CN117725220A - Method, server and storage medium for document characterization and document retrieval - Google Patents

Abstract

The application provides a method for document characterization and document retrieval, a server and a storage medium. According to the method, a plurality of documents to be processed and document association graphs of the plurality of documents are obtained, wherein the document association graphs comprise nodes corresponding to the documents and edges representing association relations among the documents; the method comprises the steps of inputting content information of a plurality of documents and a document association graph into a document representation model, performing diagrammatical learning through the document representation model based on the document association graph initialized by using semantic representations of the documents, updating feature representations of the nodes, taking the updated feature representations of the nodes as document representations of corresponding documents, learning semantic information of single documents and related information among the documents based on the document association graph, improving quality of the document representations, and further improving accuracy of document retrieval, recommendation and classification based on the document representations.

Description

Method, server and storage medium for document characterization and document retrieval

Technical field

The present application relates to computer technology, and in particular to a method, server and storage medium for document characterization and document retrieval.

Background technique

In actual application scenarios, many documents do not exist independently, such as documents containing hyperlinks, scientific papers with citation relationships, tender documents with common bidders, etc. Most of the real-life applications are related to multi-document scenarios, such as retrieval, recommendation, classification, multi-document summary and other scenarios of scientific papers, tender documents and other documents, all involving multiple documents with associated relationships.

In traditional document representation learning, technicians usually focus on the acquisition of semantic information in a single document. They mainly use language models to generate document representations based on the semantic information of a single document at the word, sentence and other levels, and cannot learn the correlation information between documents. , the quality of document representation is low, which leads to low accuracy in retrieval, recommendation, and classification based on document representation.

Contents of the invention

This application provides a method, server and storage medium for document characterization and document retrieval to solve the problem of low quality of document representation obtained by existing document characterization methods, which results in the accuracy of downstream retrieval, recommendation and classification based on document representation. low question.

In the first aspect, this application provides a document characterization method, including:

Obtain multiple documents to be processed and the document association graph of the multiple documents. The document association graph includes nodes corresponding to each of the documents and edges representing the association relationships between documents; convert the content information of the multiple documents The document association graph is input into a document representation model, and the document representation model performs graph representation learning based on the document association graph initialized using the semantic representation of each document, and updates the feature representation of each node; the updated features of each node are Represents a document representation as a corresponding document.

In the second aspect, this application provides a document retrieval method, including:

In response to the document retrieval request, obtain the query information input by the user; map the query information into a vector representation, perform similarity matching between the vector representation and the document representation of each document in the document retrieval database, and obtain a match with the query information at least one target document; output information of the at least one target document; wherein the document representation of each document in the document retrieval library is determined in the following manner: obtaining multiple documents to be represented in the document retrieval library, and the multiple documents A document association graph for each document, the document association graph including nodes corresponding to each of the documents and edges representing the association between documents; inputting the multiple documents and the document association graph into a document representation model, through the documents The representation model performs graph representation learning based on the document association graph initialized using the semantic representation of each document, and updates the feature representation of each node; the updated feature representation of each node is used as the document representation of the corresponding document.

In a third aspect, this application provides a server, including:

at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to The server is caused to execute the method described in the first aspect or the second aspect.

In a fourth aspect, the present application provides a computer-readable storage medium. Computer-executable instructions are stored in the computer-readable storage medium. When the processor executes the computer-executable instructions, the implementation as described in the first aspect or the second aspect is achieved. method described.

The document characterization and document retrieval methods, servers and storage media provided by this application obtain multiple documents to be processed and document association graphs of the multiple documents. The document association graph includes nodes corresponding to each document and represents the association between documents. edge; input the content information and document association graph of multiple documents into the document representation model, perform graph representation learning based on the document association graph initialized using the semantic representation of each document through the document representation model, update the feature representation of each node, and The updated feature representation of each node is used as the document representation of the corresponding document. Not only can the semantic information of a single document be learned, but also the correlation information between documents can be learned based on the document association graph of multiple documents, which improves the quality of the document representation and further It can improve the accuracy of document retrieval, recommendation, and classification based on document representation.

Description of the drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the application and together with the description, serve to explain the principles of the application.

Figure 1 is a schematic diagram of an example system architecture applicable to this application;

Figure 2 is a schematic diagram of another example system architecture applicable to this application;

Figure 3 is a flow chart of a document characterization method provided by an exemplary embodiment of the present application;

Figure 4 is an example architecture diagram of a document representation model provided by an exemplary embodiment of the present application;

Figure 5 is an example structural diagram of a semantic information learning module provided by an exemplary embodiment of the present application;

Figure 6 is an example structural diagram of a document representation model provided by an exemplary embodiment of the present application;

Figure 7 is a flow chart of a training method for a text representation model provided by an exemplary embodiment of the present application;

Figure 8 is an example diagram of the training framework of the document representation model provided by an exemplary embodiment of the present application;

Figure 9 is a flow chart of a document characterization method provided by an exemplary embodiment of the present application;

Figure 10 is a flow chart of a document retrieval method provided by another exemplary embodiment of the present application;

Figure 11 is a schematic structural diagram of a server provided by an embodiment of the present application.

Through the above-mentioned drawings, clear embodiments of the present application have been shown, which will be described in more detail below. These drawings and text descriptions are not intended to limit the scope of the present application's concepts in any way, but are intended to illustrate the application's concepts for those skilled in the art with reference to specific embodiments.

Detailed ways

Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the following exemplary embodiments do not represent all implementations consistent with this application. Rather, they are merely examples of apparatus and methods consistent with aspects of the application as detailed in the appended claims.

It should be noted that the user information (including but not limited to user device information, user attribute information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in this application are all It is information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards, and a corresponding operation entrance is provided for the user to choose to authorize or refuse.

First, the terms involved in this application will be explained:

Documents: When computers are used, files generated by various text editing software are generally called documents. In this embodiment, a document refers to a file that records text content. The document data includes the file content, as well as relevant information such as the creator, creation time, type, purpose, and field of the file.

Graph data (Graph): that is, the graph is an ordered tuple, expressed as G = (V, E), where V = {v ₁ ,..., v _N } represents the node set, Represents the edge set. The nodes in the graph have characteristic representations, and the edges represent the association between the two connected nodes.

Representation Learning (RL): It is a subtask in natural language processing. It mainly focuses on how to map semantic information in text into vector representations that can be used for downstream tasks.

Graph Neural Network (GNN): refers to the use of neural networks to learn graph-structured data, extract and explore features and patterns in graph-structured data, and meet graph learning tasks such as clustering, classification, prediction, segmentation, and generation. The general name of the required algorithm mainly focuses on how to apply the learning paradigm of neural network to the graph data structure, and then learn the information of the nodes and topology structure on the graph.

Wasserstein Distance (WD): It is a distance used to measure the similarity of two statistical distributions.

Graph pooling: is an operation used to decompose a graph with a similar structure into a small collection of nodes. It can select node information that represents the entire graph through global pooling operations such as maximum pooling or average pooling. The pooling operation will reduce the amount of data and reduce the number of nodes through a certain algorithm, thereby achieving layer-by-layer extraction. Global pooling only includes the readout layer, and the readout layer can use the pooling operation to read out the representation of the entire graph. Global pooling is often used in tasks such as graph classification.

Pre-trained large language model: a pre-trained model obtained by pre-training a large language model (LLM).

In traditional document representation learning, technicians usually focus on acquiring semantic information in a single document. However, in actual application scenarios, many documents do not exist independently, such as documents containing hyperlinks, scientific papers with citation relationships, tender documents with common bidders, etc. Most of the real-life applications are related to multi-document scenarios, such as retrieval, recommendation, classification, multi-document summary and other scenarios of scientific papers, tender documents and other documents, all involving multiple documents with associated relationships.

Traditional document representation learning solutions mainly use language models to generate document representations based on the semantic information of a single document at the word, sentence and other levels. However, the correlation information between documents cannot be learned, and the quality of the document representation is low, which in turn leads to document representation-based The accuracy of retrieval, recommendation, and classification is low.

In response to the above technical problems, this application provides a document characterization method to obtain multiple documents to be processed and the document association graph of the multiple documents. The document association graph includes nodes corresponding to each document and edges representing the association between documents; The content information and document association graph of multiple documents are input into the document representation model. The document representation model generates the semantic representation of each document based on the content information of each document. The semantic representation of each document is used to initialize the feature representation of each node in the document association graph. Based on The initialized document association graph is used for graph representation learning, the feature representation of each node is updated, and the updated feature representation of each node is used as the document representation of the corresponding document. Not only can the semantic information of a single document be learned, but also documents based on multiple documents can be learned. The association graph learns the correlation information between documents, which improves the quality of document representation, and thereby improves the accuracy of document retrieval, recommendation, and classification based on document representation.

Figure 1 is a schematic diagram of an example system architecture applicable to this application. As shown in Figure 1, the system architecture includes servers and end-side devices. Among them, there is a communicable communication link between the server and the end-side device, which can realize the communication connection between the server and the end-side device.

Among them, the server is a device with computing capabilities deployed in the cloud or locally, such as a cloud cluster, etc. The server stores a trained document representation model and can implement the document representation function based on the document representation model. In addition, the server can also be responsible for implementing the training of the document representation model.

End-side devices can be electronic devices that run downstream applications. Specifically, they can be hardware devices with network communication functions, computing functions, and information display functions, including but not limited to smartphones, tablets, desktop computers, local servers, cloud servers, etc. . When the end-side device runs downstream applications, it needs to use the document representation capability of the document representation model. The downstream application run by the end-side device may be an application system that implements at least one document processing task such as document retrieval, document recommendation, and document classification, such as bid document search, bid document classification, document retrieval, document classification, etc. When implementing the functions of downstream applications, it is necessary to use the document representation capability of the document representation model to generate high-quality document representations of each document based on multiple documents in a given document set. Further, based on the document representation of each document, the downstream application can be implemented Document processing tasks.

Based on the system architecture shown in Figure 1, the end-side device sends multiple documents to be processed to the server. The server receives multiple documents to be processed and constructs a document association graph of the multiple documents. The document association graph includes nodes corresponding to each document and edges representing the association between documents. Further, the server inputs the content information and document association graphs of multiple documents into the document representation model, performs graph representation learning through the document representation model based on the document association graph initialized using the semantic representation of each document, updates the feature representation of each node, and updates the The feature representation of each node is then used as the document representation of the corresponding document. The server returns the document representations of the multiple documents to the end-side device.

The end-side device receives the document representations of multiple documents returned by the server, continues to execute the document processing logic of the downstream application based on the document representations of the multiple documents, implements the document processing tasks of the downstream application, and obtains the document processing results.

In an example scenario, taking the downstream application as a document classification system as an example, when classifying multiple given documents, the end-side device needs to obtain the text representation of each document, and then classify based on the document representation of each document, obtaining Document classification results. The terminal device sends a document set to be classified to the server. The document set contains multiple documents, at least two of which have an associated relationship. The server receives the document set sent by the end-side device and constructs a document association graph based on the association between documents. The server inputs the content information and document association graph of multiple documents in the document set into the document representation model. The document representation model performs graph representation learning based on the document association graph initialized using the semantic representation of each document, updates the feature representation of each node, and updates the The feature representation of each node is then used as the document representation of the corresponding document. The server returns document representations of multiple documents in the document set to the end-side device. The end-side device receives the document representations of multiple documents in the document set returned by the server, classifies each document according to the document representation of each document, and obtains the classification result of each document.

Based on the system architecture shown in Figure 1, other document processing tasks such as document clustering and generating summaries of multiple documents can also be implemented, which are not specifically limited here. For example, classification of bid documents and documents, clustering of bid documents and documents, summary of specified multiple documents, etc.

FIG. 2 is a schematic diagram of another example system architecture applicable to this application. As shown in Figure 2, the system architecture includes servers and end-side devices. Among them, there is a communicable communication link between the server and the end-side device, which can realize the communication connection between the server and the end-side device.

Among them, the server is a device with computing capabilities deployed in the cloud or locally, such as a cloud cluster, etc. The server stores a trained document representation model and can implement the document representation function based on the document representation model. The server can use the document representation model to characterize multiple documents in a given document retrieval library, obtain the document representation of each document in the document retrieval library, and store the document representation of each document in the document retrieval library. Further, the server can implement tasks such as document retrieval and recommendation based on a document retrieval database containing document representations of each document. In addition, the server can also be responsible for implementing the training of the document representation model.

The end-side device can be an electronic device used by the user, and specifically can be a hardware device with network communication functions, computing functions, and information display functions, including but not limited to smartphones, tablet computers, desktop computers, etc.

Based on the system architecture shown in Figure 2, the user sends a document processing request to the server through the end-side device. The document processing request contains user-related input information, including at least one of the following: user-input query information, user portrait, user historical behavior ( Including but not limited to browsing, clicking, searching, collecting, citing, following) related document information. The server receives user-related input information, maps the user-related input information into a vector representation, performs similarity matching between the vector representation and the document representation of each document in the document retrieval database, and obtains at least one target document that matches the input information, and sends it to the terminal The side device returns information about the target document. The terminal device receives the information of the target document returned by the server and displays the information of the target document to the user.

In an example scenario, taking the document retrieval scenario as an example, the user sends a document retrieval request to the server through the terminal device, and the document retrieval request includes the query information input by the user. The server obtains the query information input by the user, maps the query information into a vector representation, performs similarity matching between the vector representation and the document representation of each document in the document retrieval database, obtains at least one target document that matches the query information, and obtains the document retrieval results. Further, the server returns the information of the target document to the end-side device. The end-side device receives the information of the target document returned by the server, displays the information of the target document to the user, and completes the document retrieval. Among them, the document retrieval scenario can specifically include tender document retrieval, document retrieval, retrieval of knowledge documents in various technical fields, etc. There are no specific limitations here.

Based on the system architecture shown in Figure 2, document recommendation can also be implemented. According to the first document related to the user's behavior, the target document with a higher similarity to the document representation of the first document is retrieved in the document retrieval library and recommended to the user. Personalized document recommendations based on user behavior. For example, recommendation of tender documents, recommendation of literature, recommendation of knowledge documents in various technical fields, etc. are not specifically limited here.

The technical solution of the present application and how the technical solution of the present application solves the above technical problems will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of the present application will be described below with reference to the accompanying drawings.

Figure 3 is a flow chart of a document characterization method provided by an exemplary embodiment of the present application. The execution subject of this embodiment may be a server running the document representation model, and may specifically be a server in any of the aforementioned system architectures. As shown in Figure 3, the specific steps of this method are as follows:

Step S301: Obtain multiple documents to be processed and document association graphs of the multiple documents. The document association graph includes nodes corresponding to each document and edges representing association relationships between documents.

The method of this embodiment is suitable for performing multi-document representation learning on multiple documents with certain correlations to obtain high-quality document representation of each document. It can be specifically applied to document classification, clustering, retrieval, recommendation, and multi-document summary. and other application scenarios. When applied to different application scenarios, the multiple documents to be processed can be different. For example, the multiple documents to be processed may be documents in a document set to be classified or clustered given by the user/downstream application, or all or part of the documents in a specific document retrieval library. Among them, the content of the document to be processed can include but not limited to text, images, program codes, hyperlinks and other information.

The association between documents represents the correlation between documents, which can be citations, the same author, the same target object, the same description object, etc. For example, there can be reference relationships between paper-type documents, tender-type documents can belong to the same bidder/tenderer, knowledge-type documents can describe the same item, etc. In different application scenarios, for different document sets, the relationships between documents can be different. The correlation between different documents in a given document set in a specific application scenario can be determined through data analysis and mining. The details can be obtained using existing technologies, which are not specifically limited here.

For multiple given documents, nodes corresponding to each document are constructed. According to the association between documents, an edge is established between any two document corresponding nodes with an association relationship to obtain a document association graph of multiple documents. It should be noted that the same document association graph can contain one or more different types of edges, and different types of edges represent different association relationships.

In this step, optionally, the end-side device provides multiple documents to be processed to the server. The server obtains multiple documents to be processed from the terminal device, analyzes the relationships between the multiple documents, and builds a document association graph of the multiple documents based on the relationships between the multiple documents. The document association graph includes nodes corresponding to each document and edges representing the association between documents.

Optionally, the end-side device provides multiple documents to be processed and document association graphs of the multiple documents to the server. The server receives multiple documents sent by the end-side device and the document association graph of the multiple documents. The end-side device is responsible for building a document association graph of multiple documents.

Step S302: Input the content information and document association graphs of multiple documents into the document representation model, and use the document representation model to perform graph representation learning based on the document association graph initialized using the semantic representation of each document, and update the feature representation of each node.

In this embodiment, the server inputs the content information of each document and the document association graph into the document representation model. The document representation model generates the semantic representation of each document based on the content information of each document, and uses the semantic representation of each document obtained based on a single document representation. Initialize the feature representation of each node in the document association graph, perform graph representation learning on the initialized document association graph and update the feature representation of each node, and output the updated document association graph. In this way, based on the semantic representation of each document obtained based on a single document representation, the topological structure information of the document association graph is learned through graph representation learning, that is, the correlation information between multiple documents is learned, and each element in the document association graph is improved. The quality of a node’s feature representation.

The content information of the document input into the document representation model may include the complete content information of the document, or may be partial content information of a pre-specified document, such as the document body, abstract, key paragraphs, etc., which are not specifically limited in this embodiment. .

In an optional embodiment, the server can use other models or algorithms to obtain the semantic representation of each document, use the semantic representation of each document to initialize the document association graph of multiple documents, and input the initialized document association graph into the document representation model. The representation model performs graph representation learning based on the input initialized document association graph and updates the feature representation of each node, and outputs an updated document association graph.

Step S303: Use the updated feature representation of each node as the document representation of the corresponding document.

Using the high-quality feature representation of each node in the updated document association graph as the document representation of the document corresponding to the node, a high-quality document representation of each document can be obtained.

Furthermore, based on the high-quality document representation of each document, document processing tasks such as document retrieval, recommendation, classification, clustering, and multi-document summary can be realized, thereby improving the accuracy of document processing.

Illustratively, taking document clustering as an example, the method of this embodiment can be used to obtain high-quality document representations of multiple documents to be clustered, and clustering is performed based on the obtained document representations to obtain document clustering results. Improve the accuracy of document clustering.

Illustratively, taking document retrieval as an example, a high-quality document representation of each document in the document retrieval library can be obtained through the method of this embodiment. When performing document retrieval, the vector representation of user-related input information is matched with high-quality document representations in the document retrieval database to determine the target document that matches the input information and obtain the document retrieval results, which can improve the accuracy of document retrieval. degree and recall.

The method of this embodiment obtains multiple documents to be processed and the document association graph of the multiple documents. The document association graph includes the nodes corresponding to each document and the edges representing the association relationships between documents, and combines the content information of the multiple documents with the document The association graph is input into the document representation model, and the semantic representation of each document is generated based on the content information of each document through the document representation model. The semantic representation of each document is used to initialize the feature representation of each node in the document association graph, and the graph is performed based on the initialized document association graph. Representation learning updates the feature representation of each node and uses the updated feature representation of each node as the document representation of the corresponding document. It can not only learn the semantic information of a single document, but also learn the correlation between documents based on the document association graph of multiple documents. This information improves the quality of document representation, thereby improving the accuracy of document retrieval, recommendation, and classification based on document representation.

Figure 4 is an example architecture diagram of a document representation model provided by an exemplary embodiment of the present application. In an optional embodiment, as shown in Figure 4, in the foregoing method embodiment, the document representation model includes a semantic information learning module and a related Sexual Information Learning Module. Among them, the semantic information learning module is used to perform semantic information representation learning on the content information of each document to obtain the semantic representation of each document. The correlation information learning module is used to: use the semantic representation of each document to initialize the feature representation of each node in the document association graph, perform graph representation learning based on the initialized document association graph, and update the feature representation of each node.

Based on the model architecture shown in Figure 4, the aforementioned step S302 is implemented in the following manner:

Step S3021: Input the content information of multiple documents into the semantic information learning module, perform semantic information representation learning on the content information of each document through the semantic information learning module, obtain the semantic representation of each document, and input the semantic representation of each document into the correlation Learning modules.

In this embodiment, the semantic information representation learning is performed on the content information of a single document through the semantic information learning module. The semantic information at the word, sentence and other levels within the document can be better learned, thereby obtaining a high-quality semantic representation of a single document. . The obtained high-quality semantic representation of a single document will output the subsequent correlation learning module, which is used to initialize the feature representation of each node in the document correlation graph.

Step S3022: Input the document association graph into the relevance information learning module, use the semantic representation of each document through the relevance information learning module to initialize the feature representation of each node in the document association graph, and perform graph representation learning based on the initialized document association graph. , update the feature representation of each node.

In this embodiment, the document association graph is used as the input of the relevance information learning module. The relevance information learning module uses the high-quality semantic representation of each document output by the semantic information learning module to initialize the feature representation of each node in the document association graph, such that The feature representation of the nodes in the document association graph after initialization is a high-quality semantic representation of the corresponding document. Furthermore, the correlation information learning module performs graph representation learning based on the initialized document association graph, updates the feature representation of each node, and can learn the topological structure information of the document association graph based on the multi-document document association graph, and the document association graph Topological structure information contains correlation information between documents, so the correlation information between documents can be learned in the process of updating the feature representation of nodes, so that the feature representation of the updated node not only contains the semantic information of a single document, but also contains the semantic information between documents. The correlation information of the updated node is used as the document representation of the corresponding document, which improves the quality of the document representation.

In another optional embodiment, the document representation model may also include only a correlation information learning module, which is implemented using a graph neural network GNN. When performing multi-document representation, the existing document representation method is used to perform single-document semantic representation on the content information of multiple documents to obtain the semantic representation of each document. The semantic representation of each document is used to initialize the document association graph of multiple documents. The feature representation of each node is used to obtain the initialized document association graph. The initialized document association graph is input into the relevance information learning module for graph representation learning, and the feature representation of each node in the document association graph is updated.

Figure 5 is an example structural diagram of a semantic information learning module provided by an exemplary embodiment of the present application. Based on the model architecture shown in Figure 4, in an optional embodiment, as shown in Figure 5, the semantic information learning module Modules include: entity relationship diagram building module, first graph neural network, text representation model and semantic representation fusion module.

Among them, the entity-relationship graph building module is used to construct an entity-relationship graph (Entity-Relationship Graph, E-R graph for short) of each document based on the content information of each document. The first graph neural network is used to: perform graph representation learning on the entity relationship diagram of each document to obtain the feature representation of the entities contained in each document, and fuse the feature representations of the entities contained in each document to obtain the entity-level semantics of each document express. The first graph neural network can use Graph Convolutional Network (GCN for short), Neural Network for Graphs (NN4G for short), Graph Attention Network (GAT for short), graph synchronization Implementation of Graph Isomorphism Network (GIN for short) or other graph neural networks can realize representation learning of graph data. This embodiment is not specifically limited here.

Among them, the entity-relationship graph (E-R graph) of a document is a graph structure constructed with the named entities contained in the document as nodes and the relationships between named entities as edges. Specifically, based on the content information of any document, the named entities contained in the document and the relationship between the named entities can be extracted, corresponding nodes can be created based on the named entities contained in the document, and the edges between the corresponding nodes can be constructed based on the relationship between the named entities to construct the document. Entity relationship diagram. The feature vector of the node in the entity relationship graph may be a vector representation corresponding to the named entity, which may be specifically determined based on the word vector of the named entity. For example, perform word segmentation processing for any named entity. If the named entity contains only one word after word segmentation, the word vector of the word is used as the vector representation of the named entity and as the feature representation of the node corresponding to the named entity. If the named entity contains multiple words, the word vectors of the multiple words are averaged to obtain the vector representation of the named entity, which is used as the feature representation of the node corresponding to the named entity. In addition, the feature vectors of the nodes in the entity relationship diagram can also be determined through random initialization or other methods, which are not specifically limited here.

Optionally, the first graph neural network may include a graph encoding layer and a graph pooling layer. The graph encoding layer is used to perform graph representation learning on the entity relationship graph of each document, and encode the feature representation of each node in the entity relationship graph. , update the feature representation of nodes in the entity relationship graph, and output the updated entity relationship graph. The characteristic representation of each node in the updated entity relationship graph is used as the characteristic representation of the corresponding entity, that is, the characteristic representation of the entities contained in each document is obtained. The updated entity relationship graph is input to the graph pooling layer. The graph pooling layer is used to perform graph pooling operations on the updated entity relationship graph to achieve the fusion of feature representations of each node and output a graph representation of the entity relationship. The graph representation of entity relationships is used as the semantic representation of the corresponding document. This semantic representation is obtained by fusing the feature representations of the entities contained in the document, and is a semantic representation of the entity level of the document.

Optionally, the first graph neural network may include a graph coding layer. The graph coding layer is used to perform graph representation learning on the entity relationship graph of each document, encode the feature representation of each node in the entity relationship graph, and update the entity relationship graph. Feature representation of the nodes in the node, and output the updated entity relationship graph. The characteristic representation of each node in the updated entity relationship graph is used as the characteristic representation of the corresponding entity, that is, the characteristic representation of the entities contained in each document is obtained. Furthermore, the feature representations of the entities contained in any document are averaged or spliced to achieve the fusion of the feature representations of each node and obtain the semantic representation of the entity level of the document.

In this embodiment, the text representation model is used to perform text representation on the content information of each document to obtain a text-level semantic representation of each document. The text representation model can be implemented using any existing text representation model/algorithm. For example, the text representation model can be implemented using various types of language models (Language Model, LM for short), such as Transformer-based Bidirectional Encoder Representations from Transformers (BERT for short), large language models, pre-trained language models, etc.

The semantic representation fusion module is used to fuse the entity-level semantic representation of each document and the text-level semantic representation to obtain the semantic representation of each document. Specifically, the semantic representation fusion module obtains the semantic representation of each document by averaging the semantic representation of the entity level and the semantic representation of the text level of each document; or, the semantic representation of the entity level and the semantic representation of the text level of each document are obtained. Splicing to obtain the semantic representation of each document.

Based on the structural example of the semantic information learning module shown in Figure 5, the aforementioned step S3021 can be implemented in the following manner:

Input the content information of multiple documents into the entity relationship diagram building module, and use the entity relationship diagram building module to construct the entity relationship diagram of each document based on the content information of each document; input the entity relationship diagram of each document into the first graph neural network, and use The first graph neural network performs graph representation learning on the entity relationship diagram corresponding to each document to obtain the entity-level semantic representation of each document; inputs the content information of multiple documents into the text representation model, and uses the text representation model to The content information is represented as text respectively to obtain the text-level semantic representation of each document; through the semantic representation fusion module, the entity-level semantic representation of each document and the text-level semantic representation are fused to obtain the semantic representation of each document. The semantic representation of each document is output to the correlation learning module.

In another optional embodiment, the semantic information learning module can be implemented using a text representation model, and it only needs to be able to extract the semantic information of the content text contained in a single document. For example, the semantic information learning module can be implemented using a language model (Language Model, LM for short). For example, Transformer-based bidirectional encoder representations (BidirectionalEncoder Representations from Transformers, BERT for short), large language models, pre-trained language models, etc. are implemented.

In the aforementioned step S3021, multiple documents are respectively input into the text representation model, and text representation is performed on each document through the text representation model to obtain the text-level semantic representation of each document as the semantic representation of each document.

In another optional embodiment, the semantic information learning module can be implemented based on the graph neural network GNN, and constructs the entity relationship graph of the document by extracting the named entities contained in the document and the relationships between the named entities based on the content information of a single document. The nodes in the entity relationship graph represent named entities contained in the content information of the document. The feature vectors of the nodes may be vector representations corresponding to the named entities. The edges in the entity relationship graph represent relationships between named entities. Input the entity relationship graph of the document into the graph neural network GNN of the semantic information learning module for graph representation learning, update the feature vectors of the nodes in the entity relationship graph, and obtain the encoded entity relationship graph. The features of each node in the encoded entity relationship graph are obtained. Vector as an encoding vector of named entities. Integrate the encoding vectors of named entities contained in the document to obtain the semantic representation of the document. Specifically, when the encoding vectors of the named entities contained in the document are integrated to obtain the semantic representation of the document, the encoding vectors of each named entity contained in the document can be averaged to obtain the semantic representation of the document. Or, perform graph pooling on the encoded entity-relationship graph, and read out the graph representation of the entire entity-relationship graph as the semantic representation of the corresponding document.

In the aforementioned step S3021, the entity relationship graph of each document is input into the graph neural network for semantic information learning. The entity relationship graph of each document is separately graph representation learned through the graph neural network to obtain the semantic representation of the entity level of each document as each document. Semantic representation of the document.

In any of the foregoing method embodiments, the entity relationship diagram of any document is obtained as follows: performing Named Entity Recognition (NER) and Relation Extraction (RE) on the content text of the document to obtain the content of the document. entities (that is, named entities), and the relationships between entities. Construct nodes corresponding to each entity contained in the document, construct edges based on the relationships between entities, and obtain the entity relationship graph of the document. Among them, named entity recognition NER and relationship extraction RE are performed on the content text of the document, which can be implemented using existing named entity recognition methods and relationship extraction methods, and are not specifically limited here. It should be noted that the content of the document can include but not limited to text, images, program code, hyperlinks and other information. When performing named entity recognition and relationship extraction on the content text of a document, named entity recognition and relationship extraction can be performed based on the text content in the document (including program code and hyperlinks, etc.) to obtain the entities contained in the document (that is, named entities) , and the relationships between entities. Alternatively, image recognition can be performed on the image in the document to obtain the text information contained in the image in the document. The text information contained in the image in the document is also used as the content text of the document. According to the text content in the document and the text information contained in the image, the name Entity recognition and relationship extraction are used to obtain the entities contained in the document (that is, named entities) and the relationships between entities.

In an optional embodiment, based on the model architecture shown in Figure 4, the correlation information learning module includes: an initialization module and a second graph neural network. Among them, the initialization module is used to: use the semantic representation of each document to initialize the feature representation of each node in the document association graph. The second graph neural network is used to perform graph representation learning based on the initialized document association graph and update the feature representation of each node.

In this embodiment, the aforementioned step S3022 can be implemented in the following manner:

Input the semantic representation and document association graph of each document into the initialization module, use the semantic representation of each document through the initialization module to initialize the feature representation of each node in the document association graph, and input the initialized document association graph into the second graph neural network, through The second graph neural network performs graph representation learning based on the initialized document association graph and updates the feature representation of each node.

Among them, the second graph neural network includes a graph coding layer. The graph coding layer is used to perform graph representation learning on the initialized document association graph, encode the feature representation of each node in the document association graph, and update the features of the nodes in the document association graph. Indicates that the updated document association graph is output. The feature representation of each node in the updated document association graph is used as the document representation of the corresponding document, that is, the document representation of each document is obtained.

Exemplarily, the second graph neural network can be implemented using a graph convolutional neural network GCN, a constructive graph neural network NN4G, a graph attention network GAT, a graph isomorphism network GIN or other graph neural networks, and can realize the representation of graph data. Learning is not specifically limited in this embodiment. Optionally, the second graph neural network and the graph coding layer of the first graph neural network of the semantic information learning module can be implemented using the same structure, but do not share parameters.

In another optional embodiment, the relevance information learning module may only include the second graph neural network. Before inputting the document association graph into the relevance information learning module, each node in the document association graph is initialized using the semantic representation of each document. Feature representation, obtain the initialized document association graph, and input the initialized document association graph into the correlation information learning module.

Exemplarily, FIG. 6 is an example structural diagram of a document representation model provided by an exemplary embodiment of the present application. As shown in FIG. 6 , the document representation model includes a semantic information learning module and a correlation information learning module. Among them, the semantic information learning module includes: entity relationship diagram building module, first graph neural network, language model and semantic representation fusion module. The correlation information learning module includes an initialization module and a second graph neural network. When performing multi-document representation, input the content information of multiple documents into the language model, perform text representation on the content information of each document through the language model, obtain the text-level semantic representation of each document, and input it into the semantic representation fusion module; and Input the content information of multiple documents into the entity relationship diagram building module, and use the entity relationship diagram building module to construct the entity relationship diagram of each document based on the content information of each document; input the entity relationship diagram of each document into the first graph neural network, and use The first graph neural network performs graph representation learning on the entity relationship diagram of each document, obtains the feature representation of the entities contained in each document, fuses the feature representations of the entities contained in each document, and obtains the semantic representation of the entity level of each document, and Enter the semantic representation fusion module. Through the semantic representation fusion module, the entity-level semantic representation of each document and the text-level semantic representation are fused to obtain the semantic representation of each document. The semantic representation of each document is input into the initialization module of the correlation information learning module. The semantic representation of each document is used through the initialization module to initialize the feature representation of each node in the document association graph to obtain the initialized document association graph and input it into the second graph neural network. network. The second graph neural network performs graph representation learning based on the initialized document association graph, updates the feature representation of each node, and uses the updated feature representation of each node as the document representation of the corresponding document.

Figure 7 is a flow chart of a training method for a text representation model provided by an exemplary embodiment of the present application. In an optional embodiment, as shown in Figure 7, the training process of the document representation model used in the foregoing method embodiment is as follows:

Step S701: Construct a document representation model.

In this embodiment, the document representation model based on the model architecture shown in Figure 4 can be constructed. Specifically, the model architecture in any of the aforementioned method embodiments can be used. For example, Figure 4 is combined with other optional embodiments to determine the model structure, or the model structure shown in Figure 6 The model structure shown is not specifically limited in this embodiment.

Step S702: Input the multiple document samples in the document set used for training and the document association graphs of the multiple document samples into the document representation model, and use the document representation model to mask the initial edges in the document association graphs of the multiple document samples. Process, and use the semantic representation of each document sample to initialize the feature representation of each node in the document association graph of multiple document samples, obtain the initialized document association graph, perform graph representation learning based on the initialized document association graph, and update the Feature representation.

In this embodiment, the training data used to train the text representation model includes multiple document sets. Each document set contains multiple document samples. The multiple document samples in the document set have a certain correlation, and some of the document samples are related. relation. Among them, multiple document sets can be obtained by grouping multiple documents in the existing document representation training data, and multiple documents in each group constitute a document set. The relationship between each document is the document's own characteristics, which can be determined by analyzing the document content or document-related attribute information (such as author, time, purpose, field, source, etc.), or obtained by manual annotation. In addition, the document set can be automatically divided by the server based on relevant attributes such as the neighborhood and source of the document, or the server can randomly divide the documents, or the relevant technical personnel can manually divide the document set. There is no specific limit here.

In this step, the process of obtaining document association graphs of multiple document samples based on multiple document samples in a document set is consistent with the implementation of obtaining document association graphs of multiple documents in the aforementioned step S301. For details, please refer to the aforementioned embodiments. The relevant content will not be repeated here.

The specific implementation of step S702 is similar to the aforementioned step S302. The difference is that in the process of initializing the document association graph, only the feature representation of each node is initialized in step S302, while in step S702 not only the feature representation of each node is initialized. The initial edges in the document association graph are also masked, and other processing procedures are consistent with step S302. For details, please refer to the relevant content of the foregoing embodiments, which will not be described again here.

In this embodiment, the training task of the text representation model is the task of predicting the edges between nodes in the document association graph. In order to construct the training task, in the process of initializing the document association graph of multiple document samples, not only the Semantic representation: initialize the feature representation of each node in the document association graph of multiple document samples. It is also necessary to mask the initial edges in the document association graph of multiple document samples to obtain the initialized document association graph. In subsequent step S703, based on the updated characteristic representation of each node, the association relationship between each node is predicted, that is, whether there is an edge between each node is predicted.

For example, a graph G can be represented as an ordered tuple, represented as G = (V, E), where V = {v ₁ ,..., v _N } represents the node set, Represents the edge set. A graph G can be stored as a feature matrix X∈R ^N×d associated with an adjacency matrix A∈R ^N×N . Among them, the feature matrix X is a matrix composed of feature representations of nodes in the graph, and d is the dimension of node feature representations. N is the number of nodes contained in graph G. The adjacency matrix A records whether there is an edge between any two nodes in the graph. A _ij =1 indicates that there is an edge between nodes _vi and v _j , and A _ij =0 indicates that there is no edge between nodes _vi and v _j . To mask the initial edges in the document association graph of multiple document samples, 1 in the adjacency matrix can be covered with a specified mask (such as 0 or NULL), where NULL indicates that it is uncertain whether an edge exists.

Step S703: Predict the association between nodes in the document association graph based on the updated feature representation of each node.

After obtaining the updated feature representation of each node, the feature representation of each node can be input into the first classifier, and the first classifier can be used to classify and predict whether there is an association relationship between each node in the document association graph, and determine whether there is an association relationship between the nodes in the document association graph. The association relationship between each node is used to obtain the prediction result of the association relationship between each node in the document association graph. The prediction result includes whether there are edges between nodes in the document association graph (that is, whether there is an association relationship).

Among them, the first classifier used to predict the association between nodes in the document association graph can use a multi-layer perceptron (MLP), or other classification models or algorithms, such as support vector machines. , decision tree classification algorithm, etc., there are no specific limitations here in this embodiment.

Step S704: Calculate the first loss based on the prediction result of the association relationship between nodes in the document association graph and the initial edge in the document association graph, and update the parameters of the document representation model based on the first loss.

In this step, cross-entropy (CE) is calculated based on the updated feature representation of each node, the determined prediction results of the association between nodes in the document association graph, and the initial edges in the document association graph. Loss, as the first loss. Further, according to the first loss, the parameters of the document representation model are updated through backpropagation. Specifically, the gradient descent method or other common parameter update methods can be used to update the parameters of the document representation model. After training is completed, the trained document representation model can be obtained.

It should be noted that the first classifier used in step S703 may be a pre-trained first classifier whose parameters remain unchanged during the training process of the document representation model; or, in the training process of the document representation model, During the training process, updating the parameters of the first classifier and the document representation model based on the first loss can continuously improve the classification accuracy of the first classifier during the training process, thereby improving the training effect of the document representation model.

The method of this embodiment not only performs semantic representation on a single document and obtains a semantic representation containing rich semantic information for each document, but also models the association between multiple documents, using the document as a node and the association between documents as an edge. , build a document association graph between multiple documents. The topological structure of the document association graph contains the relevance information of the document. During the training process of the text representation model, the semantic representation of each document containing rich semantic information is used to initialize the feature representation of the nodes in the document association graph, and the masking strategy is used to mask the edges in the document association graph and then input into the graph neural The network performs graph representation learning, updates the feature representation of nodes, infers the association between nodes based on the updated feature representation, and calculates the loss and updates the model parameters based on the prediction results and the initial edges of the document association graph, making the document representation model not only It has the ability to learn document semantic information and also has the ability to learn correlation information between documents, which can improve the quality of multi-document representation.

In an optional embodiment, based on the structure of the semantic information learning module shown in Figure 5, the semantic information learning module includes: an entity relationship diagram building module, a first graph neural network, a text representation model and a semantic representation fusion module. For the specific structure, please refer to the relevant content of the foregoing embodiments and will not be described again here. In the aforementioned step 702, the semantic representation of each document sample is used through the document representation model, which is specifically implemented in the following manner:

Input the content information of each document sample in the document set into the entity relationship diagram building module, and use the entity relationship diagram building module to construct the entity relationship diagram of each document sample based on the content information of each document sample; enter the entity relationship diagram of each document sample in the document set The graph is input into the first graph neural network, and graph representation learning is performed on the entity relationship diagram of each document sample through the first graph neural network to obtain the entity-level semantic representation of each document sample; and the key points of multiple document samples in the document set are obtained The content information is input into the text representation model, and the key content information of each document sample is separately represented as text through the text representation model to obtain the text-level semantic representation of each document sample; the entity-level semantic representation and text-level semantic representation of each document sample are Representation fusion is used to obtain the semantic representation of each document sample.

Ideally, entity-level information can help text-level information focus more on useful information and reduce the weight of useless information; while text-level information can enrich entity-level information. However, during the actual training process, the applicant found that training the text representation model and the first graph neural network separately would make it difficult to balance the information at the two levels.

In this embodiment, during the training process of the document representation model, in order to balance the learning between the text representation model and the first graph neural network, a second loss is added to constrain the learning of the two levels of information. Specifically, this is achieved in the following ways:

According to the entity-level semantic representation of each document sample, the category label of each document sample is predicted to obtain the first prediction result; and based on the text-level semantic representation of each document sample, the category label of each document sample is predicted to obtain the second prediction result. ; Calculate the distance between the first prediction result and the second prediction result to obtain the second loss; update the parameters of the semantic information learning module according to the second loss.

Among them, the category label of the document sample represents the category of the document sample, which can be the field, subject, type, etc. of the sample document. The actual category label of the document sample is determined through pre-annotation.

For example, when predicting the category label of each document sample based on the entity-level semantic representation of each document sample, a second classifier can be used. The entity-level semantic representation of each document sample is input into the second classifier, and the category of the document sample is classified and predicted by the second classifier to obtain the first prediction result. The first prediction result includes the distribution of category labels of each document sample. Among them, the distribution of category labels for any document sample is a discrete distribution, which can be expressed as a vector with dimension K, where K is the total number of category labels. 1 in the vector indicates that the document has the corresponding category label, and 0 in the vector indicates that the document does not Have corresponding category labels. For example, the first prediction result may be a complete discrete distribution formed by splicing distributions of category labels of each document sample. Among them, the second classifier can be implemented using a multi-layer perceptron (MLP) or other classification models or algorithms, such as classification algorithms based on support vector machines and decision trees, etc., which are not included in this embodiment. Specific limitations.

For example, when predicting the category label of each document sample based on the text-level semantic representation of each document sample, a third classifier can be used. The text-level semantic representation of each document sample is input into a third classifier, and the category of the document sample is classified and predicted by the third classifier to obtain a second prediction result. The second result contains the distribution of category labels of each document sample, which is similar to the information items contained in the first prediction result. Among them, the third classifier can be implemented using a Multi-Layer Perceptron (MLP), or other classification models or algorithms, such as classification algorithms based on support vector machines, decision trees, etc., which are not included in this embodiment. Make specific limitations.

The second classifier and the third classifier can be implemented using the same or different classifiers, and are not specifically limited here. In addition, the second classifier and the third classifier may be pre-trained classifiers, and the parameters of the second classifier and the third classifier remain unchanged during the training process of the document representation model.

Optionally, during the training process of the document representation model, the first cross-entropy loss can be calculated based on the first prediction result and the actual category label of the document sample, and the parameters of the second classifier can be updated based on the first cross-entropy loss; and based on the first cross-entropy loss; Calculate the second cross-entropy loss based on the second prediction result and the actual category label of the document sample, and update the parameters of the third classifier based on the second cross-entropy loss, thereby continuously improving the classification accuracy of the two classifiers during the training process, so that the Improve the training effect of document representation model.

Optionally, the distance between the first prediction result and the second prediction result may specifically be the Wasserstein distance or KL (Kullback-Leibler) between the two discrete distributions of the first prediction result and the second prediction result. Divergence, Manhattan distance, etc., or other information used to measure the similarity between two discrete distributions, are not specifically limited here. In a preferred embodiment, the Wasserstein distance between the first prediction result and the second prediction result is calculated, and as the second loss, the Wasserstein distance is used to balance the first graph neural network and the text representation. Learning between models can improve the effect of model training.

Optionally, the first cross-entropy loss can also be calculated based on the first prediction result and the actual category label of each document sample; and/or the second cross-entropy loss can be calculated based on the second prediction result and the actual category label of each document sample. Loss; determine the third loss based on the first cross entropy loss and/or the second cross entropy loss, and update the parameters of the semantic information learning module based on the second loss and the third loss, which can improve the performance of the semantic information learning module, thereby improving The quality of the semantic representation of the generated documents.

Alternatively, the first cross-entropy loss may be used as the third loss, or the second cross-entropy loss may be used as the third loss, or the sum of the first cross-entropy loss and the second cross-entropy loss may be used as the third loss.

Optionally, when updating the parameters of the semantic information learning module based on the second loss and the third loss, the sum of the second loss and the third loss can be used as the comprehensive semantic learning loss, or the weighted sum of the second loss and the third loss can be used As a comprehensive semantic learning loss, the parameters of the semantic information learning module are updated according to the comprehensive semantic learning loss. Among them, the weight coefficients of the second loss and the third loss can be configured and adjusted according to the needs of the actual application scenario, and are not specifically limited here.

Optionally, after the first loss, the second loss and the third loss are calculated, the comprehensive loss can be calculated based on the first loss, the second loss and the third loss, and the parameters of the entire document representation model can be updated based on the comprehensive loss. For example, the sum of the first loss, the second loss and the third loss is calculated as the comprehensive loss; or the weighted sum of the first loss, the second loss and the third loss is calculated as the comprehensive loss. The weight coefficients of the first loss, the second loss and the third loss can be configured and adjusted according to the needs of the actual application scenario, and are not specifically limited here.

Exemplarily, Figure 8 is an example diagram of the training framework of the document representation model provided by an exemplary embodiment of the present application. Based on the document representation model shown in Figure 6, a first graph corresponding to the first graph neural network is added during the training process. a classifier, a second classifier corresponding to the language model, and a third classifier corresponding to the second graph neural network. During the training process of the document representation model, the entity relationship diagram of each document sample ( _PA ...P _B as shown in Figure 8) is input into the first graph neural network for graph representation learning, and it is obtained that each document sample contains entities The feature representation (e ₁ , e ₂ , e ₅ , e ₅ , e ₆ , e ₉ shown in Figure 8 ) can further determine the entity-level semantic representation of each document sample, as shown in Figure 8 The superscript E represents the entity level. The entity-level semantic representation of each document sample is input into the first classifier for classification prediction, and the first prediction result is obtained, as shown in μ ^E in Figure 8.

Moreover, the summary information of each document sample is input into the language model for text representation, and the text-level semantic representation of each document sample is obtained, as shown in Figure 8 The superscript T represents the text level. The text-level semantic representation of each document sample is input into the second classifier for classification prediction, and the second prediction result is obtained, as shown in Figure 8 as μ ^T .

By calculating the Wasserstein distance d between the first prediction result μ ^E and the second prediction result μ ^T , it is used as the second loss to constrain the learning of the first graph neural network and language model, and the intersection is calculated in combination with the actual category label Entropy loss (not shown in Figure 8), and then determine the comprehensive semantic information learning loss l ₁ (not shown in Figure 8).

Fuse the entity-level semantic representation of each document with the text-level semantic representation, use the fused semantic representation of each document to initialize the feature representation of each node in the document association graph, and mask the edges in the document association graph (Fig. The dotted line in the document association graph shown in 8 indicates that the edge is masked, and the solid line indicates the actual existing edge or the predicted existing edge) to obtain the initialized document association graph. The initialized document association graph is input into the second graph neural network for graph representation learning, and the feature representation of each node is updated. The third classifier is used to predict whether there is an edge between each node based on the updated feature representation of each node, and the edge prediction result is obtained. According to the predicted results of the association between nodes in the document association graph and the initial edges in the document association graph, the cross-entropy loss is calculated as the first loss, as l ₂ shown in Figure 8.

Among them, l ₁ is used to update the parameters of the first graph neural network and language model, and l ₂ is used to update the parameters of the second graph neural network; or, based on l ₁ + l ₂ , update the first graph neural network, language model and second graph neural network. Parameters of graph neural networks.

It should be noted that when the structure of the document representation model is different, the processing flow of multi-document representation or model training based on the document representation model will be different. For details, please refer to the description of various optional model structures in the aforementioned embodiments. Various possible implementation modes can be learned through the combination of embodiments, and the details will not be described one by one in this embodiment.

It should be noted that in the aforementioned model training scheme, the semantic information learning module and the correlation information learning module of the document representation model are jointly trained. In other embodiments, the semantic information learning module and the correlation information learning module of the document representation model can be trained separately, and the semantic information learning module and the correlation information learning module are trained separately, and the trained semantic information learning module and correlation information learning module are Modules are combined to obtain a document representation model. In the process of training the semantic information learning module, there is no need to calculate the first loss. Only the aforementioned second loss and third loss are calculated. According to the second loss and the third loss, the parameters of the semantic information learning module are updated without calculating the first loss. In the process of training the correlation information learning module, the first loss is calculated, and the parameters of the document representation model are updated according to the first loss, without the need to calculate the second loss and the third loss. For the calculation method of each loss, please refer to the relevant content in the aforementioned model training embodiment, and will not be described again here.

FIG. 9 is a flow chart of a document characterization method provided by an exemplary embodiment of the present application. This embodiment is based on the system architecture shown in FIG. 1 , and the execution subject of this embodiment is the server in the system architecture shown in FIG. 1 . As shown in Figure 9, the specific steps of this method are as follows:

Step S901: Receive a document set submitted by the end-side device. The document set contains multiple documents to be processed.

In this embodiment, when the end-side device needs to characterize multiple documents in the document set to obtain the document representation during the process of running the downstream application, it submits the document set to the server. The server receives the document set submitted by the end-side device, thereby obtaining multiple documents to be processed contained in the document set.

For example, when the end-side device submits a document set to the server, it can upload the document set through the front-end interface; or, the end-side device submits a request containing the document set to the server, so that the server obtains the data contained in the request and returns the result. For example, the terminal device sends a calling request containing a document set to the server by calling the calling interface of the document representation service provided by the server. In addition, it can also be implemented by submitting files to the server from other terminals, which will not be described here.

Step S902: Construct a document association graph corresponding to the document set. The document association graph includes nodes corresponding to each document and edges representing association relationships between documents.

After obtaining a document set containing multiple documents to be processed, the server can automatically build a document association graph corresponding to the document set, that is, a document association graph of multiple documents in the document set. Specifically, for multiple documents included in the document set, nodes corresponding to each document are constructed; according to the association between documents, an edge is established between the nodes corresponding to any two documents with an association relationship, and the number of documents in the document set is obtained. Document association diagram. It should be noted that the same document association graph can contain one or more different types of edges, and different types of edges represent different association relationships.

Among them, the association relationship between documents represents the correlation between documents, which can be specific references, the same author, the same target object, the same description object, etc. For example, there can be reference relationships between paper-type documents, tender-type documents can belong to the same bidder/tenderer, knowledge-type documents can describe the same item, etc. In different application scenarios, for different document sets, the relationships between documents can be different. The correlation between different documents in the document set of a specific application scenario can be determined through data analysis and mining. The details can be obtained using existing technologies, and are not specifically limited here.

In this embodiment, through steps S901-S902, the process of obtaining multiple documents to be processed and the document association graphs of the multiple documents in step S301 is implemented. In another embodiment, the document association graph corresponding to the document set can also be constructed by the end-side device, and the end-side device submits the document set and the document association graph corresponding to the document set to the server.

Step S903: Input the content information and document association graphs of multiple documents into the document representation model, generate a semantic representation of each document based on the content information of each document through the document representation model, and use the semantic representation of each document to initialize the nodes in the document association graph. Feature representation, graph representation learning is performed based on the initialized document association graph, and the feature representation of each node is updated.

This step is consistent with the specific implementation manner of the aforementioned step S302. For details, please refer to the relevant content of the aforementioned embodiments and will not be described again here.

Step S904: Use the updated feature representation of each node as the document representation of the corresponding document.

This step is consistent with the specific implementation manner of the aforementioned step S303. For details, please refer to the relevant content of the aforementioned embodiments and will not be described again here.

Step S905: Send the document representation of each document in the document set to the end-side device.

In the method of this embodiment, the server receives a document set submitted by the end-side device, constructs a document association graph corresponding to the document set, and inputs the content information and document association graphs of multiple documents into the document representation model. The document representation model is based on the document representation of each document. The content information generates the semantic representation of each document, uses the semantic representation of each document to initialize the feature representation of each node in the document association graph, performs graph representation learning based on the initialized document association graph, and updates the feature representation of each node, which can achieve end-to-end Multi-document representation eliminates the need for end-side devices to provide complex data such as document association graphs.

In this embodiment, after obtaining the document representation of each document through document characterization, the server returns the document representation of each document in the document set to the end-side device. After receiving the document representation of each document in the document set, the end-side device can continue to execute the document processing logic of the downstream application based on the document representation of multiple documents, implement the document processing task of the downstream application, and obtain the document processing result.

Illustratively, taking the document classification scenario as an example, the end-side device can submit a document set containing multiple documents to be classified to the server. The server characterizes multiple documents in the document set to obtain the document representation of each document, and returns the document representation of each document to the end-side device. The end-side device performs document classification based on the document representation of each document and obtains the document classification result.

Illustratively, taking the document retrieval scenario as an example, the terminal device can submit a document set containing multiple documents in the retrieval library to the server. The server characterizes multiple documents contained in the document set to obtain document representations of multiple documents in the retrieval database, and returns the document representations of multiple documents in the retrieval database to the terminal device. The end-side device updates/stores document representations of multiple documents in the retrieval library. Further, the end-side device performs document retrieval based on the replaced retrieval database. For example, the end-side device maps the query information into a vector representation based on the query information input by the user, performs similarity matching between the vector representation and the document representation of each document in the retrieval database, and obtains one or more documents that match the query information, as Search results and output the search results.

When the method of this embodiment is applied to each downstream application, the server can continue to execute the set processing logic based on the document representations of multiple documents after obtaining the document representation of each document, obtain the document processing result, and return it to the end-side device. Document processing results.

For example, in a document classification scenario, the end-side device may submit a document set containing multiple documents to be classified to the server. After the server characterizes multiple documents in the document set to obtain the document representation of each document, the server classifies the multiple documents according to the document representation of each document, obtains the document classification result, and sends the document classification result to the end-side device.

For example, in a document association relationship prediction scenario, the end-side device may submit a document set containing multiple documents to be processed to the server. After the server characterizes multiple documents in the document set to obtain the document representation of each document, the server predicts the association between each document based on the document representation of each document, and sends the association between each document to the end-side device. Alternatively, after the server characterizes multiple documents in the document set to obtain the document representation of each document, it returns the document representation of each document in the document set to the end-side device. The end-side device predicts the association between each document based on the document representation of each document, and obtains the association between each document.

Exemplarily, in the document retrieval scenario, the server performs document representation on multiple documents in the document retrieval library based on the method of the aforementioned method embodiment, and obtains the document representation of each document; further, the server performs document representation on the multiple documents based on the document representation of the multiple documents. , update the document representation of each document in the document retrieval database.

Furthermore, the server can also implement online document retrieval processing based on the updated document retrieval database. Specifically, the server receives a document retrieval request sent by the end-side device, and the document retrieval request includes the input query information. The server maps the query information into a vector representation, performs similarity matching between the vector representation and the document representation of each document in the document retrieval database, obtains at least one target document that matches the query information, and sends at least one target document to the end-side device.

Figure 10 is a flow chart of a document retrieval method provided by another exemplary embodiment of the present application. The execution subject of this embodiment can run a local or cloud server with a document retrieval system. The server stores a document retrieval library. The document retrieval library The document representations of multiple documents in are obtained by characterizing the method of the foregoing document representation embodiment. As shown in Figure 10, the specific steps of this method are as follows:

Step S1001: Receive a document retrieval request sent by the end-side device. The document retrieval request includes the input query information.

The input query information may include at least one of the following: keywords entered by the user through the input box of the front-end interface, at least one information item selected by the user from the front-end interface (such as optional keywords, filtering conditions, etc.).

Step S1002: Map the query information into a vector representation, perform similarity matching between the vector representation and the document representation of each document in the document retrieval database, and obtain at least one target document that matches the query information.

Among them, mapping the query information into a vector representation can be implemented by any method in the prior art for converting text information into a vector representation, which is not specifically limited here.

When similarity matching is performed between the vector representation and the document representation of each document in the document retrieval database, the similarity between the vector representation of the query information and the document representation of each document in the document retrieval database can be calculated, and at least one document with higher similarity is selected. As a target document that matches the query information.

Optionally, the documents in the document retrieval database can be sorted in order of similarity from high to low, and a preset number of top-ranked documents are selected according to the sorting results as target documents matching the query information. The preset number can be configured and adjusted according to the needs of actual application scenarios and experience values, and there is no specific limit here.

Optionally, based on the similarity, documents whose similarity is greater than a preset similarity threshold can be selected as target documents matching the query information. The preset similarity threshold can be configured and adjusted according to the needs of actual application scenarios and experience values, and is not specifically limited here.

In this embodiment, the solution of the aforementioned embodiment is used to perform document characterization on multiple documents in the document retrieval database, and the document representation of each document is obtained. Specifically, multiple documents to be represented in the document retrieval database are obtained, as well as document association graphs of the multiple documents. The document association graph includes nodes corresponding to each document and edges representing the association relationships between documents; multiple documents and document association graphs are Input the document representation model, and use the document representation model to perform graph representation learning based on the document association graph initialized using the semantic representation of each document, and update the feature representation of each node; use the updated feature representation of each node as the document representation of the corresponding document. Further, based on the document representations of the multiple documents, the document representation of each document in the document retrieval database is updated. For the specific implementation process of characterizing multiple documents in the document retrieval database and obtaining the document representation of each document, please refer to the relevant content in the foregoing embodiments, which will not be described again in this embodiment.

Step S1003: Send at least one target document to the end-side device.

The server returns the retrieved at least one target document that matches the query information to the end-side device, so that the end-side device outputs relevant information of the retrieved at least one target document that matches the query information to the user. In addition, according to the configuration of the retrieval system, the server can send one or more key information of the target document that needs to be output to the user to the end-side device, so that the end-side device outputs one or more key information of the target document.

This embodiment provides a document retrieval method. By using the solution of the previous embodiment to perform document characterization on multiple documents in the document retrieval library, the document representation of each document is obtained, and the document retrieval library is updated according to the document representation of the multiple documents. The document representation of each document in the document can improve the quality of document representation in the document retrieval library, thereby improving the accuracy and recall rate of document retrieval.

Figure 11 is a schematic structural diagram of a server provided by an embodiment of the present application. As shown in Figure 11, the server in this embodiment may include: at least one processor 1101; and a memory 1102 communicatively connected to the at least one processor.

Among them, the memory 1102 stores instructions that can be executed by at least one processor 1101. The instructions are executed by at least one processor 1101, so that the server performs the method of any of the above embodiments. Its specific functions and the technical effects that can be achieved are similar. , which will not be described again here.

Optionally, the memory 1102 can be independent or integrated with the processor 1101. Optionally, as shown in Figure 11, the server also includes: a firewall 1103, a load balancer 1104, a communication component 1105, a power supply component 1106 and other other components. Only some components are schematically shown in Figure 11, which does not mean that the server only includes the components shown in Figure 11.

Embodiments of the present application also provide a computer-readable storage medium. Computer-executable instructions are stored in the computer-readable storage medium. When the processor executes the computer-executed instructions, the method of any of the foregoing embodiments is implemented. The specific functions and what can be achieved are The technical effects will not be repeated here.

An embodiment of the present application also provides a computer program product, including a computer program, which implements the method of any of the foregoing embodiments when executed by a processor. The computer program is stored in a readable storage medium. At least one processor of the server can read the computer program from the readable storage medium. At least one processor executes the computer program so that the server executes the technical solution provided by any of the above method embodiments. Specifically The functions and technical effects that can be achieved are not repeated here.

An embodiment of the present application provides a chip, which includes: a processing module and a communication interface. The processing module can execute the technical solution of the server in the foregoing method embodiment. Optionally, the chip also includes a storage module (such as a memory), the storage module is used to store instructions, the processing module is used to execute the instructions stored in the storage module, and the execution of the instructions stored in the storage module causes the processing module to perform any of the foregoing. A technical solution provided by a method embodiment.

The above integrated modules implemented in the form of software function modules can be stored in a computer-readable storage medium. The above-mentioned software function module is stored in a storage medium and includes a number of instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor to execute some steps of the methods of various embodiments of the present application.

It should be understood that the above-mentioned processor can be a processing unit (Central Processing Unit, CPU for short), or other general-purpose processor, Digital Signal Processor (Digital Signal Processor, DSP for short), Application Specific Integrated Circuit (Application Specific Integrated Circuit, for short) ASIC) etc. A general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc. The steps of the method disclosed in the application can be directly implemented by a hardware processor, or executed by a combination of hardware and software modules in the processor. The memory may include high-speed random access memory (RAM), or may also include non-volatile storage NVM, such as at least one disk memory, or it may be a USB flash drive, mobile hard disk, read-only memory, magnetic disk or optical disk wait.

The above-mentioned storage may be an object storage service (OSS).

The above memory can be implemented by any type of volatile or non-volatile storage device or their combination, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable and programmable read-only memory (EEPROM). Programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic or optical disk.

The above communication component is configured to facilitate wired or wireless communication between the device where the communication component is located and other devices. The device where the communication component is located can access wireless networks based on communication standards, such as mobile hotspots (WiFi), second-generation mobile communication systems (2G), third-generation mobile communication systems (3G), and fourth-generation mobile communication systems (4G). / Long Term Evolution (LTE), fifth generation mobile communication system (5G) and other mobile communication networks, or their combination. In an exemplary embodiment, the communication component receives broadcast signals or broadcast related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component further includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module can be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

The above-mentioned power supply component provides power to various components of the device where the power supply component is located. A power component may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the device in which the power component resides.

The above storage medium can be implemented by any type of volatile or non-volatile storage device or their combination, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable Except programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk. Storage media can be any available media that can be accessed by a general purpose or special purpose computer.

An exemplary storage medium is coupled to the processor such that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and storage medium may be located in Application Specific Integrated Circuits (ASIC). Of course, the processor and the storage medium may also exist as discrete components in an electronic device or a host control device.

It should be noted that, in this document, the terms "comprising", "comprises" or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article or device that includes a series of elements not only includes those elements, It also includes other elements not expressly listed or inherent in the process, method, article or apparatus. Without further limitation, an element defined by the statement "comprises a..." does not exclude the presence of additional identical elements in a process, method, article or apparatus that includes that element.

The order of the above embodiments of the present application is only for description and does not represent the advantages and disadvantages of the embodiments. In addition, some of the processes described in the above embodiments and drawings include multiple operations that appear in a specific order, but it should be clearly understood that these operations may not be performed in the order in which they appear in this document or may be performed in parallel. , is only used to distinguish different operations, and the sequence number itself does not represent any execution order. Additionally, these processes may include more or fewer operations, and the operations may be performed sequentially or in parallel. It should be noted that the descriptions such as "first" and "second" in this article are used to distinguish different messages, devices, modules, etc., and do not represent the order, nor do they limit "first" and "second" are different types. "Plural" means more than two, unless otherwise clearly and specifically limited.

Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus the necessary general hardware platform. Of course, it can also be implemented by hardware, but in many cases the former is better. implementation. Based on this understanding, the technical solution of the present application can be embodied in the form of a software product in essence or that contributes to the existing technology. The computer software product is stored in a storage medium (such as ROM/RAM, disk, CD), including several instructions to cause a terminal device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of various embodiments of the present application.

Other embodiments of the present application will be readily apparent to those skilled in the art from consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary technical means in the technical field that are not disclosed in this application. .

The above are only preferred embodiments of the present application, and are not intended to limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made using the contents of the description and drawings of the present application may be directly or indirectly used in other related technical fields. , are all equally included in the patent protection scope of this application.

Claims (16)

1. A method of document characterization, comprising:

acquiring a plurality of documents to be processed and a document association graph of the plurality of documents, wherein the document association graph comprises nodes corresponding to the documents and edges representing association relations among the documents;

inputting the content information of the plurality of documents and the document association graph into a document representation model, performing graph feature learning based on the document association graph initialized by using the semantic representation of each document through the document representation model, and updating the feature representation of each node;

the characteristic representation of each node after updating is used as the document representation of the corresponding document.

2. The method of claim 1, wherein the document representation model comprises: a semantic information learning module and a relevance information learning module,

the semantic information learning module is used for: carrying out semantic information characterization learning on the content information of each document to obtain semantic representation of each document;

the correlation information learning module is used for: initializing the characteristic representation of each node in the document association graph by using the semantic representation of each document, and carrying out graph feature learning based on the initialized document association graph to update the characteristic representation of each node.

3. The method of claim 2, wherein the semantic information learning module comprises: an entity relation diagram construction module, a first graph neural network, a text representation model and a semantic representation fusion module,

the entity relation diagram construction module is used for: constructing an entity relation diagram of each document according to the content information of each document;

the first graph neural network is used for: respectively performing graph sign learning on the entity relation graph of each document to obtain feature representations of the entities contained in each document, and fusing the feature representations of the entities contained in each document to obtain semantic representations of the entity levels of each document;

the text representation model is used for: respectively carrying out text representation on the content information of each document to obtain semantic representation of a text level of each document;

the semantic representation fusion module is used for: and fusing the semantic representation of the entity level and the semantic representation of the text level of each document to obtain the semantic representation of each document.

4. The method of claim 2, wherein the correlation information learning module comprises: the module and the second graph neural network are initialized,

The initialization module is used for: initializing feature representations of nodes in the document association graph by using semantic representations of the documents;

the second graph neural network is used for: and performing graph feature learning based on the initialized document association graph, and updating the feature representation of each node.

5. The method of claim 2, wherein said inputting the content information of the plurality of documents and the document association graph into a document representation model, through the document representation model, performing graph feature learning based on a document association graph initialized using semantic representations of the respective documents, updating feature representations of the respective nodes, comprises:

inputting the content information of the plurality of documents into the semantic information learning module, performing semantic information characterization learning on the content information of each document through the semantic information learning module to obtain semantic representation of each document, and inputting the semantic representation of each document into the relativity learning module;

inputting the document association graph into the correlation information learning module, initializing the characteristic representation of each node in the document association graph by using the semantic representation of each document through the correlation information learning module, and carrying out graph feature learning based on the initialized document association graph to update the characteristic representation of each node.

6. The method of claim 5, wherein the semantic information learning module comprises: an entity relationship graph construction module, a first graph neural network and a text representation model,

inputting the content information of the plurality of documents into the semantic information learning module, performing semantic information characterization learning on the content information of each document through the semantic information learning module to obtain semantic representation of each document, wherein the semantic representation comprises the following steps:

inputting the content information of the plurality of documents into the entity relation diagram construction module, and constructing an entity relation diagram of each document according to the content information of each document through the entity relation diagram construction module;

inputting the entity relation graph of each document into the first graph neural network, and respectively performing graph sign learning on the entity relation graph of each document through the first graph neural network to obtain semantic representation of the entity level of each document;

inputting the content information of the documents into the text representation model, and respectively carrying out text representation on the key content information of each document through the text representation model to obtain semantic representation of the text level of each document;

And fusing the semantic representation of the entity level and the semantic representation of the text level of each document to obtain the semantic representation of each document.

7. The method of claim 5, wherein the correlation information learning module comprises: the module and the second graph neural network are initialized,

inputting the document association graph into the correlation information learning module, initializing feature representations of nodes in the document association graph by the correlation information learning module by using semantic representations of the documents, and performing graph feature learning based on the initialized document association graph to update the feature representations of the nodes, wherein the method comprises the following steps:

inputting the semantic representation of each document and the document association graph into the initialization module, and initializing the characteristic representation of each node in the document association graph by using the semantic representation of each document through the initialization module;

inputting the initialized document association diagram into the second diagram neural network, performing diagram feature learning based on the initialized document association diagram through the second diagram neural network, and updating the feature representation of each node.

8. The method of claim 2, wherein the training process of the document representation model comprises:

Inputting a plurality of document samples in a document set for training and document associated graphs of the plurality of document samples into a document representation model, masking initial edges in the document associated graphs of the plurality of document samples through the document representation model, initializing feature representations of nodes in the document associated graphs of the plurality of document samples by using semantic representations of the document samples to obtain an initialized document associated graph, performing graph feature learning based on the initialized document associated graph, and updating the feature representations of the nodes;

predicting the association relation among the nodes in the document association graph according to the updated characteristic representation of each node;

and calculating a first loss according to a predicted result of the association relation among the nodes in the document association diagram and an initial edge in the document association diagram, and updating parameters of a document representation model according to the first loss.

9. The method of claim 8, wherein the semantic information learning module comprises: an entity relation diagram construction module, a first graph neural network, a text representation model and a semantic representation fusion module,

the using of the semantic representation of each of the document samples by the document representation model includes:

Inputting the content information of each document sample in the document set into the entity relation diagram construction module, and constructing an entity relation diagram of each document sample according to the content information of each document sample by the entity relation diagram construction module;

inputting the entity relation diagram of each document sample in the document set into a first graph neural network, and respectively performing graph sign learning on the entity relation diagram of each document sample through the first graph neural network to obtain semantic representation of the entity level of each document sample;

inputting key content information of a plurality of document samples of the document set into the text representation model, and respectively carrying out text representation on the key content information of each document sample through the text representation model to obtain semantic representation of a text level of each document sample;

and fusing the semantic representation of the entity level and the semantic representation of the text level of each document sample to obtain the semantic representation of each document sample.

10. The method of claim 8, wherein the document represents a training process of a model, comprising:

predicting class labels of the document samples according to semantic representations of entity levels of the document samples to obtain first prediction results, and predicting class labels of the document samples according to semantic representations of text levels of the document samples to obtain second prediction results;

Calculating the distance between the first predicted result and the second predicted result to obtain a second loss; calculating cross entropy loss according to the first prediction result and/or the second prediction result and the actual category labels of the document samples to obtain third loss;

and updating parameters of the semantic information learning module according to the second loss and the third loss.

11. The method of any of claims 1-10, wherein the obtaining a plurality of documents to be processed, and a document association graph of the plurality of documents, comprises:

receiving a document set submitted by a terminal side device, wherein the document set comprises a plurality of documents to be processed;

and constructing a document association diagram corresponding to the document set, wherein the document association diagram comprises nodes corresponding to the documents and edges representing association relations among the documents.

12. The method of claim 11, wherein after the step of using the updated feature representation of each node as a document representation of the corresponding document, further comprises:

classifying the plurality of documents according to the document representation of each document to obtain a document classification result, and sending the document classification result to the terminal side equipment;

Or,

and predicting the association relation among the documents according to the document representation of the documents, and sending the association relation among the documents to the terminal side equipment.

13. The method of any of claims 1-10, wherein the obtaining a plurality of documents to be processed, and a document association graph of the plurality of documents, comprises:

acquiring a plurality of documents contained in a document retrieval library and a document association graph of the plurality of documents;

after the updated characteristic representation of each node is used as the document representation of the corresponding document, the method further comprises the following steps:

updating the document representation of each document in the document retrieval library according to the document representations of the plurality of documents.

14. A document retrieval method, comprising:

responding to a document retrieval request, and acquiring query information input by a user;

mapping the query information into vector representations, and performing similarity matching on the vector representations and document representations of all documents in a document retrieval library to obtain at least one target document matched with the query information;

outputting information of the at least one target document;

wherein the document representation of each document in the document retrieval library is determined by: acquiring a plurality of documents to be characterized in a document retrieval library and document association graphs of the plurality of documents, wherein the document association graphs comprise nodes corresponding to the documents and edges representing association relations among the documents; inputting the plurality of documents and the document association graph into a document representation model, performing graph sign learning based on the document association graph initialized by using the semantic representation of each document through the document representation model, and updating the feature representation of each node; the characteristic representation of each node after updating is used as the document representation of the corresponding document.

15. A server, comprising:

at least one processor; and

a memory communicatively coupled to the at least one processor;

wherein the memory stores instructions executable by the at least one processor to cause the server to perform the method of any one of claims 1-14.

16. A computer readable storage medium having stored therein computer executable instructions which when executed by a processor implement the method of any of claims 1-14.