A multi-hop question answering system and method based on a graph neural network

Abstract

The application discloses a multi-hop question and answer system and method based on a graph neural network, and the system comprises a question screening module used for extracting training data from a multi-hop question and answer data set; a graph neural network coding-decoding module used for splitting the training data into a multi-hop question, a multi-hop answer and a support document, and then respectively performing preprocessing to obtain a paragraph, a sentence and an entity; and constructing an entity graph by taking the paragraph, the sentence and the entity as nodes; and used for interacting the entity graph with the multi-hop answer and the support document to obtain an answer of the multi-hop question; a node fusion replacement module used for calculating a conditional probability of node fusion replacement, updating the entity graph and obtaining a new multi-hop question, and used for judging the number of entities of the new multi-hop question; and the application can reduce the number of documents and the number of reasoning steps required by the multi-hop question and answer system to find a multi-hop answer, and can convert a multi-hop question with a larger number of hops into a new multi-hop question with a smaller number of hops, so that the overall accuracy of the multi-hop question and answer system is improved.

Description

Technical Field

[0001] This invention relates to the technical field of multi-hop question answering, and in particular to a multi-hop question answering system and method based on graph neural networks. Background Technology

[0002] With the continuous development of the information age, in order to obtain effective information more efficiently, question-answering systems that can provide concise and accurate answers to user input questions have become the research focus of many researchers. Among them, multi-hop question-answering systems that can answer complex questions are a research hotspot.

[0003] In multi-hop question answering systems, the number of reasoning steps required to arrive at the answer to a given complex question is called the "hop count". However, existing multi-hop question answering systems have low accuracy when dealing with some questions with very high hop counts. Summary of the Invention

[0004] The purpose of this invention is to solve the problem of low accuracy in existing multi-hop question answering systems when dealing with multi-hop questions with a high number of hops, and to provide a multi-hop question answering system and method based on graph neural networks.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A multi-hop question answering system based on graph neural networks includes a question filtering module, a graph neural network encoder-decoder module, and a node fusion and substitution module. The question filtering module extracts training data from a multi-hop question answering dataset. The graph neural network encoder-decoder module splits the training data into multi-hop questions, multi-hop answers, and supporting documents, and then preprocesses each to obtain paragraphs, sentences, and entities. It then constructs an entity graph using the paragraphs, sentences, and entities as nodes. The module also interacts with the entity graph and the multi-hop answers and supporting documents to obtain answers to the multi-hop questions. The node fusion and substitution module calculates the conditional probability of the node fusion substitution, updates the entity graph, and obtains new multi-hop questions. It also determines the number of entities in the new multi-hop questions.

[0007] In some embodiments of the present invention, the extraction of training data specifically includes: determining whether the multi-hop questions in the multi-hop question-answering dataset are labeled with the number of question hops; if the number of question hops is labeled, then determining the number of question hops in the multi-hop question; if the number of question hops is greater than or equal to 3, then extracting it as training data, otherwise discarding it; if the number of question hops is not labeled, then determining the number of entities in the multi-hop question through named entity recognition; if the number of entities is less than 4, then discarding it; if the number of entities is greater than or equal to 4, then determining the number of entity relations in the multi-hop question through sentence component analysis; if the number of entity relations is greater than or equal to 3, then extracting it as training data, otherwise discarding it.

[0008] In some embodiments of the present invention, the preprocessing includes sentence component analysis, paragraph splitting, and named entity recognition.

[0009] In some embodiments of the present invention, the graph neural network encoder-decoder module is further used to preprocess the multi-hop problem, and the preprocessing further includes labeling entities, entity relationships and multi-hop problem semantics in the multi-hop problem.

[0010] In some embodiments of the present invention, the graph neural network encoder-decoder module includes a graph neural network encoder and a graph neural network decoder. The graph neural network encoder is used to construct an entity graph using the paragraphs, sentences, and entities as nodes. The graph neural network decoder is used to interact the entity graph with the multi-hop answer and the supporting document to obtain the answer to the multi-hop question.

[0011] In some embodiments of the present invention, the graph neural network encoder-decoder module is further used to encode the paragraphs, sentences, and entities into multidimensional space vectors.

[0012] In some embodiments of the present invention, the graph neural network encoder-decoder module is further configured to use the highest-rated paragraph in the supporting document as a paragraph node and the sentences in the paragraph as sentence nodes.

[0013] In some embodiments of the present invention, updating the entity graph and obtaining a new multi-hop problem specifically includes: calculating the conditional probability of node fusion replacement, selecting the entity with the largest conditional probability value for entity replacement to obtain a new multi-hop problem, thereby reducing the number of entities contained in the new multi-hop problem by one.

[0014] In some embodiments of the present invention, the formula for calculating the conditional probability of node fusion replacement is as follows:

[0015]

[0016] Where e represents the natural logarithm, Q' represents the new multi-hop problem, and Q' is denoted as [q1, q2, ..., q]. m-1 q * q m+2 , ..., q i ], q m q m+1 p represents nodes in the entity graph that have mutual relationships. j s represents a paragraph node in the entity graph. k Represents paragraph node p j Sentence nodes in Represents paragraph node p j The corresponding entities (l = 1, 2, ..., k) of the sentence node sk in the text; q * yes A special case of p, representing the paragraph node. j Sentence node s in k Replacement entity; s(q * |q m q m+1 p j s k ) indicates that given q m q m+1 p j s k Under the condition that q* replaces node q m q m+1 The probability, Indicates that given q m q m+1 p j s k conditions, Replacement node q m q m+1 The probability of.

[0017] This invention also provides a multi-hop question answering method based on graph neural networks. Based on the multi-hop question answering system based on graph neural networks described above, the method includes the following steps: S1: Extracting training data from the multi-hop question answering dataset using the question filtering module; S2: Decomposing the training data into multi-hop questions, multi-hop answers, and supporting documents using the graph neural network encoder-decoder module, and preprocessing each to obtain paragraphs, sentences, and entities; S3: Constructing an entity graph using the paragraphs, sentences, and entities as nodes using the graph neural network encoder-decoder module; S4: Calculating the conditional probability of node fusion substitution using the node fusion substitution module, updating the entity graph, and obtaining a new multi-hop question; S5: Determining the number of entities in the new multi-hop question using the node fusion substitution module. If the number of entities in the new multi-hop question is greater than a specified value, repeating step S4; S6: Interacting the updated entity graph with the multi-hop answers and supporting documents using the graph neural network encoder-decoder module to obtain the answer to the new multi-hop question.

[0018] The present invention has the following beneficial effects:

[0019] The multi-hop question answering system proposed in this invention, based on graph neural networks, extracts training data from a multi-hop question answering dataset through a question filtering module. A graph neural network encoder-decoder module then splits the training data into multi-hop questions, multi-hop answers, and supporting documents, which are preprocessed to obtain paragraphs, sentences, and entities. An entity graph is constructed using these paragraphs, sentences, and entities as nodes. This entity graph interacts with the multi-hop answers and supporting documents to obtain the answer to the multi-hop question. A node fusion and substitution module reduces the number of entities in complex multi-hop questions with high hop counts, thereby reducing the number of documents and inference steps required for the multi-hop question answering system to find the answer. This transforms multi-hop questions with high hop counts into new multi-hop questions with fewer hop counts, thus improving the overall accuracy of the multi-hop question answering system.

[0020] Furthermore, in some embodiments of the present invention, the following beneficial effects are also achieved:

[0021] By selecting multi-hop questions with a high number of hops as training data, the quality of the training data can be improved, allowing the multi-hop question answering system to learn more about the characteristics of complex questions, thereby improving the performance of the multi-hop question answering system.

[0022] In an entity graph, each node has a corresponding position in the entity graph space after being encoded by a graph neural network, and semantic information will not be lost due to node fusion or replacement.

[0023] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0024] Figure 1 This is an overall framework diagram of the multi-hop question-answering system in an embodiment of the present invention;

[0025] Figure 2 This is a flowchart of the steps of the problem screening module in an embodiment of the present invention;

[0026] Figure 3 This is a flowchart of the steps of the graph neural network encoding-decoding module in an embodiment of the present invention;

[0027] Figure 4 This is a flowchart of the steps of the node fusion replacement module in an embodiment of the present invention;

[0028] Figure 5 This is a flowchart of the multi-hop question-and-answer system in an embodiment of the present invention. Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0030] It should be noted that the directional terms such as left, right, up, down, top, and bottom in this embodiment are only relative concepts or are based on the normal use of the product, and should not be considered as restrictive.

[0031] Definitions:

[0032] A question-answering system is a computer system that can automatically reason about a user's input question and provide an answer.

[0033] Multi-hop question answering refers to a question that requires multiple inferences and the retrieval of information from multiple paragraphs, sentences, or entities in a knowledge base before an answer can be provided.

[0034] A knowledge base is a document library containing objective knowledge. Its scope is uncertain; depending on the specific application scenario, it can be the entire Internet or a document library of a certain size.

[0035] Paragraphs, sentences, or entities in a knowledge base refer to the paragraphs, sentences, and entities (usually nouns) in each document within the knowledge base.

[0036] Graph neural network question answering systems encode questions, paragraphs, sentences, entities, etc., into nodes to construct a graph. Using deep learning technology, they reason about user-input questions and provide answers in an encoding-decoding manner.

[0037] Encoding refers to representing text as a multi-dimensional spatial vector, while decoding refers to decoding the multi-dimensional spatial vector representing the meaning of the text back into text.

[0038] A multi-level graph neural network question answering system refers to a system that encodes questions along with multiple paragraphs, sentences, entities, etc., that is, it uses paragraph-level, sentence-level, and entity-level information simultaneously during the construction of the graph neural network.

[0039] Existing multi-hop question answering systems can be broadly categorized into graph neural network-based systems and sequence prediction-based systems. Existing methods show a significant decrease in accuracy when handling questions with a large number of hops (≥3 hops) compared to 2-hop and 1-hop questions. Therefore, this invention proposes a simplified multi-hop question answering reasoning method. For graph neural network-based multi-hop question answering systems, when handling questions with 3 or more hops, the method improves the system's accuracy by fusing and replacing the nodes corresponding to entities in the node graph.

[0040] This invention chooses a multi-hop question answering system based on graph neural networks rather than one based on sequence prediction for two reasons. First, systems based on sequence prediction are limited by the model's memory capacity, resulting in limited semantic understanding and encoding-decoding capabilities for long texts. Second, systems based on sequence prediction introduce a certain degree of distortion during the fusion of entity-corresponding vector sequences. In contrast, in a system based on graph neural networks, each node in the node graph has a corresponding position in the graph space after being encoded by the graph neural network, preventing the loss of semantic information due to node fusion and replacement.

[0041] To address this, this invention proposes a multi-hop question-answering system based on graph neural networks and a simplified reasoning method for multi-hop question answering. For user-generated questions requiring multiple inference steps, relevant supporting documents are searched from a given knowledge base. By constructing nodes for multi-hop questions based on entities and calculating conditional probabilities, two related entities in the entity graph are replaced with a single entity. This reduces the number of entities and inference steps in multi-hop questions, improving encoding-decoding accuracy and the performance of the multi-hop question-answering system. The knowledge base in this embodiment contains 100,000 documents, each with 200-500 characters.

[0042] The multi-hop question answering system based on graph neural networks proposed in this invention includes a question selection module, a graph neural network encoder-decoder module, and a node fusion and substitution module, such as... Figure 1 As shown, the following configuration is defined: the question filtering module is used to extract training data from the multi-hop question-answering dataset; the graph neural network encoder-decoder module is used to split the training data into multi-hop questions, multi-hop answers, and supporting documents, and then preprocess them to obtain paragraphs, sentences, and entities; and construct an entity graph using the paragraphs, sentences, and entities as nodes; it is also used to interact the entity graph with the multi-hop answers and supporting documents to obtain the answers to the multi-hop questions; the node fusion and substitution module is used to calculate the conditional probability of the node fusion and substitution, update the entity graph and obtain new multi-hop questions, and also to determine the number of entities in the new multi-hop questions.

[0043] The specific functions are as follows:

[0044] I. Problem Screening Module

[0045] The main task of the problem filtering system is to determine the number of hops in a problem in order to filter out problems with 3 or more hops. The process is as follows: Figure 2As shown: Training data is extracted from the multi-hop question-answering dataset. If the question hop count is labeled in the initial multi-hop question-answering dataset, multi-hop question-answer pairs with 3 or more hops are selected as training data; otherwise, they are discarded. If the question hop count is not labeled in the initial multi-hop question-answering dataset, firstly, named entity recognition is used to determine the number of entities in the question. If the number of entities in the question is greater than or equal to 4, the question is likely to be a 3-hop or higher question; otherwise, it is discarded. Then, sentence component analysis is used to determine the number of relationships between entities. If the number of relationships between entities in the question is greater than or equal to 3, the question is likely to be a 3-hop or higher question; otherwise, it is discarded.

[0046] The question selection module can select multi-hop questions with a high number of hops as training data to train the multi-hop question answering system, which improves the quality of the training data and allows the multi-hop question answering system to learn more about the characteristics of complex questions, thereby improving the performance of the multi-hop question answering system.

[0047] II. Graph Neural Network Encoder-Decoder Module

[0048] The role of the graph neural network encoder-decoder module is to encode the input into a graph structure using a graph neural network model and then decode the output. The processing flow of the graph neural network encoder-decoder module is as follows: Figure 3 As shown, the training data, filtered by the question filtering module, is first input into the graph neural network encoder-decoder module. This module preprocesses the training data, including sentence component analysis, paragraph segmentation, and named entity recognition, splitting the training data into multi-hop questions, multi-hop answers, and supporting documents. The module also preprocesses the multi-hop questions, including labeling entities, entity relationships, and multi-hop question semantics. The graph neural network encoder-decoder module includes a graph neural network encoder and a neural network decoder. The training data is input into the graph neural network encoder, which encodes paragraphs, sentences, and entities into multi-dimensional spatial vectors and constructs an entity graph using paragraphs, sentences, and entities as nodes, thereby obtaining the entity graph and supporting document vector representations. The entity graph is updated through a node fusion and replacement module. The updated entity graph and supporting document vector representations are then input into the graph neural network decoder to decode and obtain the answers to the multi-hop questions.

[0049] Preferably, the graph neural network encoder-decoder module also uses the highest-rated paragraph in the supporting document as a paragraph node and the sentences in the paragraph as sentence nodes.

[0050] In the embodiments of the present invention Figure 3The example demonstrates the question: "What brand of sports car did Tom's deskmate buy?". Using named entity recognition and sentence component analysis tools in the question filtering system, the given question is encoded in a graph neural network encoder, resulting in entity graph A. "Tom," "deskmate," "sports car," and "brand" are nodes in entity graph A, and edges connecting these nodes ("Tom-deskmate," "deskmate-sports car," and "sports car-brand") represent the relationships between them. After applying the node fusion and replacement method proposed in this embodiment, entity graph B is obtained, where the nodes "Tom" and "deskmate" are replaced with "Sandy" (Tom's deskmate). Finally, entity graph B interacts with the vector representation of the supporting document, and the graph neural network decoder decodes the question to obtain the answer.

[0051] III. Node Fusion Replacement Module

[0052] The processing flow of the node fusion replacement module is as follows: Figure 4 As shown, the steps are as follows:

[0053] Step 1: Problem Extraction

[0054] Users upload task code, initial multi-hop question-answering dataset, and multi-hop question-answering model to the system, extract multi-hop questions from the initial dataset, select multi-hop questions that meet the requirements through the question filtering module, and preprocess the questions through methods such as sentence component analysis and named entity recognition.

[0055] Step 2: Construct the entity diagram

[0056] Construct nodes q1, q2, ..., q for each entity in the multi-hop problem. i If we take the modifier relation as an edge, then the multi-hop problem Q is represented as [q1, q2, ..., q...]. i (i≥3), the relationships between entities are represented as r 12 r 13 , ..., r 1i r 23 r 24 , ..., r 2i , ..., r (i-1)i (r ii’ Does it exist based on the actual situation, where r ii’ The relationship between the i-th entity and the i′-th entity (i≠i′) allows us to construct an entity graph (see [link]). Figure 3 ).

[0057] Step 3: Document Retrieval

[0058] For each entity node in the entity graph, retrieve the j highest-rated paragraphs from the supporting documents, and encode each paragraph as a paragraph node: P1, P2, ..., pj using a graph neural network encoder. j For each paragraph node P1, P2, ..., p j The k sentences are encoded into sentence nodes: S1, S2, ..., S3 using a graph neural network encoder. k .

[0059] Step 4: Probability Calculation

[0060] The conditional probability P(Q'|q) of node fusion replacement is calculated. m q n p j s k ), where Q′ represents the new multi-hop problem, denoted as [q1, q2, ..., q m-1 q * q m+2 , ..., q i ], q * For paragraph P j Sentence S k The corresponding entity in the entity graph replaces two related nodes: q m q m+1 If we select the new multi-hop question Q' with the largest conditional probability value, the number of entities contained in the new multi-hop question Q' changes from i to (i-1), which reduces the number of reasoning steps required for the multi-hop question answering system to answer the question by 1.

[0061] The formula for calculating the conditional probability of node fusion and replacement is:

[0062]

[0063] Where e is the natural logarithm. Represents paragraph node p j Sentence node s in k The corresponding entities (l = 1, 2, ..., k); q * yes A special case, representing the paragraph node p j Sentence node s in k Replacement entity; s(q * |q m q m+1 p j s k ) refers to the situation where q is given m q m+1 p j s k Under the condition that q* replaces q m qm+1 The probability of (taking values ​​[0, 1]) This refers to the situation where q is given. m q m+1 p j s k conditions, Replace q m q m+1 The probability of (values ​​[0, 1]).

[0064] Step 5: Controlling the number of reasoning steps

[0065] Depending on whether the number of entities i contained in the new multi-hop problem Q' is less than a given value (e.g., given value l = 4), if i ≤ l, proceed to step 6; if i > l, repeat steps 3-4.

[0066] Step 6: For the new multi-hop problem Q' with i ≤ l, construct an updated entity graph based on its entity correspondence (see...). Figure 3 It interacts with the multi-hop answer and supporting documentation, and uses a graph neural network decoder to decode and obtain the answer to the multi-hop question.

[0067] The multi-hop question-answering system proposed in this invention is a computer system in which users can ask questions and input them into the system. The system then uses methods such as keyword matching and word frequency retrieval to retrieve relevant supporting documents from a given knowledge base.

[0068] This invention also provides a multi-hop question answering method based on graph neural networks, based on the multi-hop question answering system based on graph neural networks described above, such as... Figure 5 As shown, the process includes the following steps: S1: Extract training data from the multi-hop question-answering dataset using the question filtering module; S2: Split the training data into multi-hop questions, multi-hop answers, and supporting documents using the graph neural network encoder-decoder module, and preprocess them to obtain paragraphs, sentences, and entities respectively; S3: Construct an entity graph using the paragraphs, sentences, and entities as nodes using the graph neural network encoder-decoder module; S4: Calculate the conditional probability of the node fusion substitution using the node fusion substitution module, update the entity graph, and obtain a new multi-hop question; S5: Determine the number of entities in the new multi-hop question using the node fusion substitution module. If the number of entities in the new multi-hop question is greater than a specified value, repeat step S4; S6: Interact the updated entity graph with the multi-hop answers and supporting documents using the graph neural network encoder-decoder module to obtain the answer to the new multi-hop question.

[0069] Preferably, the question-answering system annotates user-input questions according to the entities involved, the correspondence between entities, and the semantics of the question. The multi-hop question-answering system uses methods such as sentence component analysis, paragraph segmentation, and named entity recognition to perform preliminary processing on the user-input multi-hop questions and related supporting documents, obtaining the entities and correspondences in the multi-hop questions. It then performs paragraph segmentation and entity recognition, dividing paragraphs into sentences and labeling entities. Finally, it obtains labeled text at three levels: paragraphs, sentences, and entities.

[0070] Preferably, the question-answering system encodes the various paragraphs, sentences, and entities obtained above into multi-dimensional spatial vectors through a graph neural network, which are represented as nodes in the graph neural network.

[0071] Preferably, the multi-hop question-answering system calculates the multi-hop question Q (Q = [q1, q2, ... q...). i Two related entities in the document are merged and replaced with the conditional probabilities of entities in the supporting document, and P1, P2, ... P are determined by calculation. j S1, S2, ... S k The most suitable entity to replace entity q in Q is * The entity in the supporting document with the highest conditional probability is selected for entity replacement, resulting in the new multi-hop problem Q′.

[0072] Preferably, for the final obtained new multi-hop problem Q' and the supporting documents, paragraphs and entities with the highest degree of relevance to it, the new multi-hop problem Q' and the paragraphs and sentences are jointly represented as [Q′, P1, ..., P′, S1, ..., S′], which are used as nodes in the graph neural network and input into the graph neural network decoder to obtain the solution to the multi-hop problem.

[0073] The embodiments of the present invention have the following characteristics:

[0074] 1. Existing multi-hop question answering systems suffer from reduced accuracy in graph neural network encoding-decoding when handling questions with three or more hops. This is because the encoding of multiple entities involves multiple paragraphs, sentences, and entities, resulting in a large concatenated text length, which limits the performance of multi-hop question answering systems. The multi-hop question answering system proposed in this invention reduces the number of entities in complex questions through one or more iterations. This reduces the number of documents and inference steps required for the system to find the answer, transforming multi-level inference problems into fewer-level inference problems and improving the overall accuracy of the multi-hop question answering system.

[0075] 2. The question filtering module for estimating the number of hops in multi-hop questions proposed in this embodiment of the invention provides a reference approach for classifying and processing multi-hop questions when dealing with complex data composed of questions with different numbers of hops in a multi-hop question answering system.

[0076] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, several equivalent substitutions or obvious modifications can be made without departing from the concept of the present invention, and all such modifications, achieving the same performance or purpose, should be considered within the scope of protection of the present invention.

Claims

1. A multi-hop question-answering system based on graph neural networks, characterized in that, It includes a problem selection module, a graph neural network encoder-decoder module, and a node fusion and replacement module, among which: The question filtering module is used to extract training data from the multi-hop question-answering dataset; The graph neural network encoder-decoder module is used to split the training data into multi-hop questions, multi-hop answers, and supporting documents, and then preprocess them to obtain paragraphs, sentences, and entities; and to construct an entity graph using the paragraphs, sentences, and entities as nodes; and to interact the entity graph with the multi-hop answers and supporting documents to obtain the answers to the multi-hop questions; By selecting special entities to replace entity pairs in the entity graph, the conditional probability of node fusion replacement is calculated, thereby updating the entity graph and generating a new multi-hop problem with a reduced number of entities, and determining the number of entities in the new multi-hop problem.

2. The multi-hop question-answering system based on graph neural networks according to claim 1, characterized in that, The extraction of training data specifically includes: determining whether the multi-hop questions in the multi-hop question-answering dataset are labeled with the number of question hops; if the number of question hops is labeled, then determining the number of question hops in the multi-hop question; if the number of question hops is greater than or equal to 3, then extracting it as training data, otherwise discarding it; if the number of question hops is not labeled, then determining the number of entities in the multi-hop question through named entity recognition; if the number of entities is less than 4, then discarding it; if the number of entities is greater than or equal to 4, then determining the number of entity relations in the multi-hop question through sentence component analysis; if the number of entity relations is greater than or equal to 3, then extracting it as training data, otherwise discarding it.

3. The multi-hop question-answering system based on graph neural networks according to claim 1, characterized in that, The preprocessing includes sentence component analysis, paragraph splitting, and named entity recognition.

4. The multi-hop question-answering system based on graph neural networks according to claim 3, characterized in that, The graph neural network encoder-decoder module is also used to preprocess the multi-hop problem, and the preprocessing includes labeling entities, entity relationships and multi-hop problem semantics in the multi-hop problem.

5. The multi-hop question-answering system based on graph neural networks according to claim 1, characterized in that, The graph neural network encoder-decoder module includes a graph neural network encoder and a graph neural network decoder. The graph neural network encoder is used to construct an entity graph using the paragraphs, sentences, and entities as nodes. The graph neural network decoder is used to interact the entity graph with the multi-hop answer and supporting documents to obtain the answer to the multi-hop question.

6. The multi-hop question-answering system based on graph neural networks according to claim 5, characterized in that, The graph neural network encoder-decoder module is also used to encode the paragraphs, sentences, and entities into multidimensional space vectors.

7. The multi-hop question-answering system based on graph neural networks according to claim 6, characterized in that, The graph neural network encoder-decoder module is also used to take the highest-rated paragraph in the supporting document as a paragraph node and the sentences in the paragraph as sentence nodes.

8. The multi-hop question-answering system based on graph neural networks according to claim 1, characterized in that, The process of updating the entity graph and obtaining a new multi-hop problem specifically includes: calculating the conditional probability of node fusion replacement, selecting the entity with the largest conditional probability value for entity replacement to obtain a new multi-hop problem, thereby reducing the number of entities contained in the new multi-hop problem by one.

9. The multi-hop question-answering system based on graph neural networks according to claim 8, characterized in that, The formula for calculating the conditional probability of node fusion replacement is: ; in Represents the natural logarithm. This indicates the new multi-hop problem. Represented as , This represents nodes in the entity graph that have mutual relationships. This represents a paragraph node in the entity graph. Represents paragraph nodes Sentence nodes in Represents paragraph nodes Sentence nodes in Each corresponding entity ; yes A special case, representing a paragraph node. Sentence nodes in The replacement entity; Indicates that in a given conditions, *Alternative Node The probability, Indicates that in a given conditions, Alternate Nodes The probability of.

10. A multi-hop question answering method based on graph neural networks, characterized in that, The multi-hop question answering system based on graph neural networks according to any one of claims 1 to 9 includes the following steps: S1: Extract training data from the multi-hop question-answering dataset using the question filtering module; S2: The training data is split into multi-hop questions, multi-hop answers, and supporting documents by the graph neural network encoder-decoder module, and preprocessed to obtain paragraphs, sentences, and entities respectively; S3: Construct an entity graph using the paragraphs, sentences, and entities as nodes through the graph neural network encoder-decoder module; S4: Calculate the conditional probability of node fusion and replacement through the node fusion and replacement module, update the entity graph, and obtain a new multi-hop problem; S5: Determine the number of entities in the new multi-hop problem through the node fusion replacement module. If the number of entities in the new multi-hop problem is greater than a specified value, repeat step S4. S6: The updated entity graph is interacted with the multi-hop answer and supporting documents through the graph neural network encoder-decoder module to obtain the answer to the new multi-hop question.

Images