Local quotation recommendation method and system based on quotation statement heterogeneous network

Abstract

The invention provides a local quotation recommendation method and system based on a quotation statement heterogeneous network, and relates to the technical field of information recommendation, and the method comprises the steps: carrying out the data preprocessing and quotation relation analysis of a quotation statement sentence in a paper according to a preset measurement index and a preset semantic feature of local quotation recommendation, and obtaining a quotation statement statement; constructing a citation statement heterogeneous network; and for the citation statement heterogeneous network, performing node low-dimensional embedded learning, acquiring a citation relation path instance based on reinforcement learning, generating a citation relation meta-path diagram, performing citation recommendation and generating a recommended literature list. The technical problem of poor paper recommendation effect caused by lack of citation context information in a citation embedding network in the prior art is solved, and the technical effect of improving the citation recommendation accuracy and recommendation quality is achieved by constructing a citation statement heterogeneous network and accurately identifying the deep semantic relevance between citations.

Description

Technical Field

[0001] This invention relates to the field of information recommendation technology, specifically to a local citation recommendation method and system based on a citation statement heterogeneous network. Background Technology

[0002] With the advent of the information age, the number of academic research documents has exploded, posing researchers with a massive amount of information and a complex search environment. To improve research efficiency, literature recommendation systems have emerged, especially citation-based local citation recommendation methods, which have played a crucial role in helping researchers quickly obtain relevant literature.

[0003] Although existing works have utilized citation context and graph convolutional neural networks for citation recommendation, the separation of citation context information from the citation network and the lack of citation context information in the citation embedding network have resulted in poor paper recommendation performance. Summary of the Invention

[0004] This application provides a local citation recommendation method and system based on a citation statement heterogeneous network, which addresses the technical problem in the prior art where the separation of citation context information and citation network, and the lack of citation context information in the citation embedding network, leads to poor paper recommendation performance.

[0005] In view of the above problems, this application provides a local citation recommendation method and system based on citation statement heterogeneous networks.

[0006] Firstly, this application provides a local citation recommendation method based on a heterogeneous citation statement network, implemented through a local citation recommendation system based on a heterogeneous citation statement network. The method includes: preprocessing and analyzing citation relationships of citation statements in a paper based on preset quantitative indicators and preset semantic features for local citation recommendation, constructing a heterogeneous citation statement network; performing low-dimensional node embedding learning on the heterogeneous citation statement network, obtaining citation relationship path instances based on reinforcement learning, generating a citation relationship meta-path graph, performing citation recommendation, and generating a list of recommended documents.

[0007] Secondly, this application also provides a local citation recommendation system based on a heterogeneous citation statement network, used to execute a local citation recommendation method based on a heterogeneous citation statement network as described in the first aspect, comprising: a network construction module, used to perform data preprocessing and citation relationship analysis on citation statement sentences in a paper according to preset quantitative indicators and preset semantic features of local citation recommendation, and construct a heterogeneous citation statement network; and a citation recommendation module, used to perform low-dimensional node embedding learning on the heterogeneous citation statement network, obtain citation relationship path instances based on reinforcement learning, generate a citation relationship meta-path graph, perform citation recommendation, and generate a list of recommended documents.

[0008] The technical solution provided in this application has at least the following beneficial effects:

[0009] Based on preset metrics and semantic features for local citation recommendations, data preprocessing and citation relationship analysis are performed on citation statements in papers to construct a citation statement heterogeneous network. For this network, low-dimensional embedding learning of nodes is performed, and citation relationship path instances are obtained based on reinforcement learning to generate a citation relationship meta-path graph for citation recommendation, resulting in a recommended literature list. By constructing a citation statement heterogeneous network, utilizing graph neural networks for low-dimensional embedding learning of nodes, and combining reinforcement learning for multi-hop path reasoning, the deep semantic connections between citations are accurately identified, recommending high-value, semantically relevant literature, thus improving the accuracy and quality of citation recommendations. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0011] Figure 1 This is a flowchart illustrating a local citation recommendation method based on a heterogeneous citation statement network according to this application.

[0012] Figure 2 This is a schematic diagram of the structure of a local citation recommendation system based on a citation statement heterogeneous network according to this application.

[0013] Figure labeling: Network building module 11, citation recommendation module 12. Detailed Implementation

[0014] This application provides a local citation recommendation method and system based on a heterogeneous citation statement network to address the technical problem in existing technologies where the separation of citation context information and the citation network, coupled with the lack of citation context information in the citation embedding network, leads to poor paper recommendation performance. By constructing a heterogeneous citation statement network and utilizing graph neural networks for low-dimensional embedding learning of nodes, combined with reinforcement learning for multi-hop path reasoning, the method accurately identifies deep semantic relationships between citations, recommending high-value, semantically relevant documents, thereby improving the accuracy and quality of citation recommendation.

[0015] Example 1, as Figure 1 As shown, this application provides a local citation recommendation method based on a citation statement heterogeneous network, which specifically includes the following steps:

[0016] Based on the pre-defined metrics and semantic features recommended by local citations, data preprocessing and citation relationship analysis are performed on the citation statements in the paper to construct a heterogeneous network of citation statements.

[0017] Specifically, constructing a heterogeneous citation statement network first requires relying on pre-defined metrics and semantic features for local citation recommendations. This process aims to transform citation data into a structured network format for subsequent link prediction and recommendation tasks. First, the collection of pre-defined metrics is fundamental to building the heterogeneous network. Common metrics include the author's H-index, the source's impact factor, and citation count. These metrics help understand the influence, research field, and academic recognition of each document. For example, analyzing a paper's citation count can determine its academic influence; while the research topic helps identify the specific field covered by the paper, thus providing semantic context for subsequent citation recommendations. These designed metrics provide the necessary information for constructing the heterogeneous citation statement network through quantitative analysis.

[0018] Next, pre-defined semantic features are used to capture the semantic relationships involved in the citations. These semantic features can be textual information such as titles, abstracts, keywords, and citation statements. For example, the title of a document can intuitively reflect its research content, while the abstract further provides the research background, methods, and conclusions. Therefore, by extracting and processing this textual information, semantic connections between documents can be established, providing more accurate semantic support for recommendation tasks.

[0019] Data preprocessing is a crucial step in constructing heterogeneous networks. This process involves extracting and standardizing citations from papers to facilitate citation relationship analysis. For example, quoted statements throughout the text need to be extracted and converted to a standardized format. Standardizing references using DOI numbers not only ensures the traceability of citations within the network but also facilitates the association of citation relationships between different documents. This standardization process helps eliminate discrepancies in document wording, ensuring the accuracy of the recommendation system.

[0020] Once all relevant information has been collected and standardized, the next task is to construct a heterogeneous citation statement network by analyzing citation relationships and combining pre-defined semantic features and metrics. This network will include different types of nodes, such as paper, title, author, publication year, and citation statement. Each node type represents an element in the citation network, and the relationships between nodes are connected by edge types such as containment, citation, and association. For example, paper nodes and title nodes are connected by containment relationships, while paper nodes and citation statement nodes are connected by citation relationships. Through this diverse combination of nodes and edges, a heterogeneous citation statement network containing rich semantic information is constructed.

[0021] Through these steps, the resulting heterogeneous citation statement network not only contains semantic associations between documents but also reveals citation relationships between different documents, providing a solid foundation for subsequent reinforcement learning-based link prediction and citation recommendation.

[0022] For the aforementioned heterogeneous citation statement network, low-dimensional node embedding learning is performed, citation relationship path instances are obtained based on reinforcement learning, a citation relationship meta-path graph is generated, citation recommendation is performed, and a recommended literature list is generated.

[0023] Specifically, firstly, a graph neural network (GCN) model is used to perform low-dimensional embedding learning on the nodes in the citation statement heterogeneous network, generating a semantic representation for each node. In the low-dimensional space, each node's representation contains semantic information, including multi-dimensional information such as citation statement nodes, paper nodes, and author nodes. These embedding representations can effectively capture the semantic associations in the citation statement heterogeneous network. Further, reinforcement learning explores the nodes and relationships in the citation statement heterogeneous network, performing multi-hop path reasoning starting from the citation statement nodes. Each citation statement node acts as a source node in the network, and a reinforcement learning agent makes decisions starting from that node, aiming to select the optimal path and generate citation relationship links. The reinforcement learning agent learns how to reach the target node from the citation statement node by continuously observing the current state, performing actions, and receiving feedback. This process follows a Markov decision process. Through this path reasoning, the reinforcement learning algorithm can not only identify target nodes but also explore potential citation relationships between documents. The path instances obtained through reinforcement learning are used to generate citation relationship meta-paths. In the network, each path instance can be viewed as a path composed of multiple nodes and edges. By summarizing these paths, meta-paths with structural features are identified. These meta-paths can establish connections between different citation statement nodes, providing an operational graph structure for subsequent recommendation tasks. The goal of this process is to transform the complex heterogeneous citation statement network into multiple new graphs defined by meta-paths, making these new graphs more structured and facilitating subsequent learning and prediction.

[0024] In the aforementioned steps, based on the Graph Neural Network (GCN) model, the semantic representation of a node is achieved by embedding the features of each node into a low-dimensional vector space. This low-dimensional vector not only contains the node's basic features (such as the DOI, title, author, and abstract of a document), but also incorporates the node's relationship information with other nodes in the network. These semantic representations are propagated through the network via convolution operations, gradually aggregating information from neighboring nodes. This ensures that the semantic representation of each node accurately reflects its position and role in the entire network, thereby enabling citation recommendations based on the user's query starting point—the citation portion of a document—and generating a recommended document list.

[0025] When generating the recommended literature list, not only are the direct citation relationships between nodes considered, but the global structure of the citation network is also taken into account. The semantic representation of the nodes is used to delve into the potential connections between the literature. In this way, accurate citation recommendations can be provided.

[0026] Furthermore, this application also includes the following steps: extracting bibliometric indicators and semantic features from the citation recommendation task based on preset bibliometric indicators and preset semantic features; standardizing the cited papers contained in the citation statements identified in the papers into a DOI structured form; constructing network nodes with multiple types of nodes based on the DOI structured form; connecting nodes with relation types; reading the bibliometric indicators corresponding to each network node for node annotation; and generating the citation statement heterogeneous network.

[0027] Specifically, based on pre-defined bibliometric indicators and semantic features, the first step is to extract bibliometric indicators and semantic features from the literature. These indicators and features provide the necessary foundational data for constructing a heterogeneous citation statement network. The cited papers contained in the identified citation statements are then standardized into a DOI structured format. Specifically, citation statements are extracted from the paper; these are the parts of the paper that explicitly cite other literature, such as "As stated in reference X...". To standardize these citations, the citation information in the text needs to be converted into a unified DOI structured format. This standardization ensures consistency between citations, allowing each citation to accurately point to a specific reference, unaffected by differences in textual expression.

[0028] The next task is to construct a heterogeneous citation statement network using the DOI structure, which includes various types of nodes and different types of relation edges. Specifically, paper nodes represent each document in the network, containing its core attributes such as title, author, abstract, and publication year. Citation statement nodes represent sentences or paragraphs in the document that contain citations; each citation statement node is associated with a specific document DOI and points to the cited document. Author nodes represent the authors of the document and are usually associated with specific paper nodes. Journal or publication nodes represent the journal or publication to which each document belongs; journal nodes are connected to paper nodes through publication relationships. Paper nodes are connected to citation statement nodes through citation relationships; each paper's citation statement points to the document it cites. Relationship types include inclusion relationships and publication relationships.

[0029] When constructing heterogeneous networks, bibliometric indicators (such as H-index and journal impact factor) and semantic features (such as research topic and keywords) are labeled as additional attributes of nodes. These attributes further enrich the semantic information of the network, making it not only limited to connections but also encompassing the academic influence and research field of the documents. Based on DOI-standardized citation information, citation relationships between papers are established through citation statement nodes. Each citation statement node not only connects to the cited document but also establishes connections through other relevant nodes, such as author nodes and journal nodes, forming a complex citation network structure. Through the construction of this network, the complex relationships between documents are systematically organized into a multi-level, multi-dimensional structure. This enables the system to recommend the most relevant documents based on the information of nodes and edges in the network, combined with the semantic content and academic value of the documents.

[0030] Furthermore, this application also includes the following steps: using a graph convolutional neural network model to map the nodes in the metapath graph to a low-dimensional vector space to generate a semantic representation of each node; using reinforcement learning to perform multi-hop inference on the citation statement heterogeneous network to obtain citation relationship path instances; finding the lowest common ancestor from the citation relationship path instances to generate citation relationship metapaths, and converting the citation statement heterogeneous network into multiple new graphs defined by metapaths to generate a metapath graph.

[0031] Specifically, a Graph Convolutional Neural Network (GCN) model is used to learn low-dimensional embeddings for the nodes in these meta-path graphs. The GCN model utilizes the structural information of the graph to map the features of each node into a low-dimensional space, generating embedded representations of the nodes. These low-dimensional vectors effectively preserve the semantic information of the nodes and the relationships between them. During the embedding learning process, the GCN model fuses the information of each node's neighbors in the graph through convolutional operations, ensuring that each node's representation not only reflects its own features but also incorporates its contextual information within the network. In this way, each node in the meta-path graph is mapped to a semantically rich low-dimensional vector that contains both structural information and semantic features.

[0032] The semantic representation of each node obtained in this way provides an important basis for subsequent citation relationship prediction. Next, combining the semantic representation of each node, a reinforcement learning agent explores the entire network starting from the citation statement node. Through a series of states, actions, transitions, and feedback, the agent learns how to choose the optimal path. This process involves jumping from one node to another in the network, with each jump making a decision based on the network's structural information and the current state, until the target node is finally reached. This multi-hop reasoning approach allows the agent to gradually accumulate citation relationship information, thereby obtaining a complete citation relationship path instance. These path instances demonstrate multiple jump paths from the source node (citation statement node) to the target node, containing rich citation relationship information. Citation relationship path instances are directed acyclic paths (DAGs) from the citation statement node to the target paper node, passing through multiple intermediate nodes to finally reach the target paper node. These path instances represent citation relationship links from a citation statement to the actual literature; each path reflects the citation order between documents and can effectively reveal the academic connections between them. This path is directed acyclic, meaning that the direction of jump between each node is clear and will not form a loop. For example, starting from a citation statement node, the agent can choose to jump to a paper node, then jump to another paper node through the citation relationship, then jump to the corresponding author node, and finally return to another paper node. This process ensures the directionality and acyclicity of the path, avoids circular references, and thus maintains the logical consistency and structural clarity of the network.

[0033] After obtaining the citation relationship path instances, the next step is to find the lowest common ancestor among these path instances. The common ancestor is the earliest common connecting node among all path instances; it is a shared node across multiple paths and reflects the most fundamental connection between citations within the path. By identifying these common ancestors, we can summarize general and representative citation relationship meta-paths. These meta-paths are patterns composed of various nodes and relationship types in the graph. They not only define the structural relationships between nodes but also represent common citation patterns in citation networks. For example, a meta-path might include "Paper Node → Citation Statement Section → Paper Node," representing mutual citation relationships between documents.

[0034] Once the citation relationship metapaths are generated, the entire citation statement heterogeneous network can be transformed into multiple new graphs defined by these metapaths. These new graphs are more structured than the original network because they abstract and simplify node relationships through metapaths. The generated metapath graphs are simplified versions of the original citation statement heterogeneous network, with nodes and relationships in each graph based on the metapath definitions to facilitate subsequent computation and processing. The information contained in these graphs is crucial for citation recommendation tasks; they represent citation and relationship patterns between different types of documents and are key to the model's learning of node representations.

[0035] Furthermore, this application also includes the following steps: constructing a reinforcement learning agent, including a state, action, transition, and reward mechanism, wherein the state represents the current node type and its set of neighboring nodes, the action is a citation transition operation from the current node to the target node, and the reward is positive feedback when the citation relationship is correctly predicted; based on the reinforcement learning agent, performing multi-hop path reasoning from the citation statement node in the citation statement heterogeneous network to obtain a directed acyclic path from the citation statement node to the target paper node, which serves as an instance of the citation relationship path.

[0036] Specifically, in citation-state heterogeneous networks, the construction of reinforcement learning agents is the core of multi-hop path reasoning, aiming to generate accurate citation relationship path instances through learning the decision-making process. To enable reinforcement learning agents to perform path reasoning efficiently, clear state, action, transition, and reward mechanisms need to be designed, allowing the agent to explore the correct citation paths in the citation-state heterogeneous network.

[0037] State representation is a core component of reinforcement learning, describing the agent's current environment. In a citation statement heterogeneous network, the state includes the current node type and its set of neighboring nodes. The current node type reflects the type of node the agent is in, such as a citation statement node, paper node, author node, etc., each with different characteristics and associated information. The set of neighboring nodes includes all nodes connected to the current node; for example, starting from a citation statement node, the agent can jump to related paper nodes, author nodes, or journal nodes. Through state representation, the agent can understand its network position and the possible transition paths. Actions are the decision-making component in reinforcement learning, referring to the actions taken by the agent based on its current state. In citation statement heterogeneous networks, actions typically manifest as jumping from the current node to a target node. Each jump operation selects an appropriate edge type based on the network structure, such as jumping from a citation statement node to a paper node, from a paper node to an author node, or from a paper node to a journal node. The criterion for action selection is based on the relationship between the current node and its neighboring nodes. The agent explores these relationships to determine the optimal target for the path jump. Each action execution changes the agent's state, leading it to a new node and laying the foundation for subsequent path reasoning. The transition mechanism describes the process of state transitions. Each time an action is executed, the agent jumps from the current node to the target node, completing a state transition. In citation statement heterogeneous networks, transitions are achieved through different types of relationships within the network. For example, jumping from a citation statement node to a paper node indicates that the paper is cited in a citation; jumping from a paper node to an author node indicates that the paper was written by a particular author. Each transition establishes a new connection in the network and continuously propels the agent towards the final target state.

[0038] The reward mechanism, the feedback component in reinforcement learning, guides the agent's learning process. In citation recommendation tasks, the design of the reward mechanism is crucial, as it measures the correctness of the agent's decisions during path reasoning. In this task, rewards are typically related to the correctness of citation relationships. If the agent derives the correct citation relationship path through reasoning—that is, correctly infers a valid citation chain between documents from citation statement nodes—then the agent receives a positive feedback reward, encouraging it to continue taking similar decision paths. Conversely, if the agent's predicted citation relationship is invalid or incorrect, the system provides negative feedback, prompting the agent to adjust its strategy and avoid similar errors. Through repeated learning and adjustment, the agent gradually learns how to select appropriate paths in a heterogeneous network of citation statements.

[0039] Based on these state, action, transition, and reward mechanisms, reinforcement learning agents can perform multi-hop path reasoning in heterogeneous citation statement networks. Starting from citation statement nodes, they traverse multiple hops to ultimately generate valid citation relationship path instances. These path instances not only demonstrate the citation relationships between documents but also provide accurate evidence for subsequent citation recommendations. By continuously optimizing its decision-making strategy, the agent can eventually identify high-quality citation paths in the network, thereby improving the accuracy and relevance of the recommendation system.

[0040] Furthermore, performance testing of citation link prediction based on citation statement heterogeneous networks can be performed by constructing a validation dataset. The construction of the validation dataset relies on real citation data, typically extracted from known citation relationships in academic paper databases. This data should include citation links between documents, citation identifiers, DOI information, author and journal details, and other relevant features. During construction, the validation dataset should include positive examples (genuine citation relationships) and negative examples (incorrect or non-existent citation relationships). This data will be used to test the system's recommendation accuracy. Once the validation dataset is constructed, the next step is to use it to test the citation link prediction method based on citation statement heterogeneous networks, calculating precision, recall, accuracy, and the harmonic mean of precision and recall. These metrics provide a comprehensive evaluation of prediction performance. In addition to the aforementioned conventional evaluation metrics, ranking-related metrics (such as average rank and precision-recall curves) can be combined to further optimize the recommendation system when performing citation recommendation performance evaluation.

[0041] In summary, the local citation recommendation method based on a heterogeneous citation statement network provided in this application has the following beneficial effects:

[0042] Based on preset metrics and semantic features for local citation recommendations, data preprocessing and citation relationship analysis are performed on citation statements in papers to construct a citation statement heterogeneous network. For this network, low-dimensional embedding learning of nodes is performed, and citation relationship path instances are obtained based on reinforcement learning to generate a citation relationship meta-path graph for citation recommendation, resulting in a recommended literature list. By constructing a citation statement heterogeneous network, utilizing graph neural networks for low-dimensional embedding learning of nodes, and combining reinforcement learning for multi-hop path reasoning, the deep semantic connections between citations are accurately identified, recommending high-value, semantically relevant literature, thus improving the accuracy and quality of citation recommendations.

[0043] Example 2, based on the same inventive concept as the local citation recommendation method based on a heterogeneous citation statement network in the foregoing examples, please refer to the appendix. Figure 2 This application also provides a local citation recommendation system based on a citation statement heterogeneous network, comprising:

[0044] Network construction module 11 is used to perform data preprocessing and citation relationship analysis on citation statements in the paper based on preset quantitative indicators and preset semantic features recommended by local citations, and to construct a heterogeneous network of citation statements.

[0045] The citation recommendation module 12 is used to perform low-dimensional embedding learning of nodes for the heterogeneous citation statement network, obtain citation relationship path instances based on reinforcement learning, generate a citation relationship meta-path graph, perform citation recommendation, and generate a list of recommended documents.

[0046] Furthermore, the local citation recommendation system based on a heterogeneous citation statement network is also used to: extract bibliometric indicators and semantic features from the citation recommendation task according to preset bibliometric indicators and preset semantic features; standardize the cited papers contained in the citation statement sentences identified in the papers into a DOI structured form; construct network nodes with multiple types of nodes based on the DOI structured form; connect nodes with relation types; read the bibliometric indicators corresponding to each network node for node annotation; and generate the heterogeneous citation statement network.

[0047] Furthermore, the local citation recommendation system based on a citation statement heterogeneous network is also used to: map the nodes in the metapath graph to a low-dimensional vector space using a graph convolutional neural network model to generate a semantic representation of each node; perform multi-hop inference on the citation statement heterogeneous network using reinforcement learning to obtain citation relationship path instances; find the lowest common ancestor from the citation relationship path instances to generate citation relationship metapaths; and convert the citation statement heterogeneous network into multiple new graphs defined by metapaths to generate a metapath graph.

[0048] Furthermore, the local citation recommendation system based on a citation statement heterogeneous network is also used to: construct a reinforcement learning agent, including a state, action, transition, and reward mechanism, wherein the state represents the current node type and its set of neighboring nodes, the action is a citation transition operation from the current node to the target node, and the reward is positive feedback when the citation relationship is correctly predicted; based on the reinforcement learning agent, multi-hop path reasoning is performed from the citation statement node in the citation statement heterogeneous network to obtain a directed acyclic path from the citation statement node to the target paper node, which serves as an instance of the citation relationship path.

[0049] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The local citation recommendation method and specific examples based on citation statement heterogeneous networks in the foregoing embodiment one are also applicable to the local citation recommendation system based on citation statement heterogeneous networks in this embodiment. Through the foregoing detailed description of the local citation recommendation method based on citation statement heterogeneous networks, those skilled in the art can clearly understand the local citation recommendation system based on citation statement heterogeneous networks in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0050] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0051] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of this application and its equivalents, this application also intends to include such modifications and variations.

Claims

1. A local citation recommendation method based on a citation statement heterogeneous network, characterized in that, include: Based on the pre-defined metrics and semantic features recommended by local citations, data preprocessing and citation relationship analysis are performed on the citation statements in the paper to construct a heterogeneous network of citation statements. For the aforementioned heterogeneous citation statement network, low-dimensional node embedding learning is performed, citation relationship path instances are obtained based on reinforcement learning, a citation relationship meta-path graph is generated, citation recommendation is performed, and a recommended literature list is generated.

2. The local citation recommendation method based on a citation statement heterogeneous network as described in claim 1, characterized in that, Based on the pre-defined metrics and semantic features recommended by local citations, the citation statements in the paper are preprocessed and citation relationships are analyzed to construct a heterogeneous network of citation statements, including: Based on preset bibliometric indicators and preset semantic features, bibliometric indicators and semantic features are extracted from the citation recommendation task, and the cited papers contained in the citation statements identified in the paper are standardized into DOI structured form. Based on the DOI structure, network nodes are constructed using various node types, nodes are connected using relation types, and the bibliometric indicators corresponding to each network node are read for node annotation to generate the citation statement heterogeneous network.

3. The local citation recommendation method based on a citation statement heterogeneous network as described in claim 1, characterized in that, Based on reinforcement learning, reference relationship path instances are obtained, and a reference relationship meta-path graph is generated, including: The nodes in the metapath graph are mapped to a low-dimensional vector space using a graph convolutional neural network model to generate a semantic representation for each node. Reinforcement learning is used to perform multi-hop reasoning on the heterogeneous network of citation statements to obtain citation relationship path instances; Find the lowest common ancestor from the reference relationship path instances, generate reference relationship meta-paths, and transform the citation statement heterogeneous network into multiple new graphs defined by meta-paths, generating a meta-path graph.

4. The local citation recommendation method based on a citation statement heterogeneous network as described in claim 3, characterized in that, Reinforcement learning is used to perform multi-hop inference on the heterogeneous network of citation statements to obtain citation relationship path instances, including: Construct a reinforcement learning agent, including state, action, transition and reward mechanism. The state represents the current node type and its set of neighboring nodes. The action is the reference transition operation from the current node to the target node. The reward is the positive feedback when the reference relationship is correctly predicted. Based on the reinforcement learning agent, multi-hop path reasoning is performed from the citation statement node in the citation statement heterogeneous network to obtain a directed acyclic path from the citation statement node to the target paper node, which serves as an instance of the citation relationship path.

5. A local citation recommendation system based on a citation statement heterogeneous network, characterized in that, The steps for implementing a local citation recommendation method based on a citation statement heterogeneous network as described in any one of claims 1 to 4 include: The network construction module is used to perform data preprocessing and citation relationship analysis on citation statements in the paper based on preset quantitative indicators and preset semantic features recommended by local citations, and to construct a heterogeneous network of citation statements. The citation recommendation module is used to perform low-dimensional node embedding learning on the heterogeneous network of citation statements, obtain citation relationship path instances based on reinforcement learning, generate a citation relationship meta-path graph, perform citation recommendation, and generate a list of recommended documents.

Images