, and a composite boundary probability score is calculated to evaluate the likelihood of each position serving as a term boundary, expressed as: ; In the formula, The composite boundary probability score at position t is used to assess the likelihood of that position serving as a term boundary. The fusion coefficient is used to balance the contributions of statistical probability and semantic similarity, and a value of 0.6 is preferred. This represents the Bi-gram probability trained based on a corpus of scientific and technological achievements, reflecting the statistical characteristics of word co-occurrence. This represents the conditional probability symbol, i.e., given the preceding word. Current word The probability of occurrence; The word representing the t-th position; Indicates the t-th Location word; This represents the prefix word vector at position t, composed of words to The mean vector is used to encode the prefix context semantics; The baseline vector representing the category of technical terms is determined by k-means cluster centers and represents the typical semantic features of technical terms; This represents the cosine similarity function, used to measure the semantic similarity between vectors.

[0031] (3) Based on the weighted dependency depth change rate and composite boundary probability score of each word position, a set of semantic unit boundary positions is generated by setting dependency depth threshold and boundary probability threshold and combining modulo operation to avoid fixed interval segmentation; the set of semantic unit boundary positions is used to divide the semantic units corresponding to the text.

[0032] In practice, based on the weighted dependency depth change rate and composite boundary probability score, a set of semantic unit boundary positions is generated by setting dependency depth thresholds and boundary probability thresholds, and by combining modulo operations to avoid fixed-interval segmentation, as follows: ; In the formula, This is the set of semantic unit boundary locations, containing all segmentation points that satisfy the conditions; Represents the set construction symbol, where Representative elements, This indicates that the condition is "satisfied". This represents the logical AND operation, i.e., sets. It consists of all position indices that simultaneously satisfy all three conditions; The dependency depth threshold is preferably set to 0.310, which is used to filter syntax mutation points. The boundary probability threshold is preferably set to 0.7, which is used to filter high-probability boundaries. To avoid forced splitting every 3 words in the modulo operation, the segmentation flexibility is ensured.

[0033] Step S104 above, which involves extracting the bidirectional dependency graph convolutional features of each semantic unit using the bidirectional dependency graph convolutional module, includes: (1) Determine the word vector and relation vector corresponding to each semantic unit; the relation vector is used to represent the original dependency relation type of the word vector; calculate the semantic weight of the dependency edge between word vectors based on the word vector and relation vector of each semantic unit; In practice, the semantic weights of dependency edges are calculated based on the similarity between word vectors and relation vectors. By fusing the bidirectional similarity between words and relations, the semantic strength of grammatical relations is quantified, as follows: ; In the formula, This represents the semantic weight of the dependency edge from the i-th word to the j-th word, used to quantify the semantic strength of the grammatical relation; This represents the word vector of the i-th word, encoding the semantic information of the i-th word; This represents the word vector of the j-th word, encoding the semantic information of the j-th word; Representing relation type The corresponding embedding vector encodes the semantics of dependency relationships.

[0034] (2) Construct dependency edges between word vectors based on the semantic weights of dependency edges and the types of original dependency relations; In practical implementation, dependency edges containing relation types and semantic weights are constructed using the original dependency relation types and semantic weights, thereby enhancing the dependency graph structure, as shown below: ; In the formula, Represents the dependency edge from the i-th word to the j-th word, including the relation type and semantic weight, used to construct the graph structure; For primitive dependency relation types, such as nsubj representing a noun subject and dobj representing a direct object, define grammatical roles; The semantic weights for dependency relationships are calculated using the similarity between word vectors and relation vectors, thus enhancing the semantic information of the edges.

[0035] (3) Using multi-hop graph convolution aggregation technique to perform feature propagation and linear transformation on the dependency edges between word vectors, and applying ReLU activation function to capture long-distance dependencies, we obtain bidirectional dependency graph convolution features for each semantic unit.

[0036] In practice, multi-hop graph convolution aggregation technology is used to aggregate attribute features. Feature propagation and linear transformation are performed using a multi-hop neighbor node set and dependency edge weights. The ReLU activation function is then applied to capture long-distance dependencies, as shown below: ; In the formula, This represents the feature vector of the i-th word at the l-th layer, encoding multi-hop syntax and technical attribute information; Let k represent the set of k-hop neighbors of the i-th word, where k takes the value of 2, and expand the receptive field to capture long-distance dependencies. This represents the dependency edge from the j-th word to the i-th word, which is used as the edge weight in feature aggregation; This represents a specific weight matrix for the l-th layer, used for linear transformation, and consists of trainable parameters. This represents the feature vector of the j-th word in the (l-1)-th layer; Represents the j-th word The set of hop neighbor nodes, where A value of 2 is used to expand the receptive field to capture long-range dependencies; This represents the number of k-hop neighbors of the i-th word, used for normalization; This represents the number of k-hop neighbor nodes of the j-th word, used for normalization; This represents the linear rectification activation function, which introduces a nonlinear transformation.

[0037] Step S105 above, which involves capturing the multi-head dynamic local features of each semantic unit based on the bidirectional dependency graph convolution features of each semantic unit through the multi-head dynamic convolution gating module, includes: For each semantic unit, treat it as the current semantic unit and perform the following steps: (1) Generate the dynamic convolution kernel parameters of the current semantic unit based on the bidirectional dependency graph convolution features of the current semantic unit's context, as well as the gating transformation matrix and projection weight matrix; In practice, based on the current word context features, dynamic convolution kernel parameters are generated through the gating transformation matrix and projection weight matrix to achieve adaptive adjustment of the convolution kernel size, as shown below: ; In the formula, This represents the dynamic convolution kernel parameters at position t, which are adaptively generated based on the context. Represents the gated transformation matrix, with dimension . , used for feature dimension transformation; This represents the projection weight matrix, which maps the input features to the convolution kernel parameter space; This represents the feature vector of the [t-1, t+1]-th window in the second layer; This represents the hyperbolic tangent activation function, which controls the output range.

[0038] (2) Based on the dynamic convolution kernel parameters of the current semantic unit, dynamic convolution is performed in parallel by multiple attention heads to capture the multi-head dynamic local features corresponding to the current semantic unit.

[0039] In practice, dynamic convolution is performed in parallel by multiple attention heads. Based on the local convolution window radius and dynamically generated convolution kernel parameters, local combined patterns are captured from different feature subspaces, as represented by: ; In the formula, This represents the convolutional output feature at position t with the m-th attention head, capturing local patterns in a specific subspace; This represents the attention head index, with a value range of 1. ,in To achieve multi-view feature extraction, focus on the total number of attention heads; This represents the radius of the local convolution window, preferably 2, and defines the context of the convolution operation. Indicates the relative position offset within the window, with a value range of [value missing]. ; This represents the convolution kernel parameters dynamically generated at position t by the m-th attention head, used for feature transformation; This represents the feature vector of the (t+δ)th word in the second layer.

[0040] Step S106 above, which involves using a gated hierarchical state transition and entity recognition module to perform state transition based on the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit, generating state vectors word by word and decoding entity labels to output the structured entity sequence corresponding to the scientific and technological achievement application text, includes: (1) Determine the input feature vector based on the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit; based on the input feature vector, and combined with the weighted dependency depth change rate and composite boundary probability score of each semantic unit, obtain the boundary-enhanced forget gate of each semantic unit; In practical implementation, the forgetting gate in the standard LSTM is enhanced by using the weighted dependency depth change rate and composite boundary probability score. This allows the model to dynamically control the retention and forgetting of state information based on syntax mutation and boundary probability. Furthermore, by combining prior knowledge of technical entity distribution, the sensitivity to technical terms is improved. This can be expressed as follows: ; In the formula, Indicates the first A position-bounded enhanced forget gate is used to control the previous state. The degree of retention is indicated by a value closer to 1, which means more historical information is retained, and a value closer to 0, which means more information is forgotten. This represents the forget gate weight matrix, with dimension 1. ,in For the state dimension, The input feature dimension is used for linear transformation to fuse multi-source information; Indicates the first The position's state vector encodes the context information and entity state of the previous position, including the initial state. It is a zero vector; Indicates the first The input feature vector of the location is defined as follows: It integrates bidirectional dependency graph convolution features and multi-head dynamic convolution features to provide rich syntactic and semantic information; This represents the forget gate bias vector. These are trainable parameters used to adjust the baseline value of the forget gate. This represents the Sigmoid activation function; This represents the technical prior strength coefficient, with a preferred value of 0.1, which controls the degree of influence of technical entity priors on the forget gate; Table 1 The prior probabilities of each technical entity category are obtained statistically from a corpus of scientific and technological achievements and are used to adjust the forget gate bias.

[0041] (2) Based on the input feature vector, the input gate weight matrix and the bias vector, the input gate of each semantic unit is obtained; based on the boundary enhancement forget gate and the input gate of each semantic unit, the candidate state is modulated by combining the technical entity category vector and the boundary position set, and the state vector of each semantic unit is determined by the state transition equation. Based on the boundary-enhanced forget gate and input gate, the current state is updated through the state transition equation, and the candidate state is modulated by the technical entity category vector and the boundary position set, thereby improving the semantic representation and boundary consistency of technical terms, as shown below: ; In the formula, Indicates the first The state vector of the location encodes the context information and entity state of the current location; Indicates the first The position input gate is calculated as follows: This is used to control the degree of updating of candidate states; This represents the input gate weight matrix, with dimension 1. , are trainable parameters used for linear transformation; This represents the input gate bias vector, which is a trainable parameter used to adjust the baseline value of the input gate. This represents the candidate state weight matrix, with dimension 1. , are trainable parameters used for linear transformation to generate candidate states; This represents the candidate state bias vector, which is a trainable parameter used to adjust the candidate state baseline value. The vector representing the category of technical entities is calculated by weighted average of all entity word vectors in the corpus of scientific and technological achievements. The weights are determined by the entity frequency and are used to enhance the semantic representation of technology. This represents the boundary modulation coefficient, preferably with a value of 0.10, which controls the influence of the boundary indicator function on the state update. Indicates an indicator function, when Belongs to the set of boundary locations The value is 1 when the condition is met and 0 otherwise, which is used to enhance the state update of boundary points; This represents element-wise multiplication and is used for gating and regulating the information flow. This represents the hyperbolic tangent activation function.

[0042] (3) Based on the state vector of each semantic unit and the state vector in the corresponding local context window, multi-level information is aggregated through a multi-head attention mechanism to determine the context-aware state vector of each semantic unit; Based on the current state and the state within the local context window, multi-level information is aggregated through a multi-head attention mechanism, and the entity label probability distribution is output. Simultaneously, technical entity co-occurrence features are incorporated to improve decoding accuracy, represented as: ; In the formula, Indicates the first The context-aware state vector of the location aggregates local context information for entity label prediction, enhancing the hierarchical representation; This represents the radius of the local context window, with a preferred value of 3, which defines the size of the context window and balances local and global information. This represents the number of attention heads, with a preferred value of 4, used for multi-view information aggregation and capturing features from different subspaces. Indicates the first The first one in the attention. Position state to the first The attention weight of a location is calculated as follows: , used for weighted aggregation of related states; Indicates the first The attention score for each attention head is calculated as follows: ,in For the first The weight matrix of each attention head. For the first The weight vector of each attention head. For the co-occurrence feature vector of technical entities, through the first... The entity and the first The co-occurrence frequency of each entity is encoded.

[0043] (4) Output the entity label probability distribution based on the context-aware state vector of each semantic unit.

[0044] Furthermore, the entity label probability is calculated as follows: ,in, This indicates that the output weight matrix is ​​a trainable parameter. This represents the output bias vector, which are trainable parameters. This represents the softmax activation function. This means that given the entire input sequence Under the condition that, the entity label at position t The probability distribution is the probability value that the model predicts for the t-th word to belong to each entity category.

[0045] The training process of the above entity recognition model is described below: First, obtain the training dataset. The specific process is as follows: Systematically collect application materials for scientific and technological achievements, including scientific and technological project application forms, technical achievement appraisal reports, patent specifications, award certificates, and related technical documents.

[0046] Then, an initial corpus is built by crawling public or open-source materials. During the training dataset construction phase, the original text is cleaned and standardized, including removing irrelevant tables and images, unifying the format of numbers and units, and correcting typos and grammatical errors. Then, based on the structural features of the scientific and technological achievement application text (such as technical background, innovation points, implementation methods, effect verification, etc.), a team of domain experts annotates the data. The annotation categories strictly correspond to the subsequent entity recognition and relation extraction requirements, mainly including five categories: technical terms (such as algorithm models, hardware equipment), technical indicators (such as accuracy, efficiency, power consumption), innovation attributes (such as originality, breakthrough), application scenarios (such as equipment diagnosis, intelligent buildings), and research and development entities (such as institutions, teams).

[0047] The annotation process employs a dual-person independent annotation and cross-validation mechanism, with discrepancies in samples arbitrated by senior experts, ultimately forming a high-quality annotated dataset that provides a training foundation for subsequent composite nested boundary detection and entity recognition models.

[0048] In the recognition of entities in scientific and technological achievements, syntactic structure and semantic information are often intertwined in feature representations, causing models to confuse the boundaries and categories of technical entities. Conventional cross-entropy loss cannot explicitly decouple these feature components. At the same time, the imbalanced distribution of technical entity categories exacerbates the difficulty of model training.

[0049] This embodiment improves the model's ability to distinguish technical entities by designing a syntax-semantic feature decoupling loss and a category-balanced focus loss. Specifically, during the training process of the entity recognition model, the total model loss is determined by using a weighted sum of the syntax-semantic feature decoupling loss and the category-balanced focus loss. The syntax-semantic feature decoupling loss is determined as follows: based on the bidirectional dependency graph convolutional features and multi-head dynamic convolutional features of each semantic unit, the orthogonality loss between syntax features and semantic features is calculated, and combined with the covariance matrix penalty term, the syntax-semantic feature decoupling loss is determined. In practice, based on dependency graph convolution features and multi-head dynamic convolution features, the orthogonal loss between syntactic and semantic features is calculated, and the model is forced to learn decoupled feature representations through a covariance matrix penalty term, as follows: ; In the formula, This represents the syntax-semantic feature decoupling loss, which is used to encourage syntax and semantic features to be independent of each other and reduce confusion caused by feature interleaving; Indicates the total number of words in the text; Indicates the first The syntactic feature vector of position is defined as follows: Features derived from bidirectional dependency graph convolutions encode syntactic structure information; Indicates the first The semantic feature vector of location is defined as follows: Features derived from multi-head dynamic convolution encode semantic information; Represents the L2 norm; This represents the covariance penalty coefficient, with a preferred value of 0.01, which controls the weight of the covariance term in the loss. The grammatical feature matrix consists of all Stacked row by row, with a dimension of ; The semantic feature matrix is ​​composed of all Stacked row by row, with a dimension of ; This represents the Frobenius norm.

[0050] The category-balanced focus loss is determined by adjusting the cross-entropy loss based on the category prediction probability, category weight, focus parameter, prior adjustment coefficient, and category technical importance weight for each semantic unit.

[0051] To address the imbalance in entity categories in scientific and technological achievements, class weights and focusing parameters are introduced, and the cross-entropy loss is adjusted to make the model focus on hard-to-classify samples. Furthermore, prior knowledge of the distribution of technical entities is incorporated to improve the ability to identify rare entities, as expressed below: ; In the formula, This represents the class-balanced focus loss, used to handle class imbalance and hard sample learning, thereby improving model robustness; Indicates the total number of entity categories; This represents the weight of the c-th category, calculated as follows: ,in The total number of samples, Let c be the number of samples in the c-th category; Indicates the first The predicted probability that a location belongs to the c-th category is based on get; This represents the focusing parameter, with a preferred value of 2, which is used to reduce the loss contribution of easily classified samples and make the model pay more attention to difficult-to-classify samples. This represents the prior adjustment coefficient, with a preferred value of 0.2, which controls the influence of the technology importance weight. The weight representing the technological importance of the c-th category is calculated based on expert annotations in the field of scientific and technological achievements and is used to enhance the loss contribution of key entity categories.

[0052] Furthermore, the final total loss is ,in This is the decoupling loss weight, used to balance the entity recognition loss and the decoupling loss; a value of 0.10 is preferred. This is the total loss function used for model training.

[0053] Based on the aforementioned modular architecture and loss function, the entity recognition model is trained end-to-end. During training, the constructed labeled dataset is divided into training, validation, and test sets in an 8:1:1 ratio. Model parameters are initialized using the Xavier method, with the AdamW optimizer selected. The initial learning rate is set to 0.001 and dynamically adjusted using a cosine annealing strategy.

[0054] Each iteration computes the bidirectional dependency graph convolutional features, multi-head dynamic convolutional features, and gated hierarchical state transition results via forward propagation, and then uses the syntax-semantic feature decoupling loss defined above. Balanced focus loss with category Calculate total loss The gradient is calculated using the backpropagation algorithm, and gradient clipping (threshold 1.0) is used to prevent gradient explosion, and all trainable parameters are updated.

[0055] Training termination conditions are dynamically determined based on validation set performance: if the entity recognition F1 score on the validation set does not improve for 10 consecutive iterations, or the total loss value decreases by less than 0.001 for 10 consecutive iterations, an early termination mechanism is triggered. At the same time, the optimal model parameters are recorded to ensure that the model achieves stable recognition performance in scenarios such as syntactic structural mutations, nested entities, and long-distance referencing.

[0056] In one embodiment, the comprehensive performance of different entity recognition methods in scientific and technological achievement application texts is evaluated. Four entity recognition methods are compared experimentally, including the conventional bidirectional long short-term memory network conditional random field model, the bidirectional encoder representation transformation based on the pre-trained language model combined with the bidirectional long short-term memory network conditional random field model, the rule-based method, and the method proposed in this embodiment.

[0057] The conventional bidirectional long short-term memory (LSTM) network conditional random field (CRRF) model is a classic deep learning method for sequence labeling tasks. The combination of a pre-trained language model with a bidirectional encoder representation transformation and a CRRF network represents the current state-of-the-art pre-trained language model approach. Rule-based methods, on the other hand, rely on rules manually defined by domain experts for entity recognition. In terms of experimental configuration, all methods were trained and evaluated on the same training, validation, and test sets, using the same entity labeling specifications. The evaluation metric was accuracy, i.e., the proportion of correctly identified entities out of the total number of entities. Figure 2 The experimental results show that the method in this embodiment exhibits superior performance across all entity categories. In terms of technical terminology recognition, the composite nested boundary detection mechanism employed in this embodiment effectively handles complex nested entities, resulting in significantly higher accuracy compared to other methods. Regarding technical indicator recognition, this embodiment captures the dependency relationship between technical indicators and numerical values ​​through a bidirectional dependency graph convolutional feature extraction module. For entity categories such as innovation attributes, application scenarios, and R&D entities, the gated hierarchical state transition mechanism in this embodiment combines boundary information and prior entity knowledge, significantly improving recognition accuracy.

[0058] In one embodiment, the performance of different methods in handling nested entities is analyzed. Nested entities are a common phenomenon in scientific and technological achievement application texts; for example, "quantum computing chip" contains multiple layers of entities, including both "quantum" and "computing chip." The experiment set up four nesting levels, ranging from simple single-layer entities to complex four-layer and higher nested entities. The evaluation metric used was the F1 score, a comprehensive evaluation index for entity recognition, which considers both precision and recall. Figure 3 As shown in the experimental results, the performance of all methods declines with increasing nesting levels, but the method in this embodiment shows the smallest decline. In single-layer entity recognition, the performance differences between methods are relatively small. However, when processing double-layer nested entities, the performance of the conventional bidirectional long short-term memory network conditional random field model and the bidirectional encoder representation transformation based on a pre-trained language model combined with the bidirectional long short-term memory network conditional random field model begins to decline significantly. The method in this embodiment, through a gated hierarchical state transition mechanism, effectively maintains high recognition performance. When processing three or more layers of nested entities, the advantages of the method in this embodiment are even more significant. This is due to its hierarchical entity decoding module's ability to aggregate multi-level information and improve decoding accuracy by combining technical entity co-occurrence features. In contrast, conventional sequence labeling methods perform poorly in complex nested entity recognition tasks because they struggle to effectively capture hierarchical structures.

[0059] In another embodiment, the performance of the method in entity boundary detection is analyzed in detail using heatmaps. Boundary detection is a crucial step in entity recognition, directly impacting its accuracy. Experiments examined four boundary types: start boundary, end boundary, internal boundary, and external boundary, covering all boundary locations in entity recognition. Furthermore, the performance of five entity categories in boundary detection was analyzed, with boundary detection accuracy serving as the evaluation metric. See [link to relevant documentation] Figure 4 As shown, the color intensity in the heatmap intuitively reflects the accuracy level of boundary detection; the darker the color, the higher the accuracy. Overall, the method in this embodiment maintains a high accuracy rate across various boundary detection tasks. Particularly noteworthy is the highest accuracy achieved for all entity categories in external boundary detection, indicating that the method effectively distinguishes between entity and non-entity regions, reducing false identification. The method performs exceptionally well in boundary detection for technical terms and technical indicators, thanks to the terminology boundary probability fusion mechanism, which accurately identifies the boundary positions of technical entities by combining statistical probability with semantic similarity. For innovative attribute entities, boundary detection is slightly more challenging due to their diverse expressions, but the method in this embodiment still maintains good performance.

[0060] In another embodiment, the impact of text complexity on entity recognition performance is analyzed. The complexity of scientific and technological achievement application texts varies greatly, ranging from simple technical descriptions to complex technical arguments. See [link to relevant documentation]. Figure 5 As shown, the experiment categorized texts into four complexity levels based on text length and structural complexity, ranging from simple to extremely complex texts. The average text length for each complexity level was recorded in terms of word count. The performance of the four methods was compared, with the F1 score (a comprehensive evaluation metric for entity recognition) used as the evaluation metric. The results show that the performance of all methods decreases with increasing text complexity, but the method in this embodiment exhibits the most gradual decline. For simple and moderately complex texts, the performance differences between the methods are relatively small. However, when processing complex texts, the method in this embodiment effectively captures long-distance dependencies and local combination features through bidirectional dependency graph convolution and multi-head dynamic convolution gating modules, thus minimizing performance degradation. Even when processing extremely complex texts, the method in this embodiment maintains a relatively high performance level, demonstrating its robustness. In contrast, rule-based methods perform reasonably well in simple texts, but experience the most significant performance degradation in complex texts, indicating that rule-based methods are ill-suited to complex linguistic phenomena. Conventional bidirectional long short-term memory network conditional random field models and bidirectional encoder representation transformation based on pre-trained language models combined with bidirectional long short-term memory network conditional random field models also exhibit significant performance degradation in complex text processing.

[0061] Based on the above entity recognition method embodiments, this application embodiment also provides a knowledge graph construction method for scientific and technological achievement application, see [link to relevant documentation]. Figure 6 As shown, the method includes: Step S601: Output the structured entity sequence corresponding to the text of the scientific and technological achievement application using the entity recognition method for scientific and technological achievement application. The specific implementation process of the entity recognition method for scientific and technological achievement application is described in the aforementioned method embodiment.

[0062] In simple terms, the preprocessed text of the scientific and technological achievement application is input into the trained entity recognition model. Specifically, firstly, composite nested boundary detection is performed: semantic units are dynamically divided through dependency depth mutation detection and term boundary probability fusion (e.g., "image recognition system based on deep convolutional neural network" is treated as a complete unit). Then, each semantic unit is sequentially processed by a bidirectional dependency graph convolution module to extract syntactic association features, and then a multi-head dynamic convolutional gating module is used to capture local combination patterns. Finally, based on a gating hierarchical state transition mechanism, the model combines the boundary position set and prior knowledge of entity categories to generate state vectors word by word and decode entity labels. The final output is a structured entity sequence, for example: Technical terms: quantum computing chip, deep convolutional neural network; Technical specifications: Recognition accuracy 98.2%, power consumption ≤10W; Innovation attributes: International first, technological breakthrough; Application scenarios: image analysis, autonomous driving; Research and development entity: XX Company Artificial Intelligence Laboratory.

[0063] Step S602: Through dependency path enhancement and multi-hop graph convolutional aggregation, the relationship is extracted from the text of scientific and technological achievement application, and the relationship information between entities is established. In practice, relation extraction is completed through the following steps: (1) Perform compound nesting boundary detection on the text of scientific and technological achievement application, dynamically divide the semantic units corresponding to the text of scientific and technological achievement application; extract the syntactic path connecting the semantic units to obtain the dependency syntactic path connecting the entity pairs; such as "precision-[nsubj]-system".

[0064] (2) Through the bidirectional dependency graph convolution feature extraction module, multi-hop feature propagation is performed on the nodes on the path between entities to integrate syntactic relations and technical attribute features to obtain bidirectional dependency graph convolution features between entity pairs; such as the measurement relationship between "98.2%" and "precision".

[0065] (3) Input the bidirectional dependency graph convolution features of entity pairs and their contexts into the multilayer perceptron to predict the predefined relationship categories and obtain the inter-entity association information.

[0066] For example, inputting entity pairs and their contextual features (including dependency path encoding and relative position embedding) into a multilayer perceptron can predict predefined relationship categories, mainly including: 1> Technical indicator association (e.g., "Image recognition system - has indicators - recognition accuracy 98.2%)); 2> Technological composition relationship (e.g., "quantum computing chip - components - qubit array"); 3> Application scenario adaptation (e.g., "Autonomous driving system - applied to - logistics and transportation"); 4> Attribution of innovation attributes (e.g., "Multimodal fusion algorithm - belongs to - original technology"); 5> Contributions from the main R&D entity (e.g., “XX Laboratory - R&D - Neural Network Accelerator”).

[0067] Step S603: Construct a knowledge graph for scientific and technological achievement application based on structured entity sequences and inter-entity association information.

[0068] (1) Perform entity alignment and disambiguation on the structured entity sequence to identify multiple target entities; Synonymous entities (such as "CNN" and "convolutional neural network") are normalized, and ambiguous entities (such as "apple" as a company name and fruit) are distinguished based on context.

[0069] (2) Based on the inter-entity association information, the entity relationship is transformed to obtain triples; Transform entity-relation pairs into triples of <head entity, relation, tail entity> (e.g., <quantum computing chip, technology composition, qubit array>).

[0070] (3) Based on a predefined graph pattern, triples are stored in a graph database to obtain a knowledge graph for scientific and technological achievement applications, such as... Figure 7 As shown.

[0071] The ontology architecture is designed based on the characteristics of scientific and technological achievements. Core classes include TechnologyTerm, TechnicalIndicator, and InnovationType, with relational attributes inheriting the previously defined classification system. The Neo4j graph database is used to store triples, and entity descriptions are enriched through node attributes (such as the numerical range of technical indicators and the level of innovation attributes). Finally, an interactive knowledge graph is generated, supporting queries and visualizations based on technical fields, innovation dimensions, and other criteria, providing structured knowledge support for the evaluation of scientific and technological achievements and the analysis of technical routes.

[0072] Based on the above-described entity recognition method embodiments, this application also provides an entity recognition device for technology achievement applications, see [link to relevant documentation]. Figure 8 As shown, the device includes: a text acquisition module 81, used to acquire the text of a scientific and technological achievement application to be identified; a preprocessing module 82, used to preprocess the text of the scientific and technological achievement application and input the preprocessed text into a pre-trained entity recognition model; the preprocessing includes: data cleaning and standardization; the entity recognition model includes: a composite nested boundary detection module, a bidirectional dependency graph convolution module, a multi-head dynamic convolution gating module, a gating hierarchical state transition module, and an entity recognition module; and a model recognition module 83, used to perform composite nested boundary detection on the text through the composite nested boundary detection module, dynamically... The system divides the text into semantic units; extracts the bidirectional dependency graph convolution features of each semantic unit through a bidirectional dependency graph convolution module; captures the multi-head dynamic local features of each semantic unit based on its bidirectional dependency graph convolution features through a multi-head dynamic convolution gating module; and performs state transitions based on a gating hierarchical state transition mechanism according to the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit through a gating hierarchical state transition and entity recognition module, generating state vectors word by word and decoding entity labels to output the structured entity sequence corresponding to the scientific and technological achievement application text.

[0073] Furthermore, the aforementioned model recognition module 83 is also used to obtain the dependency syntax tree depth value of each word position in the text, and calculate the weighted dependency depth change rate of each word position based on the dependency syntax tree depth value of each word position and the Gaussian weighted sliding window; calculate the composite boundary probability score of each word position based on the obtained prefix word vector, the benchmark vector of the technical term category, and the Bi-gram probability trained based on the scientific and technological achievement corpus; the composite boundary probability score is used to evaluate the possibility of the position as a term boundary; based on the weighted dependency depth change rate and composite boundary probability score of each word position, by setting the dependency depth threshold and the boundary probability threshold, and combining the modulo operation to avoid fixed interval segmentation, a set of semantic unit boundary positions is generated; the set of semantic unit boundary positions is used to divide the semantic units corresponding to the text.

[0074] Furthermore, the aforementioned model recognition module 83 is also used to determine the word vector and relation vector corresponding to each semantic unit; the relation vector is used to characterize the original dependency relation type of the word vector; based on the word vector and relation vector of each semantic unit, the semantic weight of the dependency edge between word vectors is calculated; based on the semantic weight of the dependency edge between word vectors and the original dependency relation type, the dependency edge between word vectors is constructed; using multi-hop graph convolution aggregation technology to aggregate attribute features, the dependency edge between word vectors is subjected to feature propagation and linear transformation, and the ReLU activation function is applied to capture long-distance dependencies, thereby obtaining the bidirectional dependency graph convolution feature of each semantic unit.

[0075] Furthermore, the aforementioned model recognition module 83 is also used to treat each semantic unit as the current semantic unit and perform the following steps: generate dynamic convolution kernel parameters of the current semantic unit based on the bidirectional dependency graph convolution features of the current semantic unit's context, as well as the gating transformation matrix and projection weight matrix; and capture the multi-head dynamic local features corresponding to the current semantic unit by performing dynamic convolution in parallel through multiple attention heads based on the dynamic convolution kernel parameters of the current semantic unit.

[0076] Furthermore, the aforementioned model recognition module 83 is also used to determine the input feature vector based on the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit; based on the input feature vector, and combined with the weighted dependency depth change rate and composite boundary probability score of each semantic unit, to obtain the boundary-enhanced forget gate of each semantic unit; based on the input feature vector, the input gate weight matrix, and the bias vector, to obtain the input gate of each semantic unit; based on the boundary-enhanced forget gate and the input gate of each semantic unit, combined with the technical entity category vector and the boundary position set to modulate the candidate state, to determine the state vector of each semantic unit through the state transition equation; based on the state vector of each semantic unit and the state vector within the corresponding local context window, to aggregate multi-level information through a multi-head attention mechanism to determine the context-aware state vector of each semantic unit; and based on the context-aware state vector of each semantic unit, to output the entity label probability distribution.

[0077] Furthermore, the aforementioned device also includes a model training module, used to determine the total model loss during the training process of the entity recognition model by using a weighted sum of the syntax-semantic feature decoupling loss and the category balancing focus loss. The syntax-semantic feature decoupling loss is determined as follows: based on the bidirectional dependency graph convolutional features and multi-head dynamic convolutional features of each semantic unit, the orthogonality loss between syntax features and semantic features is calculated, and combined with the covariance matrix penalty term to determine the syntax-semantic feature decoupling loss. The category balancing focus loss is determined as follows: based on the category prediction probability, category weight, focus parameter, prior adjustment coefficient, and technical importance weight of each semantic unit, the cross-entropy loss is adjusted to obtain the category balancing focus loss.

[0078] Based on the above embodiments of the knowledge graph construction method, this application also provides a knowledge graph construction device for scientific and technological achievement applications. See [link to relevant documentation]. Figure 9 As shown, the device includes: an entity recognition device 91 for scientific and technological achievement applications, which outputs a structured entity sequence corresponding to the text of the scientific and technological achievement application; the specific structure of the entity recognition device 91 for scientific and technological achievement applications is as follows. Figure 8 As shown. The relation extraction module 92 is used to extract relations from the text of scientific and technological achievement application through dependency path enhancement and multi-hop graph convolution aggregation, and to establish the relationship information between entities; the graph construction module 93 is used to construct a knowledge graph of scientific and technological achievement application based on the structured entity sequence and the relationship information between entities.

[0079] Furthermore, the aforementioned relation extraction module 92 is used to perform compound nesting boundary detection on the technical achievement application text, dynamically divide the semantic units corresponding to the technical achievement application text; extract the syntactic paths connecting the semantic units to obtain the dependency syntactic paths connecting entity pairs; through the bidirectional dependency graph convolution feature extraction module, perform multi-hop feature propagation on the nodes on the path between entities to fuse syntactic relations and technical attribute features to obtain the bidirectional dependency graph convolution features between entity pairs; input the bidirectional dependency graph convolution features of the entity pairs and their contexts into a multilayer perceptron to predict the predefined relation categories and obtain the association information between entities.

[0080] Furthermore, the aforementioned graph construction module 93 is used to perform entity alignment and disambiguation on the structured entity sequence to identify multiple target entities; based on the inter-entity association information, it performs entity relationship transformation to obtain triples; based on the predefined graph pattern, it uses a graph database to store the triples to obtain a knowledge graph for scientific and technological achievement applications.

[0081] The device provided in this application embodiment has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts of the device embodiment not mentioned can be referred to the corresponding content in the aforementioned method embodiment.

[0082] This application also provides an electronic device, such as... Figure 10 The diagram shown is a structural schematic of the electronic device, which includes a processor 101 and a memory 100. The memory 100 stores computer-executable instructions that can be executed by the processor 101, which executes the computer-executable instructions to implement the aforementioned method. Figure 10 In the illustrated embodiment, the electronic device further includes a bus 102 and a communication interface 103, wherein the processor 101, the communication interface 103, and the memory 100 are connected via the bus 102.

[0083] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are called and executed by a processor, the computer-executable instructions cause the processor to implement the above-described method. For specific implementation details, please refer to the foregoing method embodiments, which will not be repeated here.

[0084] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the technical scope disclosed in this application. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for identifying entities in a science and technology achievement declaration, characterized in that, The method includes: Obtain the application text for scientific and technological achievements to be identified; The text submitted for scientific and technological achievement application is preprocessed, and the preprocessed text is input into a pre-trained entity recognition model. The preprocessing includes data cleaning and standardization. The entity recognition model includes a composite nested boundary detection module, a bidirectional dependency graph convolution module, a multi-head dynamic convolution gating module, a gating hierarchical state transition module, and an entity recognition module. The composite nesting boundary detection module is used to detect the composite nesting boundary of the text and dynamically divide the semantic units corresponding to the text. The bidirectional dependency graph convolution module is used to extract the bidirectional dependency graph convolution features of each semantic unit. The multi-head dynamic convolution gating module captures the multi-head dynamic local features of each semantic unit based on the bidirectional dependency graph convolution features of each semantic unit. Through the gated hierarchical state transition and entity recognition module, state transition based on the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit is performed. State vectors are generated word by word and entity labels are decoded to output the structured entity sequence corresponding to the scientific and technological achievement application text.

2. The method according to claim 1, characterized in that, The steps of performing compound nesting boundary detection on the text and dynamically dividing the text into semantic units using the compound nesting boundary detection module include: Obtain the dependency syntax tree depth value for each word position in the text, and calculate the weighted dependency depth change rate for each word position based on the dependency syntax tree depth value for each word position and the Gaussian weighted sliding window. Based on the prefix word vector, the benchmark vector of the technical term category, and the Bi-gram probability trained on the corpus of scientific and technological achievements for each word position, a composite boundary probability score is calculated for each word position; the composite boundary probability score is used to evaluate the likelihood of the position as a term boundary. Based on the weighted dependency depth change rate and the composite boundary probability score for each word position, a set of semantic unit boundary positions is generated by setting dependency depth thresholds and boundary probability thresholds and combining modulo operations to avoid fixed-interval segmentation; the set of semantic unit boundary positions is used to divide the semantic units corresponding to the text.

3. The method according to claim 1, characterized in that, The step of extracting the bidirectional dependency graph convolutional features of each semantic unit through the bidirectional dependency graph convolutional module includes: Determine the word vector and relation vector corresponding to each semantic unit; the relation vector is used to characterize the original dependency relation type of the word vector; calculate the semantic weight of the dependency edge between word vectors based on the word vector and relation vector of each semantic unit; Based on the semantic weights of the dependency edges between the word vectors and the types of the original dependency relations, construct the dependency edges between the word vectors; The multi-hop graph convolution aggregation technique is used to perform feature propagation and linear transformation on the dependency edges between the word vectors, and the ReLU activation function is applied to capture long-distance dependencies, so as to obtain the bidirectional dependency graph convolution features of each semantic unit.

4. The method according to claim 1, characterized in that, The step of capturing the multi-head dynamic local features of each semantic unit based on the bidirectional dependency graph convolution features of each semantic unit through the multi-head dynamic convolution gating module includes: For each semantic unit, treat it as the current semantic unit and perform the following steps: Based on the bidirectional dependency graph convolution features of the current semantic unit's context, as well as the gating transformation matrix and projection weight matrix, the dynamic convolution kernel parameters of the current semantic unit are generated. Based on the dynamic convolution kernel parameters of the current semantic unit, dynamic convolution is performed in parallel by multiple attention heads to capture the multi-head dynamic local features corresponding to the current semantic unit.

5. The method according to claim 1, characterized in that, The steps of using the gated hierarchical state transition and entity recognition module to perform state transition based on the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit, generating state vectors word by word and decoding entity labels, and outputting the structured entity sequence corresponding to the scientific and technological achievement application text include: The input feature vector is determined based on the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit; Based on the input feature vector, and combined with the weighted dependency depth change rate and composite boundary probability score of each semantic unit, the boundary-enhanced forget gate of each semantic unit is obtained. The input gate for each semantic unit is obtained based on the input feature vector, the input gate weight matrix, and the bias vector. Based on the boundary-enhanced forget gate and input gate of each semantic unit, the candidate state is modulated by combining the technical entity category vector and the boundary position set, and the state vector of each semantic unit is determined by the state transition equation. Based on the state vector of each semantic unit and the corresponding state vector within the local context window, multi-level information is aggregated through a multi-head attention mechanism to determine the context-aware state vector of each semantic unit. Based on the context-aware state vector of each semantic unit, the entity label probability distribution is output.

6. The method according to claim 1, characterized in that, During the training of the entity recognition model, the total model loss is determined by a weighted sum of the syntax-semantic feature decoupling loss and the category balancing focus loss. The syntax-semantic feature decoupling loss is determined as follows: based on the bidirectional dependency graph convolutional features and multi-head dynamic convolutional features of each semantic unit, the orthogonality loss between syntax features and semantic features is calculated, and combined with a covariance matrix penalty term to determine the syntax-semantic feature decoupling loss. The category balancing focus loss is determined as follows: based on the category prediction probability, category weight, focus parameter, prior adjustment coefficient, and category technical importance weight of each semantic unit, the cross-entropy loss is adjusted to obtain the category balancing focus loss.

7. A method for constructing a knowledge graph for scientific and technological achievement application, characterized in that, The method includes: The entity recognition method for scientific and technological achievement application text as described in any one of claims 1-6 is used to output a structured entity sequence corresponding to the scientific and technological achievement application text. By using dependency path enhancement and multi-hop graph convolutional aggregation, relation extraction is performed on the text of the scientific and technological achievement application to establish the relationship information between entities; Based on the structured entity sequence and the association information between the entities, a knowledge graph for scientific and technological achievement application is constructed.

8. The method according to claim 7, characterized in that, The steps for extracting relationships from the technical achievement application texts and establishing inter-entity association information through dependency path enhancement and multi-hop graph convolutional aggregation include: The text of the application for scientific and technological achievements is subjected to compound nesting boundary detection, and the semantic units corresponding to the text of the application for scientific and technological achievements are dynamically divided; the syntactic paths connecting the semantic units are extracted to obtain the dependency syntactic paths connecting the entity pairs. The bidirectional dependency graph convolution feature extraction module performs multi-hop feature propagation on nodes on the path between entities to fuse syntactic relations and technical attribute features, thereby obtaining bidirectional dependency graph convolution features between entity pairs. The convolutional features of the bidirectional dependency graph of entity pairs and their contexts are input into a multilayer perceptron to predict predefined relationship categories and obtain information on the association between entities.

9. The method according to claim 7, characterized in that, The steps for constructing a knowledge graph for scientific and technological achievement applications based on the structured entity sequence and the inter-entity association information include: The structured entity sequence is aligned and disambiguated to identify multiple target entities; Based on the inter-entity association information, entity relationship transformation is performed to obtain triples; Based on a predefined graph pattern, the triples are stored in a graph database to obtain a knowledge graph for scientific and technological achievement applications.

10. A physical recognition device for scientific and technological achievement applications, characterized in that, The device includes: The text acquisition module is used to acquire the text of scientific and technological achievement application to be identified; The preprocessing module is used to preprocess the text of the scientific and technological achievement application and input the preprocessed text into a pre-trained entity recognition model. The preprocessing includes data cleaning and standardization. The entity recognition model includes a composite nested boundary detection module, a bidirectional dependency graph convolution module, a multi-head dynamic convolution gating module, a gating hierarchical state transition module, and an entity recognition module. The model recognition module is used to perform compound nesting boundary detection on the text through the compound nesting boundary detection module, dynamically dividing the semantic units corresponding to the text; extracting the bidirectional dependency graph convolution features of each semantic unit through the bidirectional dependency graph convolution module; capturing the multi-head dynamic local features of each semantic unit based on the bidirectional dependency graph convolution features of each semantic unit through the multi-head dynamic convolution gating module; and performing state transition based on the gating hierarchical state transition mechanism according to the bidirectional dependency graph convolution features and multi-head dynamic local features of each semantic unit through the gating hierarchical state transition and entity recognition module, generating state vectors word by word and decoding entity labels, and outputting the structured entity sequence corresponding to the scientific and technological achievement application text.

11. A knowledge graph construction device for scientific and technological achievement application, characterized in that, The device includes: The entity recognition device for scientific and technological achievement application type as described in claim 10 outputs the structured entity sequence corresponding to the scientific and technological achievement application type text. The relation extraction module is used to extract relations from the technical achievement application text through dependency path enhancement and multi-hop graph convolutional aggregation, and to establish the relationship information between entities. The graph construction module is used to construct a knowledge graph for scientific and technological achievement applications based on the structured entity sequence and the association information between the entities.

12. An electronic device, characterized in that, The method includes a processor and a memory, the memory storing computer-executable instructions executable by the processor, the processor executing the computer-executable instructions to implement the method of any one of claims 1 to 9.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions that, when invoked and executed by a processor, cause the processor to perform the method according to any one of claims 1 to 9.

Images