Project evaluation and review method and system integrating natural language processing

By building a project knowledge graph and using deep learning technology, implicit evaluation rules and evaluation knowledge are extracted and high-quality evaluation opinions are generated, which solves the comprehensive understanding of project evaluation and the quality of opinion generation in the existing technology, and improves the accuracy of evaluation and interpretability.

CN118780767BActive Publication Date: 2024-12-06BEIJING BEILONG QINGYUN SOFTWARE CO LTD
View PDF 5 Cites 0 Cited by

Patent Information

Application Number
CN202411273552.X
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-09-12
Publication Date
2024-12-06
Estimated Expiration
2044-09-12

AI Technical Summary

Technical Problem

The existing technology lacks comprehensive understanding and representation in project review, ignores multimodal information and complex semantic relationships, insufficient utilization of review knowledge, low quality of evaluation opinion generation, poor generalization ability and insufficient interpretability.

Method used

By constructing the knowledge graph of the items to be reviewed, using the pre-trained heterogeneous graph neural network model for initial screening, using the spectral clustering algorithm to obtain semantic similar sub-graphs, combining the heterogeneous graph attention network to extract implicit evaluation rules, and dynamic memory network Search review knowledge, and generate review results through the learning model of the few sample graphs, and dynamically adjust the model to improve the review quality.

Benefits of technology

It has achieved a comprehensive understanding of project review and high-quality opinion generation, improved the pertinence, accuracy and interpretability of reviews, overcome the problem of insufficient data samples, and continuously optimized the review experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118780767B_ABST
    Figure CN118780767B_ABST
Patent Text Reader

Abstract

The present invention provides a project evaluation and review method and system integrating natural language processing, which relates to the field of natural language technology, including constructing a knowledge graph of a project to be reviewed; obtaining a preliminary screening result of the project to be reviewed by embedding a node representation output in a heterogeneous graph neural network model; if the project to be reviewed enters a review session, obtaining k semantically similar subgraphs in a knowledge graph of historical reviewed projects that are semantically similar to the project to be reviewed by a spectral clustering algorithm; obtaining implicit evaluation rules for the project to be reviewed; inputting the knowledge graph of the project to be reviewed and the extracted implicit evaluation rules into a dynamic memory network, retrieving review knowledge related to the project to be reviewed by a content-related addressing mechanism, and aggregating review opinions and suggestions for the project to be reviewed by an adaptive fusion mechanism; generating a review result of the project to be reviewed, and dynamically adjusting a few-sample graph learning model according to the review result and feedback from review experts.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to natural language technology, and in particular to a project evaluation and review method and system integrating natural language processing. Background Art

[0002] In recent years, artificial intelligence technology has developed rapidly, among which natural language processing technology has made significant progress in text mining, sentiment analysis, knowledge reasoning, etc. Applying natural language processing technology to the field of project review can effectively process project text materials, automatically extract project features, mine implicit evaluation rules, and generate objective and comprehensive review opinions. In the existing technology, there have been some attempts to apply natural language processing to project review.

[0003] However, the existing technology still has the following shortcomings: First, there is a lack of comprehensive understanding and representation of projects, ignoring the multimodal information and complex semantic relationships of projects; second, the review knowledge is insufficiently utilized, failing to fully explore and utilize the experience and rules contained in historical review data; third, the quality of review opinion generation needs to be improved, and the generated review opinions have problems such as poor generalization ability and lack of interpretability. Summary of the invention

[0004] The embodiments of the present invention provide a project evaluation and review method and system integrating natural language processing, which can at least solve some problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] Provide a project evaluation and review method that integrates natural language processing, including:

[0007] Based on the pre-processed structured and unstructured data of the project to be reviewed, a knowledge graph of the project to be reviewed is constructed; the knowledge graph of the project to be reviewed is embedded into a pre-trained heterogeneous graph neural network model to obtain an output node embedding representation, and according to the node embedding representation, a graph representation of the knowledge graph of the project to be reviewed is obtained through a graph pooling layer, and the graph representation is input into a graph classifier to obtain a preliminary screening result of the project to be reviewed; and according to the preliminary screening result, whether the project to be reviewed enters the meeting review stage is determined;

[0008] If the project to be reviewed enters the review phase, based on the knowledge graph of the project to be reviewed, according to the preset number of similar subgraphs k, a spectral clustering algorithm is used to obtain k semantically similar subgraphs in the knowledge graph of historical reviewed projects that are semantically similar to the project to be reviewed; the k semantically similar subgraphs are fused to construct a heterogeneous meta-path set of the project to be reviewed, and the heterogeneous meta-path set is input into a pre-trained heterogeneous graph attention network, and cross-modal interaction features are extracted through a multi-head attention mechanism to obtain implicit evaluation rules for the project to be reviewed;

[0009] The knowledge graph of the project to be reviewed and the extracted implicit evaluation rules are input into the dynamic memory network, and the review knowledge related to the project to be reviewed is retrieved through the content-related addressing mechanism, and the review opinions and suggestions for the project to be reviewed are aggregated through the adaptive fusion mechanism; the knowledge graph of the project to be reviewed and the review opinions and suggestions are input into the few-shot graph learning model to generate the review results of the project to be reviewed, and the few-shot graph learning model is dynamically adjusted according to the review results and the feedback from the review experts.

[0010] According to a second aspect of the embodiments of the present invention,

[0011] The first unit is used to construct a knowledge graph of the project to be reviewed based on the pre-processed structured and unstructured data of the project to be reviewed; embed the knowledge graph of the project to be reviewed into a pre-trained heterogeneous graph neural network model to obtain an output node embedding representation, obtain a graph representation of the knowledge graph of the project to be reviewed through a graph pooling layer based on the node embedding representation, input the graph representation into a graph classifier to obtain a preliminary screening result of the project to be reviewed; determine whether the project to be reviewed will enter the meeting review stage based on the preliminary screening result;

[0012] The second unit is used for, if the project to be reviewed enters the review stage, based on the knowledge graph of the project to be reviewed, according to the preset number of similar subgraphs k, obtaining k semantically similar subgraphs in the knowledge graph of historical reviewed projects that are semantically similar to the project to be reviewed through a spectral clustering algorithm; fusing the k semantically similar subgraphs to construct a heterogeneous meta-path set of the project to be reviewed, inputting the heterogeneous meta-path set into a pre-trained heterogeneous graph attention network, extracting cross-modal interaction features through a multi-head attention mechanism, and obtaining implicit evaluation rules for the project to be reviewed;

[0013] The third unit is used to input the knowledge graph of the project to be reviewed and the extracted implicit evaluation rules into the dynamic memory network, retrieve the review knowledge related to the project to be reviewed through the content-related addressing mechanism, and aggregate the review opinions and suggestions for the project to be reviewed through the adaptive fusion mechanism; input the knowledge graph of the project to be reviewed and the review opinions and suggestions into the few-shot graph learning model to generate the review results of the project to be reviewed, and dynamically adjust the few-shot graph learning model according to the review results and the feedback from the review experts.

[0014] According to a third aspect of the embodiments of the present invention,

[0015] An electronic device is provided, comprising:

[0016] processor;

[0017] a memory for storing processor-executable instructions;

[0018] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0019] A fourth aspect of the embodiments of the present invention is:

[0020] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0021] This application constructs a knowledge graph of the projects to be reviewed based on preprocessed structured and unstructured data, effectively integrating the multi-source heterogeneous information of the projects and providing a comprehensive data foundation for subsequent reviews. The node embedding representation is obtained using a pre-trained heterogeneous graph neural network model, and the initial screening of the projects to be reviewed is achieved through a graph pooling layer and a graph classifier, which can quickly screen out projects that do not meet the basic conditions and improve review efficiency.

[0022] By using the spectral clustering algorithm, we can obtain subgraphs with similar semantics to the project to be reviewed from the knowledge graph of historical reviewed projects and construct a set of heterogeneous meta-paths, which fully utilizes the historical review knowledge and provides an effective way to mine implicit evaluation rules. By using the pre-trained heterogeneous graph attention network and multi-head attention mechanism, we can fully extract cross-modal interaction features, explore the implicit evaluation rules of the project to be reviewed, and improve the pertinence and accuracy of the review.

[0023] The dynamic memory network is introduced to retrieve the review knowledge related to the project to be reviewed through the content-related addressing mechanism, and the review opinions and suggestions are aggregated through the adaptive fusion mechanism, which realizes the accurate matching and fusion of the review knowledge and improves the quality and interpretability of the review opinions. The few-sample graph learning model is used to generate the review results of the project to be reviewed, which overcomes the problem of insufficient review data samples and improves the reliability of the review results. At the same time, the learning model is dynamically adjusted according to the feedback from the review experts to achieve the continuous accumulation and optimization of review experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 A schematic diagram of a process flow of a project evaluation and review method integrating natural language processing according to an embodiment of the present invention;

[0025] Figure 2 This is a schematic diagram of the structure of a project evaluation and review system that integrates natural language processing in an embodiment of the present invention. DETAILED DESCRIPTION

[0026] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0027] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0028] Figure 1 FIG. 1 is a flow chart of a project evaluation and review method integrating natural language processing according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0029] S101. Based on the pre-processed structured and unstructured data of the project to be reviewed, a knowledge graph of the project to be reviewed is constructed; the knowledge graph of the project to be reviewed is embedded into a pre-trained heterogeneous graph neural network model to obtain an output node embedding representation, and according to the node embedding representation, a graph representation of the knowledge graph of the project to be reviewed is obtained through a graph pooling layer, and the graph representation is input into a graph classifier to obtain a preliminary screening result of the project to be reviewed; and according to the preliminary screening result, whether the project to be reviewed enters the meeting review stage is determined;

[0030] S102. If the project to be reviewed enters the review phase, based on the knowledge graph of the project to be reviewed, according to the preset number of similar subgraphs k, a spectral clustering algorithm is used to obtain k semantically similar subgraphs in the knowledge graph of historical reviewed projects that are semantically similar to the project to be reviewed; the k semantically similar subgraphs are fused to construct a heterogeneous meta-path set of the project to be reviewed, and the heterogeneous meta-path set is input into a pre-trained heterogeneous graph attention network, and cross-modal interaction features are extracted through a multi-head attention mechanism to obtain implicit evaluation rules for the project to be reviewed;

[0031] S103. Input the knowledge graph of the project to be reviewed and the extracted implicit evaluation rules into the dynamic memory network, retrieve the review knowledge related to the project to be reviewed through the content-related addressing mechanism, and aggregate the review opinions and suggestions for the project to be reviewed through the adaptive fusion mechanism; input the knowledge graph of the project to be reviewed and the review opinions and suggestions into the few-shot graph learning model to generate the review results of the project to be reviewed, and dynamically adjust the few-shot graph learning model according to the review results and the feedback from the review experts.

[0032] This application constructs a knowledge graph of the projects to be reviewed based on preprocessed structured and unstructured data, effectively integrating the multi-source heterogeneous information of the projects and providing a comprehensive data foundation for subsequent reviews. The node embedding representation is obtained using a pre-trained heterogeneous graph neural network model, and the initial screening of the projects to be reviewed is achieved through a graph pooling layer and a graph classifier, which can quickly screen out projects that do not meet the basic conditions and improve review efficiency.

[0033] By using the spectral clustering algorithm, we can obtain subgraphs with similar semantics to the project to be reviewed from the knowledge graph of historical reviewed projects and construct a set of heterogeneous meta-paths, which fully utilizes the historical review knowledge and provides an effective way to mine implicit evaluation rules. By using the pre-trained heterogeneous graph attention network and multi-head attention mechanism, we can fully extract cross-modal interaction features, explore the implicit evaluation rules of the project to be reviewed, and improve the pertinence and accuracy of the review.

[0034] The dynamic memory network is introduced to retrieve the review knowledge related to the project to be reviewed through the content-related addressing mechanism, and the review opinions and suggestions are aggregated through the adaptive fusion mechanism, which realizes the accurate matching and fusion of the review knowledge and improves the quality and interpretability of the review opinions. The few-sample graph learning model is used to generate the review results of the project to be reviewed, which overcomes the problem of insufficient review data samples and improves the reliability of the review results. At the same time, the learning model is dynamically adjusted according to the feedback from the review experts to achieve the continuous accumulation and optimization of review experience.

[0035] In an optional embodiment,

[0036] Based on the pre-processed structured and unstructured data of the projects to be reviewed, a knowledge graph of the projects to be reviewed is constructed, including:

[0037] The preprocessed structured data and unstructured data are aligned using a multi-strategy entity alignment method that combines string matching, word vector similarity calculation, rule matching and transfer learning, and the target strategy is selected to map entities in different source data to unique nodes in the knowledge graph; among them, string matching is implemented using the edit distance algorithm and Jaccard similarity based on part of speech and word form, word vector similarity calculation is implemented using a word vector model that integrates dictionary tree index and a transformation model represented by a bidirectional encoder, rule matching is implemented using a knowledge-driven regular expression generation method and a conditional random field model optimized by reinforcement learning, and transfer learning is implemented using a cross-domain entity alignment model based on a graph attention network;

[0038] Extract semantic relationships between entities from aligned unstructured text data, and organize the extracted entities and semantic relationships into semantically similar subgraphs including project portrait subgraph, technology route subgraph, team capability subgraph and resource demand subgraph according to the multi-level concept hierarchy and fine-grained relationship types defined by the fusion domain ontology and industry standards; among them, the project portrait subgraph adopts dynamic heterogeneous graph neural network to model the evolution of project attributes, the technology route subgraph adopts citation network to determine the evolution of technology topics, the team capability subgraph adopts temporal multi-view learning method to represent the dynamic changes of team capabilities, and the resource demand subgraph adopts knowledge reasoning and reinforcement learning method to optimize resource allocation;

[0039] An improved transfer embedding knowledge representation learning algorithm and a heterogeneous graph neural network algorithm are applied to each semantically similar subgraph for representation learning. The entity and semantic relationship vectors obtained by representation learning are used to calculate the similarity between different semantically similar subgraphs through a preset knowledge graph alignment method. The head entity, semantic relationship and constraint rules are used as input, and the missing implicit associations between semantically similar subgraphs are predicted through multi-hop relationship path reasoning. Confidence estimation and Markov logic network are used to verify and optimize the implicit associations predicted by reasoning, and the associated triples are obtained and added to the knowledge base to obtain the knowledge graph of the project to be reviewed.

[0040] For example, in the technical solution proposed in this article, the process of constructing a knowledge graph based on the pre-processed structured and unstructured data of the project to be reviewed is mainly divided into three steps: entity alignment, semantic relationship extraction and organization, and knowledge graph construction. The technical solution will be elaborated in detail around these three steps.

[0041] Step 1: Entity alignment;

[0042] The purpose of entity alignment is to map entities in different source data to unique nodes in the knowledge graph. This paper adopts an entity alignment method that combines multiple strategies, including string matching, word vector similarity calculation, rule matching, and transfer learning.

[0043] String matching is achieved using the edit distance algorithm and the Jaccard similarity based on part-of-speech and word form. The edit distance algorithm measures the similarity between strings by calculating the minimum number of edit operations required to transform one string into another. The Jaccard similarity based on part-of-speech and word form takes into account part-of-speech and word form information, and measures the similarity by calculating the proportion of part-of-speech and word form pairs shared by two strings to the total number of part-of-speech and word form pairs.

[0044] The word vector similarity calculation is implemented by using the word vector model of the fusion dictionary tree index and the transformation model represented by the bidirectional encoder. The word vector model of the fusion dictionary tree index introduces the dictionary tree index structure on the basis of the traditional word vector model, which can quickly retrieve words with similar semantics to the target word, improving the efficiency and accuracy of the word vector similarity calculation. The transformation model represented by the bidirectional encoder uses a bidirectional recurrent neural network to encode the entity name to obtain the vector representation of the entity, and then measures the similarity between entities by calculating the cosine similarity between vectors.

[0045] Rule matching is implemented using a knowledge-driven regular expression generation method and a conditional random field model optimized by reinforcement learning. The knowledge-driven regular expression generation method automatically generates regular expression rules for entity matching based on domain knowledge, which can effectively improve the coverage and accuracy of the rules. The conditional random field model optimized by reinforcement learning improves the performance of entity alignment by modeling the entity alignment problem as a sequence labeling task and optimizing the parameters of the conditional random field model using a reinforcement learning algorithm.

[0046] Transfer learning is implemented using a cross-domain entity alignment model based on a graph attention network. The model first constructs entity relationship graphs in the source domain and the target domain respectively, then uses the graph attention network to learn the cross-domain representation of entities, and finally achieves cross-domain entity alignment by calculating the similarity between entity representations. Compared with traditional entity alignment methods, this model can effectively utilize the knowledge of the source domain and improve the effect of entity alignment in the target domain.

[0047] In the entity alignment process, this paper selects a target strategy to map entities in different source data to unique nodes in the knowledge graph. The selection of the target strategy takes into account the confidence of the entity alignment results, the fit of domain knowledge, and the results of manual verification to ensure the accuracy of entity alignment.

[0048] Step 2: semantic relationship extraction and organization;

[0049] Based on entity alignment, semantic relations between entities are extracted from unstructured text data, and the extracted entities and semantic relations are organized into semantically similar subgraphs according to the multi-level concept hierarchy and fine-grained relationship types defined by domain ontologies and industry standards.

[0050] Semantic relationship extraction uses a deep learning-based method. By training an end-to-end neural network model, it can automatically identify and extract semantic relationships between entities from text. Specifically, the named entity recognition model is first used to identify entities in the text, and then the relationship classification model is used to determine whether there is a semantic relationship between entity pairs and the type of relationship. The relationship classification model uses a convolutional neural network based on the attention mechanism. By introducing the attention mechanism, it can automatically learn the key information between entity pairs and improve the accuracy of relationship classification.

[0051] Based on the extraction of semantic relations, the extracted entities and semantic relations are organized into semantic similarity subgraphs according to the multi-level concept hierarchy and fine-grained relationship types defined by domain ontology and industry standards. This paper constructs four types of semantic similarity subgraphs, namely project portrait subgraph, technology route subgraph, team capability subgraph and resource demand subgraph.

[0052] The project portrait subgraph uses a dynamic heterogeneous graph neural network to model the evolution of project attributes. The dynamic heterogeneous graph neural network introduces the time dimension to model the changes of project attributes over time, and uses heterogeneous graph neural networks to process the complex associations between project attributes, thereby comprehensively depicting the dynamic evolution of project portraits.

[0053] The technology roadmap sub-graph uses citation networks to determine the evolution of technology themes. By analyzing the references in the project proposal, a citation network is constructed, and a community discovery algorithm is used to identify technology theme communities in the citation network to reveal the evolution of technology themes.

[0054] The team capability subgraph uses a time-series multi-view learning method to represent the dynamic changes of team capabilities. The time-series multi-view learning method integrates multi-view information such as the academic resume, project experience, and patent achievements of team members to learn the time-series representation of team capabilities and depict the dynamic changes of team capabilities.

[0055] The resource demand subgraph uses knowledge reasoning and reinforcement learning methods to optimize resource allocation. First, knowledge reasoning techniques such as ontology reasoning and rule reasoning are used to infer the resource requirements of the project based on project attributes and technical routes. Then, reinforcement learning algorithms are used to optimize resource allocation strategies to achieve efficient resource utilization.

[0056] Step 3: Knowledge Graph Construction

[0057] Based on the construction of semantically similar subgraphs, representation learning and relational reasoning techniques are used to construct a knowledge graph of the projects to be reviewed.

[0058] The improved transfer embedding knowledge representation learning algorithm and heterogeneous graph neural network algorithm are applied to each semantically similar subgraph for representation learning. The improved transfer embedding knowledge representation learning algorithm realizes the migration of knowledge representation between different domains by introducing cross-domain alignment regularization terms, thereby improving the generalization ability of representation learning. The heterogeneous graph neural network algorithm learns high-order interaction information in heterogeneous graph data by designing aggregation functions of different types of nodes and edges, thereby enhancing the expressive power of node representation. The entity and semantic relationship vectors obtained by representation learning are used to calculate the similarity between different semantically similar subgraphs through a preset knowledge graph alignment method.

[0059] Taking the head entity, semantic relationship and constraint rules as input, multi-hop relationship path reasoning is used to predict the missing implicit associations between semantically similar subgraphs. Multi-hop relationship path reasoning constructs multi-hop relationship paths between entities, uses the combined features of the paths to learn the representation of implicit associations, and uses a neural network model based on the attention mechanism to predict the existence probability of implicit associations. At the same time, confidence estimation and Markov logic network are used to verify and optimize the implicit associations predicted by reasoning to improve the reliability of association prediction. Finally, the predicted implicit associations are represented as association triples and added to the knowledge base to form a knowledge graph of the project to be reviewed.

[0060] Through the above three steps, the technical solution proposed in this paper realizes the construction of a knowledge graph of the project to be reviewed based on preprocessed structured and unstructured data. This knowledge graph integrates multi-source heterogeneous information of the project, depicts different aspects of the project through semantically similar subgraphs, and uses representation learning and relational reasoning technology to mine implicit associations between subgraphs, forming a comprehensive and in-depth semantic representation of the project, providing important knowledge support for subsequent project reviews.

[0061] In an optional embodiment,

[0062] Taking the head entity, semantic relations and constraint rules as input, the missing implicit associations between semantically similar subgraphs are predicted through multi-hop relational path reasoning, including:

[0063] Calculate the attention weight between the head entity embedding and all outgoing edge relationship embeddings, select the top M outgoing edge relationships most relevant to the head entity according to the attention weight, and use the head entity node and the selected outgoing edge relationship as the starting point of the current reasoning path; the attention weight is calculated by the dot product of the head entity embedding and the outgoing edge relationship embedding and normalized;

[0064] According to the last relation node of the current reasoning path, a relation-aware entity neighbor subgraph is constructed, which includes the entity nodes directly connected to the last relation and their one-hop neighbors; the entity neighbor subgraph is encoded using a graph attention network to obtain the hidden state representation of the entity node; each newly generated reasoning path is sequence modeled using a gated recurrent unit network, and the entity nodes and relations in the path are embedded in the sequence into the gated recurrent unit network to obtain the hidden state representation of the reasoning path; finally, the hidden state of the path and the hidden state of the last entity neighbor subgraph are aggregated through the attention mechanism to predict the next-hop relationship;

[0065] Iteratively execute the above steps of constructing a relationship-aware entity neighbor subgraph and subsequent steps according to the last relationship node of the current reasoning path until the preset hop count threshold is reached or the potential tail entity node is reached, and multiple candidate reasoning paths are generated; a multi-path exploration strategy based on beam search is adopted to retain the N relationship paths with the highest confidence from the candidate reasoning paths; wherein the confidence is determined based on the reasoning path length, the prediction probability of each hop and the semantic matching degree of the node relationship; post-processing and optimizing the candidate association paths, using the learned path constraint rules to filter and reorder the candidate paths, filter out the paths that do not meet the rule constraints, and obtain the final cross-subgraph association path set as the missing implicit association between semantically similar subgraphs.

[0066] For example, in the technical solution proposed in this paper, in order to predict the missing implicit associations between semantically similar subgraphs, a method based on multi-hop relational path reasoning is adopted. Taking the head entity, semantic relations and constraint rules as input, the reasoning path is modeled by recursively constructing a relation-aware entity neighbor subgraph and a gated recursive unit network, and the attention mechanism is used to aggregate the path and subgraph information, and finally a set of cross-subgraph association paths is generated. The key steps of this method will be elaborated in detail below.

[0067] First, the attention weights between the head entity embedding and all outgoing edge relationship embeddings are calculated. The attention weights are calculated by taking the dot product of the head entity embedding and the outgoing edge relationship embedding and normalizing them. The dot product operation measures the similarity between two vectors, and a higher score indicates a stronger correlation between the head entity and the outgoing edge relationship. The normalization operation converts the dot integral value into a weight distribution, so that the sum of the weights of all outgoing edge relationships is 1.

[0068] Based on the calculated attention weights, select the first M outgoing relations that are most relevant to the head entity. Here M is a preset hyperparameter, indicating the number of relations selected in each hop of reasoning. The selected M relations and the head entity node are used as the starting point of the current reasoning path.

[0069] According to the last relation node of the current reasoning path, a relation-aware entity neighbor subgraph is constructed. The subgraph includes the entity nodes directly connected to the last relation and their one-hop neighbors. One-hop neighbors refer to other entity nodes that are connected to the directly connected entity nodes by edges. By introducing one-hop neighbors, the information scope of the subgraph can be expanded and more context information can be captured.

[0070] In order to encode the entity neighbor subgraph, a graph attention network is used to learn the representation of the nodes in the subgraph. The graph attention network introduces the attention mechanism to adaptively assign weights to different neighbor nodes and aggregate neighbor information to obtain the hidden state representation of the node. Specifically, each node in the graph attention network updates its own hidden state by the weighted sum of the hidden state of its neighbor nodes and the attention weight. The multi-head attention mechanism is used to capture various interaction modes between nodes and improve the expressiveness of node representation.

[0071] For each newly generated reasoning path, a gated recurrent unit network is used for sequence modeling. The entity nodes and relationships in the path are embedded into a sequence in the order of the path and input into the gated recurrent unit network. By introducing update gates and reset gates, the gated recurrent unit network can effectively capture long-distance dependencies in the sequence and adaptively determine the retention and forgetting of historical information. The hidden state of the network is updated at each time step, and finally the hidden state representation of the entire reasoning path is obtained.

[0072] In order to predict the next-hop relationship, it is necessary to comprehensively consider the hidden state of the current reasoning path and the hidden state of the end entity neighbor subgraph. The attention mechanism is used to aggregate the two hidden states. First, the attention weight between the hidden state of the reasoning path and the hidden state of each node in the subgraph is calculated. Then, the attention weight is used to weighted sum the hidden states of the subgraph nodes to obtain the aggregated subgraph hidden state. Finally, the reasoning path hidden state and the aggregated subgraph hidden state are concatenated and passed through a feedforward neural network to obtain the predicted probability distribution of the next-hop relationship.

[0073] The above steps are iterated until the recursion reaches the preset hop count threshold or reaches the potential tail entity node, generating multiple candidate reasoning paths. The hop count threshold is a preset hyperparameter used to control the maximum length of the reasoning path. Potential tail entity nodes refer to entity nodes that are semantically related to the head entity and meet certain conditions, such as being in the same semantically similar subgraph as the head entity and having a certain association strength.

[0074] In order to improve the efficiency and accuracy of reasoning, a multi-path exploration strategy based on beam search is adopted. In each hop reasoning process, the N relationship paths with the highest confidence are retained as the starting point for the next hop exploration. The confidence comprehensively considers the length of the reasoning path, the prediction probability of each hop, and the semantic matching degree of the node relationship. Shorter paths, paths with higher prediction probabilities, and paths with higher semantic matching degrees of node relationships will be retained first.

[0075] After generating candidate association paths, post-processing optimization is performed to improve the quality of association prediction. Candidate paths are filtered and reordered using the learned path constraint rules. Path constraint rules can be logical rules defined based on domain knowledge or statistical rules automatically learned from data. Paths that do not meet the rule constraints will be filtered out, and paths that meet the rule constraints will be reordered according to the confidence of the rule to obtain the final set of cross-subgraph association paths.

[0076] Through the above steps, the multi-hop relational path reasoning method proposed in this paper can effectively predict the missing implicit associations between semantically similar subgraphs. This method makes full use of prior knowledge such as head entities, semantic relations, and constraint rules, recursively constructs relation-aware entity neighbor subgraphs and gated recursive unit networks to model the reasoning path, and introduces an attention mechanism to aggregate path and subgraph information to generate a high-quality set of cross-subgraph association paths. Compared with traditional association prediction methods based on graph embedding, this method can explicitly model the reasoning process and provide more transparent and explainable prediction results. At the same time, through multi-path exploration and post-processing optimization, this method can improve prediction efficiency while ensuring prediction quality, providing an effective solution for association discovery between semantically similar subgraphs.

[0077] In an optional embodiment,

[0078] The k semantically similar subgraphs are integrated to construct a heterogeneous meta-path set of the project to be reviewed, including:

[0079] Based on graph convolutional neural network, low-dimensional dense feature vector representation of nodes in semantically similar subgraphs is extracted. The semantic similarity of isomorphic nodes between different semantically similar subgraphs is calculated through cosine similarity function. Isomorphic nodes with semantic similarity exceeding the preset similarity threshold and the same node type are merged. The node feature vector is updated on the basis of retaining the original attribute information of the node to obtain the fused heterogeneous knowledge graph.

[0080] Applying a meta-path-constrained random walk algorithm on the fused heterogeneous knowledge graph to generate a node sequence containing heterogeneous information, constructing a transition probability matrix according to three predefined meta-path modes: project-personnel-project, project-institution-project, and project-research field-project, calculating the transition probability from the current node to different types of neighboring nodes, and adjusting the random degree and meta-path preference weight in the walk strategy in combination with the node concentration and web page ranking value;

[0081] The principal component analysis method is used to reduce the dimension of the node sequence, and the essential pattern of the node sequence is extracted. The frequent pattern mining algorithm is used to identify the node subsequences that conform to the three meta-path patterns from the compressed node sequence. The minimum length of the subsequence and the node heterogeneity are controlled by setting the support and confidence thresholds of the meta-path matching to generate candidate meta-path instances. The quality of the candidate meta-path instances is evaluated based on the utilization rate of heterogeneous information and structural stability, and the m meta-path instances with the best quality are screened out to form the heterogeneous meta-path set of the project to be reviewed.

[0082] For example, in the technical solution proposed in this paper, in order to construct a heterogeneous meta-path set of the project to be reviewed, it is necessary to fuse k semantically similar subgraphs to generate a heterogeneous knowledge graph, and then mine high-quality meta-path instances based on it. The key steps of this method will be described in detail below.

[0083] First, the graph convolutional neural network is used to learn the features of each semantically similar subgraph and extract the low-dimensional dense feature vector representation of the node. The graph convolutional neural network updates the feature representation of the node by aggregating the neighbor information of the node, which can effectively capture the structure and semantic information of the subgraph. Specifically, the forward propagation process of the graph convolution layer is as follows: the feature vector of each node is aggregated with the feature vectors of its neighbor nodes, and the aggregation function can be summation, average or maximum pooling, etc.; the aggregated feature vector is transformed nonlinearly to obtain the new feature vector of the node; the stacking of multiple layers of graph convolutional layers can capture the high-order neighbor information of the node and obtain a more global node representation.

[0084] After obtaining the feature vector representation of the node, the semantic similarity of isomorphic nodes between different semantically similar subgraphs is calculated by the cosine similarity function. The cosine similarity function measures the similarity between two vectors by calculating the cosine value of the angle between them, and the value range is between -1 and 1. For isomorphic nodes with semantic similarity exceeding the preset threshold and the same node type, they are merged. The merged node retains the attribute information of the original node and updates its feature vector to the weighted average of the feature vector of the original node. In this way, the heterogeneity and semantic characteristics of the node can be retained while fusing the information of different subgraphs. Finally, the fused heterogeneous knowledge graph is obtained.

[0085] On the fused heterogeneous knowledge graph, the random walk algorithm with meta-path constraints is applied to generate node sequences containing heterogeneous information. The random walk algorithm generates node sequences by randomly jumping on the graph, while the meta-path constraints introduce restrictions on node types and edge types, so that the generated node sequences conform to specific semantic patterns.

[0086] According to the three predefined meta-path patterns of project-personnel-project, project-institution-project and project-research field-project, a transition probability matrix is ​​constructed. Each element of the transition probability matrix represents the probability of jumping from the current node to a specific type of neighbor node. Specifically, for each node, the transition probability is calculated according to the type of its neighbor node and the weight of the connecting edge. At the same time, the transition probability is adjusted by combining the node's concentration (such as degree centrality) and web page ranking value (such as PageRank value) to balance the randomness and meta-path preference in the walking strategy. The higher the concentration of the node, the greater the probability of it being visited in the walking; the higher the web page ranking value, the more important the node is in the walking.

[0087] Through random walks constrained by meta-paths, a large number of node sequences containing heterogeneous information are generated. In order to extract the essential patterns of the node sequences, the principal component analysis method is used to reduce the dimensionality of the node sequences. Principal component analysis maps high-dimensional data to low-dimensional space through linear transformation and extracts the main features of the data. By retaining the first few principal components, the noise and redundant information in the node sequence can be removed, and the compressed node sequence representation can be obtained.

[0088] On the compressed node sequence, the frequent pattern mining algorithm is used to identify node subsequences that conform to the predefined meta-path pattern. The frequent pattern mining algorithm discovers patterns that appear frequently and meet specific conditions by setting support and confidence thresholds. The support threshold controls the minimum frequency of occurrence of the subsequence, and the confidence threshold controls the diversity of node types in the subsequence. By setting these two thresholds, the universality and heterogeneity of meta-path instances can be balanced to obtain a set of candidate meta-path instances.

[0089] In order to evaluate the quality of candidate meta-path instances, two indicators, heterogeneous information utilization and structural stability, are introduced. Heterogeneous information utilization measures the degree of coverage of meta-path instances for different types of nodes and edges in heterogeneous knowledge graphs. The higher the utilization, the more fully the meta-path instance can mine heterogeneous information. Structural stability measures the consistency of the meta-path instance in different node sequences. The higher the stability, the more the meta-path instance can reflect the inherent pattern of the knowledge graph.

[0090] Taking into account the utilization rate of heterogeneous information and structural stability, the candidate meta-path instances are evaluated and ranked for quality. The m meta-path instances with the best quality are selected to form the heterogeneous meta-path set of the project to be reviewed. Here, m is a preset hyperparameter, which indicates the number of selected meta-path instances. By controlling the size of m, the information coverage and computational efficiency of the meta-path set can be balanced.

[0091] Through the above steps, the heterogeneous meta-path construction method proposed in this paper can effectively fuse semantically similar subgraphs and mine meta-path instances containing heterogeneous information, thereby constructing a high-quality heterogeneous meta-path set for the project to be reviewed. This method makes full use of graph convolutional neural networks, random walk algorithms, and frequent pattern mining techniques to capture the structure and semantic information of heterogeneous knowledge graphs while extracting representative and discriminative meta-path patterns. The heterogeneous meta-path set lays the foundation for subsequent project representation learning, feature extraction, and similarity calculation, which helps to improve the accuracy and comprehensiveness of project review. Compared with the traditional meta-path construction method based on homogeneous networks, this method can fully mine the semantic information in heterogeneous networks, provide richer and more fine-grained project feature representation, and provide strong support for project review analysis.

[0092] In an optional embodiment,

[0093] The heterogeneous meta-path set is input into the pre-trained heterogeneous graph attention network, and the cross-modal interaction features are extracted through the multi-head attention mechanism to obtain the implicit evaluation rules of the items to be reviewed, including:

[0094] The heterogeneous meta-path set is input into the pre-trained heterogeneous graph attention network, which uses a multi-head attention mechanism to update the node embedding representation of each layer of the network, calculates the correlation between the neighbor nodes and the target node through the attention function based on the scaled dot product attention to obtain the attention score, and uses a multi-mask mechanism to filter the attention score, retain some neighbor nodes, and generate node-level attention weights;

[0095] A meta-path attention layer is set at the last layer of the heterogeneous graph attention network. The meta-path level attention mechanism based on the gating mechanism is used to calculate the embedding representation vector of each meta-path instance. The meta-path level attention mechanism controls the matching degree between the node embedding representation contained in the meta-path and the query vector through the gating unit, and generates the meta-path level attention weight based on the heterogeneity and centrality characteristics of the nodes. According to the prior knowledge of the meta-path type, the meta-path importance weight is calculated based on the frequency, coverage, and uniqueness indicators.

[0096] The meta-path importance weights are normalized, and the hierarchical analysis method is introduced in combination with expert knowledge to determine the upper and lower bounds of the weights of different meta-path types. The meta-path embedding representations output by the heterogeneous graph attention network are weighted summed according to the normalized meta-path importance weights, and the sequential dependencies between meta-path embeddings are modeled using a bidirectional gated recurrent unit network to obtain a low-dimensional embedding representation vector of the project to be reviewed that integrates local semantics, structural features and cross-modal association information. The implicit evaluation rules in the low-dimensional embedding representation vector are determined through a multi-layer cascaded interpreter network.

[0097] For example, in the technical solution proposed in this paper, in order to obtain the implicit evaluation rules of the project to be reviewed, it is necessary to input the heterogeneous meta-path set into the pre-trained heterogeneous graph attention network, and extract cross-modal interaction features through the multi-head attention mechanism. The key steps of this method will be described in detail below.

[0098] Heterogeneous Graph Attention Network is a deep learning model based on graph neural network, which can effectively process heterogeneous network data. Different from traditional graph neural network, Heterogeneous Graph Attention Network introduces a multi-head attention mechanism to adaptively extract and aggregate features of different types of nodes and edges in the network.

[0099] At each layer of the heterogeneous graph attention network, a multi-head attention mechanism is used to update the node embedding representation. Specifically, for each target node, its relevance to neighbor nodes is calculated through an attention function based on scaled dot product attention. The scaled dot product attention function performs a dot product operation on the query vector of the target node and the key vector of the neighbor node, and then divides it by the square root of the embedding dimension to obtain an attention score. The higher the attention score, the greater the influence of the neighbor node on the target node.

[0100] In order to improve the robustness and interpretability of the attention mechanism, a multi-mask mechanism is introduced to filter the attention scores. The multi-mask mechanism generates a mask matrix based on prior knowledge such as node type, edge type, and node degree, and performs element-level product operations on the attention scores to retain some important neighbor nodes and suppress the influence of noise nodes. The filtered attention scores are normalized by Softmax to obtain the node-level attention weights.

[0101] The node-level attention weight is weighted and summed with the embedding representation of neighboring nodes to obtain the new embedding representation of the target node. The multi-head attention mechanism captures multiple interaction modes between nodes and improves the expressiveness of node representation by executing multiple attention functions in parallel.

[0102] In the last layer of the heterogeneous graph attention network, a meta-path attention layer is set to calculate the embedding representation vector of each meta-path instance using a meta-path level attention mechanism based on a gating mechanism. The meta-path level attention mechanism aims to capture the semantic associations and structural features between nodes in a meta-path instance.

[0103] For each meta-path instance, a gating unit is used to control the degree of match between the node embedding representation contained in the meta-path and the query vector. The query vector can be a randomly initialized vector or a high-level feature extracted from the middle layer of the heterogeneous graph attention network. The gating unit compresses the matching degree to between 0 and 1 through the Sigmoid function as a modulation factor of the meta-path level attention weight.

[0104] At the same time, the meta-path level attention weights are generated by combining the heterogeneity and centrality features of the nodes. The heterogeneity of the nodes measures the diversity of node types. The higher the heterogeneity, the more node types the meta-path instance covers. The centrality of the nodes measures the importance of the nodes in the network. The higher the centrality, the more key nodes the meta-path instance contains. After linear transformation and nonlinear activation, the heterogeneity and centrality features are multiplied element-by-element with the output of the gated unit to obtain the final meta-path level attention weights.

[0105] The meta-path level attention weights are weighted and summed with the embedding representations of the nodes in the meta-path instance to obtain the embedding representation vector of the meta-path instance. In this way, the meta-path attention layer can adaptively extract the key information of the meta-path instance and filter out redundant and noisy information.

[0106] In order to further improve the quality of meta-path embedding representation, the meta-path importance weight is introduced to distinguish different types of meta-paths. The meta-path importance weight is calculated based on frequency, coverage, uniqueness and other indicators. The frequency indicator measures the number of times a meta-path instance appears in a heterogeneous knowledge graph. The higher the frequency, the more common the meta-path instance is; the coverage indicator measures the proportion of nodes and edges covered by the meta-path instance. The higher the coverage, the more representative the meta-path instance is; the uniqueness indicator measures the ability of the meta-path instance to distinguish different nodes. The higher the uniqueness, the more distinguishable the meta-path instance is.

[0107] The importance weight of each meta-path type is calculated by comprehensively considering the frequency, coverage, and uniqueness indicators. In order to introduce expert knowledge, the hierarchical analysis method is used to determine the upper and lower bounds of the weights of different meta-path types. The hierarchical analysis method compares the importance of meta-path types pairwise, constructs a judgment matrix, and calculates the eigenvector corresponding to the maximum eigenvalue of the matrix to obtain the relative value of the weight. The relative value is normalized to the preset upper and lower bounds to obtain the final meta-path type weight.

[0108] After normalizing the importance weights of the meta-paths, they are weighted summed with the meta-path embedding representations output by the heterogeneous graph attention network to obtain the project embedding representation that integrates different meta-path information. In order to further capture the sequential dependencies between meta-paths, a bidirectional gated recurrent unit network is used to model the weighted summed embedding representation. By introducing a bidirectional propagation mechanism, the bidirectional gated recurrent unit network can simultaneously capture the forward and backward dependencies between meta-paths, and obtain a more comprehensive and accurate project embedding representation.

[0109] After obtaining the low-dimensional embedding representation vector of the project to be reviewed that integrates local semantics, structural features and cross-modal association information, the implicit evaluation rules are determined through a multi-layer cascaded interpreter network. The interpreter network is a deep learning model based on the attention mechanism, which can extract key features and rules from the embedded representation.

[0110] Each layer of the interpreter network consists of an attention mechanism and a feedforward neural network. The attention mechanism calculates the importance weight of each dimension in the embedding representation based on the query vector. The higher the weight, the more critical the feature corresponding to the dimension. The feedforward neural network performs a nonlinear transformation on the weighted embedding representation to extract high-level features. The cascade of multiple layers of interpreter networks can gradually transform the low-dimensional embedding representation into implicit evaluation rules.

[0111] In the output layer of the interpreter network, the Softmax function is used to normalize the scores of implicit evaluation rules into probability distribution. The higher the probability value, the more important the corresponding implicit evaluation rule. By setting the probability threshold, the most critical implicit evaluation rules are screened out as the implicit evaluation basis for the project to be reviewed.

[0112] Through the above steps, the heterogeneous graph attention network proposed in this paper can effectively process heterogeneous meta-path sets, extract cross-modal interaction features, and obtain implicit evaluation rules for the projects to be reviewed. This method makes full use of technologies such as multi-head attention mechanism, meta-path level attention mechanism and interpreter network, and extracts discriminative and interpretable evaluation rules while capturing heterogeneous information. Compared with the traditional evaluation method based on artificial feature engineering, the method in this paper can automatically learn implicit patterns in heterogeneous networks, reduce dependence on domain knowledge, and improve the objectivity and accuracy of evaluation. At the same time, by introducing expert knowledge to guide the calculation of meta-path importance weights, this method takes into account both data-driven and knowledge-driven, and enhances the interpretability and credibility of the evaluation results. The heterogeneous graph attention network provides a novel and effective idea for mining implicit evaluation rules for projects to be reviewed, which is expected to promote the intelligence and standardization of project review.

[0113] In an optional embodiment,

[0114] The knowledge graph of the project to be reviewed and the extracted implicit evaluation rules are input into the dynamic memory network. The review knowledge related to the project to be reviewed is retrieved through the content-related addressing mechanism. The review opinions and suggestions for the project to be reviewed are aggregated through the adaptive fusion mechanism, including:

[0115] The heterogeneous knowledge graph of the project to be reviewed and the implicit evaluation rule vector extracted through the multi-layer cascaded interpreter network are input into the dynamic memory network, wherein the dynamic memory network includes an external memory matrix storing existing review knowledge, a write head component for dynamically updating the external memory matrix, and a read head component for retrieving review knowledge, wherein the read head component retrieves the review knowledge related to the project to be reviewed through a content-related addressing mechanism, and the write head component aggregates and generates an updated external memory matrix through an adaptive fusion mechanism;

[0116] For the input heterogeneous knowledge graph, the graph attention mechanism is introduced in the reading head component. The importance representation of nodes in the heterogeneous knowledge graph is learned through the attention function based on the graph structure. The node semantics are aggregated using the attention weight to obtain the project embedding vector including heterogeneous information and structural features. The cosine similarity function is used to calculate the correlation between the project embedding vector and each review knowledge vector in the external memory matrix. The review knowledge is soft-addressed according to the correlation score to obtain the n review knowledge vectors that are most relevant to the project to be reviewed. In the writing head component, the implicit evaluation rule vector is fused with the retrieved review knowledge vector through the adaptive gating mechanism. The implicit evaluation rule and the importance weight of the review knowledge are self-learned through the parameterized gating function to fuse the review information, and the fused review information is written back to the external memory matrix.

[0117] Using the memory reading method based on the key-value pair attention mechanism, the review knowledge most relevant to the query key is retrieved from the updated external memory matrix through the attention aggregation strategy, where the node features in the heterogeneous knowledge graph are used as the query key, and the vectors in the memory matrix are used as key-value pairs to obtain the context vector containing the key review information; the context vector is input into the pointer generation network for automatic generation of review opinions, and the global and local attention mechanisms are integrated in the decoding process to form review opinion suggestions for the items to be reviewed.

[0118] For example, in the technical solution proposed in this paper, in order to generate review opinions and suggestions for the project to be reviewed, it is necessary to input the knowledge graph of the project to be reviewed and the extracted implicit evaluation rules into the dynamic memory network, retrieve the review knowledge related to the project to be reviewed through the content-related addressing mechanism, and aggregate the final review opinions through the adaptive fusion mechanism. The key steps of this method will be described in detail below.

[0119] Dynamic memory network is a deep learning model based on external memory, which can dynamically store, update and retrieve knowledge. Different from traditional neural networks, dynamic memory network introduces an external memory matrix to store existing review knowledge. At the same time, dynamic memory network contains a write head component and a read head component, which are used to dynamically update the external memory matrix and retrieve review knowledge respectively.

[0120] At the input end of the dynamic memory network, the heterogeneous knowledge graph of the project to be reviewed and the implicit evaluation rule vector extracted by the multi-layer cascade interpreter network are input. The heterogeneous knowledge graph describes the semantic information and structural characteristics of the project, and the implicit evaluation rule vector contains the key rules and judgment criteria for project review.

[0121] In the reading head component, the graph attention mechanism is introduced to process heterogeneous knowledge graphs. The graph attention mechanism learns the importance representation of nodes in heterogeneous knowledge graphs through an attention function based on the graph structure. Specifically, for each node, the attention weight is calculated based on the characteristics of its neighbor nodes and the weight of the connecting edge. The higher the attention weight, the greater the contribution of the node to the item embedding vector.

[0122] The semantic features of nodes are aggregated using attention weights to obtain project embedding vectors containing heterogeneous information and structural features. The project embedding vector is a compressed representation of heterogeneous knowledge graphs, which can fully describe the key information of the project to be reviewed.

[0123] In order to retrieve the review knowledge related to the project to be reviewed, a content-related addressing mechanism is adopted. Specifically, the cosine similarity function is used to calculate the correlation between the project embedding vector and each review knowledge vector in the external memory matrix. The higher the correlation score, the higher the correlation between the review knowledge and the project to be reviewed.

[0124] According to the relevance score, the review knowledge is soft-addressed to obtain the n review knowledge vectors that are most relevant to the project to be reviewed. The soft addressing mechanism weights the review knowledge through continuous relevance scores, avoiding the information loss of the hard addressing mechanism and improving the flexibility and robustness of the retrieval.

[0125] In the writing head component, the implicit evaluation rule vector is fused with the retrieved review knowledge vector through an adaptive gating mechanism. The adaptive gating mechanism uses a parameterized gating function to self-learn the importance weights of implicit evaluation rules and review knowledge, and adaptively fuses the review information.

[0126] Specifically, the gating function receives the implicit evaluation rule vector and the review knowledge vector as input, and compresses the importance weight to between 0 and 1 through the Sigmoid activation function. The higher the importance weight, the greater the contribution of the corresponding review information to the generation of review opinions and suggestions. The parameters of the gating function are learned through the back-propagation algorithm, so that the fused review information can contain the key features and review rules of the project to be reviewed to the greatest extent.

[0127] The implicit evaluation rule vector and the review knowledge vector are weighted and summed according to the importance weights to obtain the fused review information vector. The fused review information vector contains both the implicit evaluation rules of the project to be reviewed and the historical review knowledge related to the project, which can provide strong support for generating comprehensive and accurate review opinions and suggestions.

[0128] Finally, the fused review information is written back to the external memory matrix to update the review knowledge base. By dynamically updating the external memory matrix, the dynamic memory network can continuously learn and accumulate new review knowledge and improve the quality and reliability of review opinions and suggestions.

[0129] When generating review opinions and suggestions, a memory reading method based on the key-value pair attention mechanism is used to retrieve the review knowledge most relevant to the query key from the updated external memory matrix. The node features in the heterogeneous knowledge graph are used as query keys, and the vectors in the memory matrix are used as key-value pairs. Through the attention aggregation strategy, the correlation between the query key and the key-value pair is calculated to obtain the context vector containing the key review information.

[0130] The context vector is a summary and refinement of the review knowledge in the external memory matrix, which contains the review rules, standards and suggestions most relevant to the project to be reviewed. The context vector is input into the pointer generation network to automatically generate review opinions.

[0131] The Pointer Generation Network is a text generation model based on the Seq2Seq model, which can simultaneously generate new words and copy keywords in the input. In the decoding process, the Pointer Generation Network combines the global attention mechanism and the local attention mechanism. The global attention mechanism calculates the importance of each review knowledge in the context vector according to the hidden state of the decoder and dynamically adjusts the weight of the review knowledge; the local attention mechanism focuses on the local information of the current decoding step and calculates the probability distribution of the next word according to the generated review opinion context.

[0132] By fusing global and local information, the pointer generation network can flexibly select keywords in the copy review knowledge or generate new review opinions according to the characteristics of the project to be reviewed. The decoding process is continuously iterated until a complete review opinion suggestion is generated.

[0133] Through the above steps, the dynamic memory network proposed in this paper can effectively integrate the knowledge graph of the project to be reviewed and the implicit evaluation rules, retrieve and integrate relevant review knowledge, and automatically generate comprehensive and accurate review opinions and suggestions. This method makes full use of graph attention mechanism, content-related addressing, adaptive fusion and other technologies to dynamically learn and update review knowledge while capturing project characteristics. Compared with the traditional template-based review opinion generation method, this method can adaptively generate personalized review opinions, reduce manual intervention, and improve review efficiency. At the same time, through the dynamic update of the external memory matrix, this method can continuously learn new review knowledge and continuously improve the quality and reliability of review opinions. The dynamic memory network provides a novel and effective idea for intelligent and automated project review, which is expected to promote the standardization and normalization of project review.

[0134] Figure 2 FIG. 1 is a schematic diagram of the structure of a project evaluation and review system integrating natural language processing according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0135] The first unit is used to construct a knowledge graph of the project to be reviewed based on the pre-processed structured and unstructured data of the project to be reviewed; embed the knowledge graph of the project to be reviewed into a pre-trained heterogeneous graph neural network model to obtain an output node embedding representation, obtain a graph representation of the knowledge graph of the project to be reviewed through a graph pooling layer based on the node embedding representation, input the graph representation into a graph classifier to obtain a preliminary screening result of the project to be reviewed; determine whether the project to be reviewed will enter the meeting review stage based on the preliminary screening result;

[0136] The second unit is used for, if the project to be reviewed enters the review stage, based on the knowledge graph of the project to be reviewed, according to the preset number of similar subgraphs k, obtaining k semantically similar subgraphs in the knowledge graph of historical reviewed projects that are semantically similar to the project to be reviewed through a spectral clustering algorithm; fusing the k semantically similar subgraphs to construct a heterogeneous meta-path set of the project to be reviewed, inputting the heterogeneous meta-path set into a pre-trained heterogeneous graph attention network, extracting cross-modal interaction features through a multi-head attention mechanism, and obtaining implicit evaluation rules for the project to be reviewed;

[0137] The third unit is used to input the knowledge graph of the project to be reviewed and the extracted implicit evaluation rules into the dynamic memory network, retrieve the review knowledge related to the project to be reviewed through the content-related addressing mechanism, and aggregate the review opinions and suggestions for the project to be reviewed through the adaptive fusion mechanism; input the knowledge graph of the project to be reviewed and the review opinions and suggestions into the few-shot graph learning model to generate the review results of the project to be reviewed, and dynamically adjust the few-shot graph learning model according to the review results and the feedback from the review experts.

[0138] According to a third aspect of the embodiments of the present invention,

[0139] An electronic device is provided, comprising:

[0140] processor;

[0141] a memory for storing processor-executable instructions;

[0142] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0143] A fourth aspect of the embodiments of the present invention is:

[0144] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0145] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0146] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A project evaluation and review method integrating natural language processing, characterized in that: include: Based on the pre-processed structured and unstructured data of the projects to be reviewed, a knowledge graph of the projects to be reviewed is constructed; The knowledge graph of the project to be reviewed is embedded into the pre-trained heterogeneous graph neural network model to obtain the output node embedding representation, and according to the node embedding representation, the graph representation of the knowledge graph of the project to be reviewed is obtained through the graph pooling layer, and the graph representation is input into the graph classifier to obtain the preliminary screening results of the project to be reviewed; according to the preliminary screening results, it is determined whether the project to be reviewed will enter the review stage; If the project to be reviewed enters the review phase, based on the knowledge graph of the project to be reviewed, according to the preset number of similar subgraphs k, a spectral clustering algorithm is used to obtain k semantically similar subgraphs in the knowledge graph of historical reviewed projects that are semantically similar to the project to be reviewed; the k semantically similar subgraphs are fused to construct a heterogeneous meta-path set of the project to be reviewed, and the heterogeneous meta-path set is input into a pre-trained heterogeneous graph attention network, and cross-modal interaction features are extracted through a multi-head attention mechanism to obtain implicit evaluation rules for the project to be reviewed; The knowledge graph of the project to be reviewed and the extracted implicit evaluation rules are input into the dynamic memory network, and the review knowledge related to the project to be reviewed is retrieved through the content-related addressing mechanism, and the review opinions and suggestions for the project to be reviewed are aggregated through the adaptive fusion mechanism; Input the knowledge graph of the project to be reviewed and the review opinions and suggestions into the few-shot graph learning model to generate the review results of the project to be reviewed, and dynamically adjust the few-shot graph learning model according to the review results and the feedback from the review experts; The knowledge graph of the project to be reviewed and the extracted implicit evaluation rules are input into the dynamic memory network. The review knowledge related to the project to be reviewed is retrieved through the content-related addressing mechanism. The review opinions and suggestions for the project to be reviewed are aggregated through the adaptive fusion mechanism, including: The heterogeneous knowledge graph of the project to be reviewed and the implicit evaluation rule vector extracted through the multi-layer cascaded interpreter network are input into the dynamic memory network, wherein the dynamic memory network includes an external memory matrix storing existing review knowledge, a write head component for dynamically updating the external memory matrix, and a read head component for retrieving review knowledge, wherein the read head component retrieves the review knowledge related to the project to be reviewed through a content-related addressing mechanism, and the write head component aggregates and generates an updated external memory matrix through an adaptive fusion mechanism; For the input heterogeneous knowledge graph, the graph attention mechanism is introduced in the reading head component. The importance representation of nodes in the heterogeneous knowledge graph is learned through the attention function based on the graph structure. The node semantics are aggregated using the attention weight to obtain the project embedding vector including heterogeneous information and structural features. The cosine similarity function is used to calculate the correlation between the project embedding vector and each review knowledge vector in the external memory matrix. The review knowledge is soft-addressed according to the correlation score to obtain the n review knowledge vectors that are most relevant to the project to be reviewed. In the writing head component, the implicit evaluation rule vector is fused with the retrieved review knowledge vector through the adaptive gating mechanism. The implicit evaluation rule and the importance weight of the review knowledge are self-learned through the parameterized gating function to fuse the review information, and the fused review information is written back to the external memory matrix. Using the memory reading method based on the key-value pair attention mechanism, the most relevant review knowledge to the query key is retrieved from the updated external memory matrix through the attention aggregation strategy, where the node features in the heterogeneous knowledge graph are used as the query key, and the vectors in the memory matrix are used as the key-value pairs to obtain the context vector containing the key review information; the context vector is input into the pointer generation network to automatically generate the review opinions, and the global and local attention mechanisms are integrated in the decoding process to form the review opinion suggestions for the project to be reviewed; The knowledge graph of the project to be reviewed and the review opinions and suggestions are input into the few-shot graph learning model to generate the review results of the project to be reviewed, including: The heterogeneous knowledge graph of the project to be reviewed and the review opinions and suggestions are input into the few-shot graph learning model. The few-shot graph learning model adopts a meta-learning paradigm to construct a multi-granularity neighbor aggregation mechanism for the heterogeneous knowledge graph, wherein a heterogeneous neighbor aggregation function is used at the node level to aggregate the embedding representations of different types of neighbor nodes through an attention network and obtain a node-level embedding vector. At the semantic level, the semantic evolution mechanism of the graph isomorphic network is used to iteratively aggregate the semantic representations of multi-hop neighbors through a polling mechanism and obtain a semantic-level embedding vector. The node-level embedding vector is fused with the semantic-level embedding vector to obtain a final node embedding vector. The pre-trained knowledge distillation model compresses the review opinions and suggestions into a semantic latent space, models the multi-level interaction between the semantic latent vector and the node embedding vector through a dynamic routing mechanism, uses the attention-aware gating unit to adjust the importance of information of different granularities, and updates the node embedding vector; in the twin network, the historical project knowledge graph of the support set is used to learn the cluster centroid as the prototype vector of the category of the review result, and the discriminative graph-level representation vector is extracted by alternately optimizing the weights of the graph attention network and the twin network; Calculate the similarity between the graph-level representation vector and the prototype vector of each category in Euclidean space and semantic space, filter out noise by setting a distance threshold, and take the category corresponding to the prototype vector with the highest similarity as the predicted category of the project to be reviewed; generate multiple augmented views of the knowledge graph of the project to be reviewed from three granularities: node level, edge level, and graph level, determine a composite optimization target including the contrast center loss of the twin network and the multi-view consistency loss, optimize the composite optimization target using a meta-learning algorithm, input the graph-level representation vector, the predicted category, and the node embedding representation into a review report generation module, and output a review result report including the project comprehensive score, prediction confidence, and key node explanation through visualization technology.

2. The method according to claim 1, characterized in that: Based on the pre-processed structured and unstructured data of the projects to be reviewed, a knowledge graph of the projects to be reviewed is constructed, including: The preprocessed structured data and unstructured data are aligned using a multi-strategy entity alignment method that combines string matching, word vector similarity calculation, rule matching and transfer learning, and the target strategy is selected to map entities in different source data to unique nodes in the knowledge graph; among them, string matching is implemented using the edit distance algorithm and Jaccard similarity based on part of speech and word form, word vector similarity calculation is implemented using a word vector model that integrates dictionary tree index and a transformation model represented by a bidirectional encoder, rule matching is implemented using a knowledge-driven regular expression generation method and a conditional random field model optimized by reinforcement learning, and transfer learning is implemented using a cross-domain entity alignment model based on a graph attention network; Extract semantic relationships between entities from aligned unstructured text data, and organize the extracted entities and semantic relationships into semantically similar subgraphs including project portrait subgraph, technology route subgraph, team capability subgraph and resource demand subgraph according to the multi-level concept hierarchy and fine-grained relationship types defined by the fusion domain ontology and industry standards; among them, the project portrait subgraph adopts dynamic heterogeneous graph neural network to model the evolution of project attributes, the technology route subgraph adopts citation network to determine the evolution of technology topics, the team capability subgraph adopts temporal multi-view learning method to represent the dynamic changes of team capabilities, and the resource demand subgraph adopts knowledge reasoning and reinforcement learning method to optimize resource allocation; An improved transfer embedding knowledge representation learning algorithm and a heterogeneous graph neural network algorithm are applied to each semantically similar subgraph for representation learning. The entity and semantic relationship vectors obtained by representation learning are used to calculate the similarity between different semantically similar subgraphs through a preset knowledge graph alignment method. The head entity, semantic relationship and constraint rules are used as input, and the missing implicit associations between semantically similar subgraphs are predicted through multi-hop relationship path reasoning. Confidence estimation and Markov logic network are used to verify and optimize the implicit associations predicted by reasoning, and the associated triples are obtained and added to the knowledge base to obtain the knowledge graph of the project to be reviewed.

3. The method according to claim 2, characterized in that Taking the head entity, semantic relations and constraint rules as input, the missing implicit associations between semantically similar subgraphs are predicted through multi-hop relational path reasoning, including: Calculate the attention weight between the head entity embedding and all outgoing edge relationship embeddings, select the top M outgoing edge relationships most relevant to the head entity according to the attention weight, and use the head entity node and the selected outgoing edge relationship as the starting point of the current reasoning path; the attention weight is calculated by the dot product of the head entity embedding and the outgoing edge relationship embedding and normalized; According to the last relation node of the current reasoning path, a relation-aware entity neighbor subgraph is constructed, which includes the entity nodes directly connected to the last relation and their one-hop neighbors; the entity neighbor subgraph is encoded using a graph attention network to obtain the hidden state representation of the entity node; each newly generated reasoning path is sequence modeled using a gated recurrent unit network, and the entity nodes and relations in the path are embedded in the sequence into the gated recurrent unit network to obtain the hidden state representation of the reasoning path; finally, the hidden state of the path and the hidden state of the last entity neighbor subgraph are aggregated through the attention mechanism to predict the next-hop relationship; Iteratively execute the above steps of constructing a relationship-aware entity neighbor subgraph and subsequent steps according to the last relationship node of the current reasoning path until the preset hop count threshold is reached or the potential tail entity node is reached, and multiple candidate reasoning paths are generated; a multi-path exploration strategy based on beam search is adopted to retain the N relationship paths with the highest confidence from the candidate reasoning paths; wherein the confidence is determined based on the reasoning path length, the prediction probability of each hop and the semantic matching degree of the node relationship; post-processing and optimizing the candidate association paths, using the learned path constraint rules to filter and reorder the candidate paths, filter out the paths that do not meet the rule constraints, and obtain the final cross-subgraph association path set as the missing implicit association between semantically similar subgraphs.

4. The method according to claim 1, characterized in that The k semantically similar subgraphs are integrated to construct a heterogeneous meta-path set of the project to be reviewed, including: Based on graph convolutional neural network, low-dimensional dense feature vector representation of nodes in semantically similar subgraphs is extracted. The semantic similarity of isomorphic nodes between different semantically similar subgraphs is calculated through cosine similarity function. Isomorphic nodes with semantic similarity exceeding the preset similarity threshold and the same node type are merged. The node feature vector is updated on the basis of retaining the original attribute information of the node to obtain the fused heterogeneous knowledge graph. Applying a meta-path-constrained random walk algorithm on the fused heterogeneous knowledge graph to generate a node sequence containing heterogeneous information, constructing a transition probability matrix according to three predefined meta-path modes: project-personnel-project, project-institution-project, and project-research field-project, calculating the transition probability from the current node to different types of neighboring nodes, and adjusting the random degree and meta-path preference weight in the walk strategy in combination with the node concentration and web page ranking value; The principal component analysis method is used to reduce the dimension of the node sequence, and the essential pattern of the node sequence is extracted. The frequent pattern mining algorithm is used to identify the node subsequences that conform to the three meta-path patterns from the compressed node sequence. The minimum length of the subsequence and the node heterogeneity are controlled by setting the support and confidence thresholds of the meta-path matching to generate candidate meta-path instances. The quality of the candidate meta-path instances is evaluated based on the utilization rate of heterogeneous information and structural stability, and the m meta-path instances with the best quality are screened out to form the heterogeneous meta-path set of the project to be reviewed.

5. The method according to claim 4, characterized in that The heterogeneous meta-path set is input into the pre-trained heterogeneous graph attention network, and the cross-modal interaction features are extracted through the multi-head attention mechanism to obtain the implicit evaluation rules of the items to be reviewed, including: The heterogeneous meta-path set is input into the pre-trained heterogeneous graph attention network, which uses a multi-head attention mechanism to update the node embedding representation of each layer of the network, calculates the correlation between the neighbor nodes and the target node through the attention function based on the scaled dot product attention to obtain the attention score, and uses a multi-mask mechanism to filter the attention score, retain some neighbor nodes, and generate node-level attention weights; A meta-path attention layer is set at the last layer of the heterogeneous graph attention network. The meta-path level attention mechanism based on the gating mechanism is used to calculate the embedding representation vector of each meta-path instance. The meta-path level attention mechanism controls the matching degree between the node embedding representation contained in the meta-path and the query vector through the gating unit, and generates the meta-path level attention weight based on the heterogeneity and centrality characteristics of the nodes. According to the prior knowledge of the meta-path type, the meta-path importance weight is calculated based on the frequency, coverage, and uniqueness indicators. The meta-path importance weights are normalized, and the hierarchical analysis method is introduced in combination with expert knowledge to determine the upper and lower bounds of the weights of different meta-path types. The meta-path embedding representations output by the heterogeneous graph attention network are weighted summed according to the normalized meta-path importance weights, and the sequential dependencies between meta-path embeddings are modeled using a bidirectional gated recurrent unit network to obtain a low-dimensional embedding representation vector of the project to be reviewed that integrates local semantics, structural features and cross-modal association information. The implicit evaluation rules in the low-dimensional embedding representation vector are determined through a multi-layer cascaded interpreter network.

6. A project evaluation and review system integrating natural language processing, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to construct a knowledge graph of the project to be reviewed based on the pre-processed structured and unstructured data of the project to be reviewed; The knowledge graph of the project to be reviewed is embedded into the pre-trained heterogeneous graph neural network model to obtain the output node embedding representation, and according to the node embedding representation, the graph representation of the knowledge graph of the project to be reviewed is obtained through the graph pooling layer, and the graph representation is input into the graph classifier to obtain the preliminary screening results of the project to be reviewed; according to the preliminary screening results, it is determined whether the project to be reviewed will enter the review stage; The second unit is used for, if the project to be reviewed enters the review stage, based on the knowledge graph of the project to be reviewed, according to the preset number of similar subgraphs k, obtaining k semantically similar subgraphs in the knowledge graph of historical reviewed projects that are semantically similar to the project to be reviewed through a spectral clustering algorithm; fusing the k semantically similar subgraphs to construct a heterogeneous meta-path set of the project to be reviewed, inputting the heterogeneous meta-path set into a pre-trained heterogeneous graph attention network, extracting cross-modal interaction features through a multi-head attention mechanism, and obtaining implicit evaluation rules for the project to be reviewed; The third unit is used to input the knowledge graph of the project to be reviewed and the extracted implicit evaluation rules into the dynamic memory network, retrieve the review knowledge related to the project to be reviewed through the content-related addressing mechanism, and aggregate the review opinions and suggestions of the project to be reviewed through the adaptive fusion mechanism; The knowledge graph of the project to be reviewed and the review opinions and suggestions are input into the few-shot graph learning model to generate the review results of the project to be reviewed, and the few-shot graph learning model is dynamically adjusted according to the review results and the feedback from the review experts.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.

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

Citation Information

Patent Citations

  • Government affair informatization project multi-expert opinion machine integration method based on knowledge fusion

    CN113255364A

  • Project review method and system

    CN116703328A

  • Project review expert screening method and system based on knowledge graph updating and medium

    CN116756373A

  • Project review expert discovery and recommendation system based on knowledge graph

    CN117909509A

  • Medical plan recommendation system and method based on knowledge graph representation learning

    WO2021189971A1