Complex declaration fact checking method and system based on multi-hop retrieval and reasoning

Abstract

The invention provides a complex declaration fact checking method and system based on multi-hop retrieval and reasoning, and the method comprises the steps: obtaining a to-be-checked declaration statement, and constructing an evidence set corresponding to the declaration statement through a plurality of retrieval rounds; in each retrieval round, the declaration statement and an evidence statement in a current evidence set are combined, retrieval is conducted from a database through a constraint generation type retriever, the retrieved evidence statement is obtained, and the evidence statement is added into the evidence set; combining the declaration statement and each evidence statement in the evidence set into a declaration evidence pair, and inputting the declaration evidence pair into a recursive memory model for intrasentence reasoning to obtain a corresponding declaration evidence pair vector; constructing a multi-hop reasoning sub-graph corresponding to each evidence statement based on the declaration evidence pair vector, and obtaining an inter-sentence reasoning vector based on all the multi-hop reasoning sub-graphs; and inputting the inter-sentence reasoning vector into a preset MLP model to obtain a check result of the declaration statement.

Description

Technical Field

[0001] The present invention relates to the technical field of fact-checking, and in particular, to a method and system for complex claim fact-checking based on multi-hop retrieval and reasoning. Background Art

[0002] With the rapid development of social networks, the speed and breadth of information dissemination have been significantly improved. However, this has also given rise to the large-scale spread of false information. Fact-checking systems aim to effectively curb the spread of false information by accurately retrieving relevant evidence and systematically verifying the authenticity of claims, thereby providing guarantees for the credibility and transparency of the information environment. Their main goal is to assist humans in judging the authenticity of information through automated means, thereby improving the reliability of social media content. Traditional fact-checking methods usually adopt a two-stage paradigm: first, retrieve evidence related to the claim from a large-scale corpus, and then judge the authenticity of the claim by analyzing the semantic relationship between the claim and the evidence.

[0003] However, most traditional methods rely on the one-to-one relevance between claims and evidence. In existing methods, single-hop retrieval usually only focuses on the semantic similarity between a claim and a single piece of evidence, and it is difficult to effectively capture the interaction information between multiple pieces of evidence; although multi-hop retrieval attempts to integrate evidence through an iterative approach, it faces problems such as intention drift, noise accumulation, and dependence on hyperlinks. At the same time, the challenge in the reasoning stage lies in the difficulty of simultaneously capturing local semantic features and the global structural relationship between multiple pieces of evidence, resulting in limited accuracy in complex claim verification. Summary of the Invention

[0004] In view of this, embodiments of the present invention provide a method for complex claim fact-checking based on multi-hop retrieval and reasoning to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present invention provides a method for complex claim fact-checking based on multi-hop retrieval and reasoning. The steps of the method include:

[0006] Obtain the claim statement to be fact-checked, and construct an evidence set corresponding to the claim statement through multiple retrieval rounds;

[0007] In each retrieval round, combine the claim statement and the evidence statements in the current evidence set, and retrieve from the database through a constrained generative retriever to obtain the retrieved evidence statements, and add the evidence statements to the evidence set;

[0008] Combine the claim statement with each evidence statement in the evidence set into a claim-evidence pair, and input each claim-evidence pair into a recursive memory model for intra-sentence reasoning to obtain the corresponding claim-evidence pair vector;

[0009] Construct a multi-hop inference subgraph corresponding to each evidence statement for the vector based on the stated evidence, and obtain an inter-sentence inference vector based on all the multi-hop inference subgraphs;

[0010] Input the inter-sentence inference vector into a pre-set MLP model to obtain the verification result of the statement sentence.

[0011] Adopting the above solution, this solution focuses on solving the intention deviation in multi-hop retrieval and the complex association modeling problem in multi-hop inference. These problems are the core challenges that lead to the low retrieval efficiency and limited inference ability of existing models when dealing with complex statements. Specifically, this solution designs a hierarchical structure of intra-sentence inference and inter-sentence inference. In intra-sentence inference, it uses recurrent memory to capture long-distance semantic dependencies, and in inter-sentence inference, it constructs a multi-hop evidence subgraph based on latent edge learning and graph neural networks to model complex cross-evidence semantic associations, capturing local semantic features and the global structural relationship between multiple pieces of evidence.

[0012] In some embodiments of the present invention, in the step of combining the statement sentence and the evidence sentences in the current evidence set, and retrieving the retrieved evidence sentences from the database through a constrained generative retriever:

[0013] Combine the statement sentence and the evidence sentences in the current evidence set, and input the combined sentence into a pre-set query compressor to obtain a compressed combined sentence;

[0014] Input the compressed combined sentence into a constrained generative retriever to obtain the retrieved evidence sentences.

[0015] In some embodiments of the present invention, in the step of combining each statement sentence with each evidence sentence in the evidence set into a statement-evidence pair, and inputting each statement-evidence pair into a recursive memory model for intra-sentence inference to obtain a corresponding statement-evidence pair vector, each statement-evidence pair is input into a recursive memory model including multiple transformer layers to obtain a corresponding statement-evidence pair vector.

[0016] In some embodiments of the present invention, in the step of sequentially numbering the evidence sentences in the evidence set based on the order of addition of the evidence sentences, and in the step of constructing a multi-hop inference subgraph corresponding to each evidence sentence based on the statement-evidence pair vector:

[0017] In the multi-hop inference subgraph corresponding to the evidence sentence, it includes the corresponding nodes of any one of the evidence sentences with the pre-order numbers of the number of this evidence sentence, and multiple path edges are sequentially connected to each other based on the numbers of the evidence sentences in the multi-hop inference subgraph;

[0018] Construct potential connections between nodes based on the similarity of the vectors of the declarative evidences corresponding to each node, and obtain the final multi-hop inference subgraph.

[0019] In some embodiments of the present invention, in the step of constructing potential connections between nodes based on the similarity of the vectors of the declarative evidences corresponding to each node and obtaining the final multi-hop inference subgraph, calculate the similarity based on the vectors of the declarative evidences corresponding to the nodes, and compare it with a preset similarity threshold to determine whether to construct potential connections between two nodes.

[0020] In some embodiments of the present invention, in the step of obtaining the inter-sentence inference vector based on all the multi-hop inference subgraphs, process each multi-hop inference subgraph through a graph attention layer, and aggregate all the multi-hop inference subgraphs to obtain the inter-sentence inference vector.

[0021] In some embodiments of the present invention, the steps of the method further include pre-training a query compressor, and the steps of pre-training the query compressor include distillation and alignment processing;

[0022] During the distillation process, use a preset large language model to process the combined statements of declarative statements and evidence statements to generate a compressed text as the learning target of the query compressor;

[0023] Use a preset large language model to calculate the semantic matching degree between the compressed text and the original declarative statement, take the compressed text with the highest score as the positive sample, and take the remaining compressed texts as negative samples;

[0024] Train the query compressor based on the positive samples and negative samples.

[0025] In some embodiments of the present invention, in the step of training the query compressor based on the positive samples and negative samples, input the combined statements corresponding to the positive samples and negative samples into the query compressor to obtain the compressed combined statements, calculate the loss function based on the compressed combined statements and the positive samples and negative samples, and train the query compressor based on the loss function value.

[0026] In some embodiments of the present invention, in the step of calculating the loss function based on the compressed combined statements and the positive samples and negative samples, calculate the loss function based on the following formula:

[0027]

[0028] where, L a represents the value of the loss function, N represents the number of combined statements, n represents the number of negative samples, Q i represents the i-th compressed combined statement, Q + represents the positive sample, Denote the j-th negative sample, f(·) denote the encoder, and τ denote the temperature coefficient.

[0029] The second aspect of the present invention further provides a complex claim fact-checking system based on multi-hop retrieval and reasoning. The system includes a computer device, which includes a processor and a memory. Computer instructions are stored in the memory, and the processor is configured to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method described above.

[0030] The third aspect of the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps implemented by the aforementioned complex claim fact-checking method based on multi-hop retrieval and reasoning.

[0031] The additional advantages, objectives, and features of the present invention will be partially elaborated in the following description, and will become partially apparent to those of ordinary skill in the art after studying the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the present invention can be pointed out and obtained specifically in the description and the accompanying drawings.

[0032] Those skilled in the art will understand that the objectives and advantages that can be achieved by the present invention are not limited to the above specific descriptions, and the above and other objectives that the present invention can achieve will be more clearly understood according to the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings described herein are used to provide a further understanding of the present invention, form a part of this application, and do not limit the present invention.

[0034] Figure 1 It is a schematic diagram of an embodiment of the complex claim fact-checking method based on multi-hop retrieval and reasoning of the present invention;

[0035] Figure 2 It is a schematic diagram of the processing architecture of the complex claim fact-checking method based on multi-hop retrieval and reasoning of the present invention;

[0036] Figure 3 It is a schematic diagram of the processing of the pre-trained query compressor of the present invention;

[0037] Figure 4 It is a schematic diagram of the processing of the recursive memory model of the present invention;

[0038] Figure 5 It is a schematic diagram of the multi-hop inference subgraph of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0039] To make the objectives, technical solutions, and advantages of the present invention more clear and understandable, the present invention will be further described in detail below in conjunction with the embodiments and the drawings. Herein, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but not to limit the present invention.

[0040] Herein, it should also be noted that in order to avoid obscuring the present invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the present invention are shown in the drawings, while other details less relevant to the present invention are omitted.

[0041] Introduction to the prior art:

[0042] Evidence retrieval is a fundamental step in fact verification, and its goal is to identify supporting or refuting evidence related to a claim from a large-scale corpus. Traditional methods such as TF-IDF and BM25 achieve retrieval by evaluating the independent relevance between a claim and a single piece of evidence, but fail to capture the interaction information between evidence. Therefore, single-hop retrieval cannot effectively handle complex claims that require comprehensive verification of multiple pieces of evidence. To make up for this deficiency, multi-hop retrieval has gradually become a research hotspot. Currently, most of these methods rely on inter-document hyperlinks or dense retrieval strategies, and gradually integrate the information from the previous hop through iterative retrieval to construct a complete evidence chain. For example, HESM uses inter-document hyperlinks to achieve multi-hop retrieval to improve accuracy, and GMR uses the retrieved evidence to expand the original input query, thereby enhancing the retrieval of subsequent evidence.

[0043] The core of claim verification is to determine whether a claim is supported or refuted by analyzing the semantic relationship between the claim and the retrieved evidence. Existing methods are mainly divided into two categories: based on Transformer and based on graph structure. MLA adopts a fact-checking framework based on Transformer, which uses a multi-level attention mechanism to better capture the relevance and importance of different pieces of evidence related to the claim. Transformer-based methods, such as BERT and RoBERTa, can capture the local semantic associations between a claim and evidence, but perform poorly in dealing with complex interactions between multiple pieces of evidence. In contrast, graph-structure-based methods construct an evidence graph (such as at the sentence level or the lexical level) and use graph neural networks to reason about the global associations between multiple pieces of evidence, thereby better expressing the global semantics to support the verification of complex claims. GET constructs a fine-grained token-level evidence graph to detect the semantic associations between texts, so as to conduct a more detailed understanding and analysis of the evidence semantics and ensure that subtle but crucial clues are not overlooked.

[0044] In addition, to bridge the gap between evidence retrieval and claim verification, some studies have attempted to integrate these two stages by jointly optimizing retrieval and reasoning to improve performance. Both FER and RAV adopt a joint training mechanism, using claim verification results to provide optimized training supervision signals for evidence retrieval, enabling the retriever to focus more on feature information directly relevant to the claim verification goal.

[0045] Disadvantages of the prior art:

[0046] Currently, complex claim verification poses higher requirements on existing fact-checking systems, which usually require the joint support of multiple pieces of evidence and logical reasoning. Although existing methods have made some progress in this field, there are still significant challenges in dealing with multi-hop information retrieval and complex reasoning.

[0047] At the evidence retrieval level, most existing methods adopt a single-hop approach, only considering the semantic similarity between the claim and a single piece of evidence. This method ignores the potential interaction information between evidence and is difficult to effectively handle complex claims that require joint verification of multiple pieces of evidence. Most multi-hop retrieval methods rely on hyperlink or dense retrieval strategies: Hyperlink-based multi-hop retrieval strongly depends on entity linking or hyperlinks between documents, severely limiting the application scenarios. Dense retrieval-based methods combine the original claim and the evidence retrieved in the previous steps to enhance the original claim and then perform subsequent retrieval. In this way, as the number of retrieval hops increases, additional irrelevant information may gradually accumulate, thus having a negative impact on the retrieval quality; the accumulated context will also cause the query semantics to gradually deviate from the core intention of the original claim, thereby reducing the accuracy of subsequent retrievals.

[0048] At the claim verification level, existing methods mainly include Transformer-based sequence reasoning and graph structure-based reasoning. Sequence reasoning methods are good at capturing local context information but are insufficient in dealing with global interactions between multiple pieces of evidence. For the complex logical associations of multi-hop evidence, sequence models are difficult to provide sufficient semantic integration. Although graph structure methods can capture multiple evidence nodes and their relationships, in sparse graph structures, their ability to learn semantic relationships between nodes may be limited. In addition, most current reasoning methods fail to integrate the advantages of sequences and graph structures and are difficult to capture both local intra-sentence features and cross-sentence structured relationships simultaneously. This limitation makes the model perform poorly in dealing with complex claims.

[0049] In summary, the disadvantages of the prior art mainly focus on the following aspects: First, in the multi-hop retrieval process, existing methods usually generate the next-hop query through a simple text splicing method, which easily leads to the problem of semantic intention deviation. As the number of retrieval hops increases, the relevance of the retrieval results decreases significantly. Second, in the inference stage, sequence models are difficult to capture the global relationships across evidence, while graph models, although able to construct the structured relationships between evidence, lack the ability to finely model the intra-sentence semantics.

[0050] As Figure 1 and 2 shown, the present invention proposes a complex claim fact-checking method based on multi-hop retrieval and inference. The steps of this method include:

[0051] Step S100, obtain the claim statement to be verified, and construct an evidence set corresponding to the claim statement through multiple retrieval rounds;

[0052] In each retrieval round, combine the claim statement and the evidence statements in the current evidence set, and retrieve from the database through a constrained generative retriever to obtain the retrieved evidence statements, and add the evidence statements to the evidence set.

[0053] Most traditional methods rely on one-to-one claim-evidence relevance. This single-hop retrieval strategy is difficult to fully integrate the semantic interactions of multiple pieces of evidence, resulting in poor performance when dealing with complex claims. In addition, with the increase in the scale of the corpus and the complexity of information, the retrieval accuracy and inference depth of existing methods are also limited to a certain extent. The multi-hop method provides a new solution for the verification of complex claims. Multi-hop retrieval selects evidence step by step and iteratively, not only considering the claim itself, but also comprehensively using the retrieval results of the previous hops, thereby constructing a more comprehensive evidence chain. This method can capture the deep semantic relationships across multiple pieces of evidence and effectively support the verification of complex claims. However, the multi-hop method faces two major challenges in practical applications: one is that the accumulated context may lead to intention drift, making the retrieval results gradually deviate from the core semantics of the claim; the other is that the complex interaction relationships between multi-hop evidence significantly increase the difficulty of the subsequent inference process. Therefore, designing an effective multi-hop retrieval and inference mechanism is of great significance for improving the overall performance of the fact-checking system.

[0054] In the specific implementation process, this solution constructs a multi-hop retrieval scheme through multiple retrieval rounds, which can gradually retrieve the evidence related to the claim, construct a multi-hop retrieval chain, and effectively alleviate the problem of intention deviation during the retrieval process.

[0055] Step S200, combine the claim statement with each evidence statement in the evidence set to form a claim-evidence pair, and input each claim-evidence pair into a recursive memory model for intra-sentence inference to obtain the corresponding claim-evidence pair vector;

[0056] Step S300: Based on the statement evidence, construct a multi-hop inference subgraph for each evidence statement in the vector, and obtain an inter-sentence inference vector based on all the multi-hop inference subgraphs.

[0057] Step S400: Input the inter-sentence inference vector into a pre-set MLP model to obtain the verification result of the statement sentence.

[0058] In the specific implementation process, the verification result of the statement sentence includes support, negation, or insufficient information.

[0059] The present invention aims at the fact verification task, aiming to verify the authenticity of a statement sentence through a multi-hop retrieval and inference model. The goal of fact verification is to determine whether a statement sentence can be supported, refuted, or insufficient in information ("Supports", "Refutes", or "Not Enough Information") according to a statement sentence c and an evidence set E = {e1, e2,..., e n}.

[0060] In the specific implementation process, this solution mainly includes two core modules: a generative multi-hop retriever and a hierarchical interactive reasoner. The generative multi-hop retrieval module of the overall framework of the present invention is responsible for gradually generating an evidence chain related to the statement, dynamically adjusting the query content during the iterative retrieval process to ensure that the query for each hop maintains semantic consistency with the original statement, thereby avoiding the problem of information drift. The hierarchical interactive reasoning module performs reasoning analysis on the retrieved multi-hop evidence, combines the intra-sentence and inter-sentence semantic relationships, and deeply explores the complex interactions between the evidences, so as to accurately judge the authenticity of the statement.

[0061] Adopting the above solution, this solution focuses on solving the problems of intention deviation in multi-hop retrieval and complex association modeling in multi-hop reasoning. These problems are the core challenges that lead to low retrieval efficiency and limited reasoning ability of existing models when dealing with complex statements. Specifically, this solution designs a hierarchical structure for intra-sentence reasoning and inter-sentence reasoning. In intra-sentence reasoning, it uses recurrent memory to capture long-distance semantic dependencies. In inter-sentence reasoning, based on latent edge learning and graph neural networks, it constructs a multi-hop evidence subgraph to model complex cross-evidence semantic associations, capturing local semantic features and the global structural relationship between multiple evidences.

[0062] In some embodiments of the present invention, in the step of combining the statement sentence and the evidence sentences in the current evidence set and retrieving the retrieved evidence sentences from the database through a constrained generative retriever:

[0063] Combine the statement and the evidence statements in the current evidence set, and input the combined statement into a preset query compressor to obtain a compressed combined statement;

[0064] Input the compressed combined statement into a constrained generative retriever to obtain the retrieved evidence statements.

[0065] In a specific implementation process, the query compressor uses the BART model, and the constrained generative retriever uses the seq2seq model.

[0066] In a specific implementation process, given an original statement c, the generative multi-hop retriever aims to iteratively select a set of evidence E = {e1, e2,..., e n} through a generative model. Specifically, in the t-th hop retrieval process, the query compressor generates a compressed query Q 1:t-1 by concatenating the statement c with the previously retrieved evidence statements E t-1 = {e1, e2,..., e t} to maintain semantic consistency with the original statement. The constrained generative retriever then retrieves the next-hop evidence e t from the corpus based on Q t using the seq2seq model. This iterative process uses historical retrieved evidence to optimize subsequent retrievals, thereby comprehensively generating a complete evidence chain to support the robust verification of the statement's authenticity.

[0067] In some embodiments of the present invention, each evidence statement e i represents a single piece of evidence retrieved from the document set D of the database. The fact verification process not only requires accurately understanding the semantic relationship between the statement and the evidence but also efficiently retrieving relevant evidence from a large number of documents. For this task, the present invention gradually constructs a complete evidence chain E through a generative multi-hop retrieval module and combines it with a hierarchical interaction reasoning module to uniformly model the intra-sentence and inter-sentence relationships of the evidence chain, and finally generates a verification result for the statement. The ultimate goal of the present invention is to ensure that the system can efficiently and accurately retrieve and infer the multi-hop evidence relationship related to the statement in a complex fact verification task by optimizing the retrieval and reasoning mechanisms, thereby judging the authenticity label y of the statement.

[0068] Specifically, during the retrieval process, directly concatenating the statement with multi-hop evidence may result in an overly long input query, thereby weakening the semantic integrity of the statement and reducing the accuracy of evidence retrieval. To solve this problem, the present invention introduces a query compressor to compress the query text while retaining the core semantic relationship between the statement and the evidence.

[0069] In the t-th hop retrieval, as Figure 2As shown, the query compressor utilizes the original claim c and the evidence E retrieved from the previous t - 1 hops 1:t-1 ={e1, e2,..., e t-1} as input to generate the compressed query Q t for the t - th hop retrieval. Specifically, given the input query I t ={c; e1; e2;...; e t-1}(where ";" represents the concatenation operation), the encoder first processes I t and transforms it into the context representation h t =Encoder C (I t ). In each decoding step j, the decoder generates the j - th word q t based on h t,<j and the previously generated sequence q t,j :

[0070] P(q t,j |q t,<j , h t ) = Decoder C (q t,<j , h t ), j ≤ L

[0071] where L is the predefined maximum compression length.

[0072] During the retrieval process, to enhance the interaction between evidences, the present invention adopts an autoregressive generation method. In each step of retrieval, the generative retriever takes the original query and historical evidences as input and uses the seq2seq model to generate the next - hop evidence.

[0073] First, the query compressor compresses the input sequence I t into the query Q t . During the generation process, the compressed query Q t is encoded into the context representation Subsequently, the t - th hop evidence e t is generated step by step. Specifically, in each decoding step j, the decoder generates the current j - th word e t,<j based on the previously generated sequence e and t,j :

[0074]

[0075] The generation process of the t - th hop evidence can be expressed as:

[0076]

[0077] where L represents the length of e t .

[0078] To prevent the problem of generating invalid candidate evidence, the present invention uses a constrained generation strategy to limit the range of words generated by the model in each decoding step. A prefix tree is used to store candidate evidence, where each node represents a word in the evidence and the path between nodes can represent a string. The prefix tree constraint significantly reduces the number of candidate words, thus improving the generation efficiency.

[0079] In some embodiments of the present invention, in the step of combining the claim statement with each evidence statement in the evidence set into a claim-evidence pair and inputting each said claim-evidence pair into a recursive memory model for intra-sentence reasoning to obtain a corresponding claim-evidence pair vector, each said claim-evidence pair is input into a recursive memory model including multiple Transformer layers to obtain a corresponding claim-evidence pair vector.

[0080] In the specific implementation process, first, intra-sentence reasoning is performed on the text sequence of each claim-evidence pair to initialize the context features. As the number of retrieval hops increases, the distance between the retrieved evidence and the original claim gradually expands. Traditional Transformer often blurs the global information due to the dispersion of context when processing long sequences. To solve this problem, the present invention proposes to use a recursive memory Transformer, which supports the transmission and retention of global information by introducing memory tokens, thereby enhancing the ability to capture global context in long sequences.

[0081] As Figure 4 shown, assume X is the text sequence of a claim-evidence pair, which is divided into S segments, i.e., X = {X1, X2,..., X S}. To effectively model long-term dependencies, special memory tokens are added to enhance each segment X τ . Specifically, the memory token is appended to the beginning and end of each segment, i.e., for each segment X τ , its initial input representation is: where "°" represents the concatenation operation. Then, the enhanced sequence is processed using a standard Transformer, i.e., where N1 represents the number of Transformer layers.

[0082] To enable the sequence representation to pay attention to the memory state generated by X τ , the end token segment of is used as X τ to read the memory unit To introduce the memory information of the previous segment into the subsequent segment representation, The starting token segment is defined as writing to the memory cell and taking it as X τ+1 's memory token That is Then X τ+1 's initial input representation is:

[0083] Finally, the comprehensive representation H of the sentence is obtained by combining the read memory cells of each segment This recursive memory process effectively models long-term dependencies and ensures the accurate transmission of information.

[0084] In some embodiments of the present invention, based on the order of addition of evidence statements in the evidence set, the evidence statements are sequentially numbered. In the step of constructing a multi-hop inference subgraph corresponding to each evidence statement based on the declarative evidence pair vectors:

[0085] In the multi-hop inference subgraph corresponding to an evidence statement, any evidence statement corresponding to the previous number of the number of this evidence statement is included in the corresponding node. In the multi-hop inference subgraph, multiple path edges are sequentially connected based on the numbers of the evidence statements for each node;

[0086] Based on the similarity of the declarative evidence pair vectors corresponding to each node, potential connection edges between the nodes are constructed to obtain the final multi-hop inference subgraph.

[0087] In some embodiments of the present invention, in the step of constructing potential connection edges between nodes based on the similarity of the declarative evidence pair vectors corresponding to each node to obtain the final multi-hop inference subgraph, the similarity is calculated based on the declarative evidence pair vectors corresponding to the nodes and compared with a preset similarity threshold to determine whether to construct potential connection edges between two nodes.

[0088] In some embodiments of the present invention, in the step of obtaining the inter-sentence inference vector based on all the multi-hop inference subgraphs, each multi-hop inference subgraph is processed through a graph attention layer, and all the multi-hop inference subgraphs are aggregated to obtain the inter-sentence inference vector.

[0089] In the specific implementation process, in the step of aggregating all the multi-hop inference subgraphs to obtain the inter-sentence inference vector, a normalization layer is used for aggregation.

[0090] As Figure 5 shown, in order to capture the semantic connection between the claim and the evidence, the present invention constructs an evidence graph according to the retrieval path, and denotes the t-th hop evidence graph as G t. To alleviate the limitation of the sparse graph structure on node feature learning, a latent edge learning technique is used to infer potential connection relationships. According to the network homogeneity principle, similar nodes are more likely to establish connections with each other. Therefore, potential edges are inferred by calculating the similarity of node features.

[0091] Suppose is the initial representation of G t , where represents the i-th node in G t . To integrate semantic and structural features in latent edge learning, first, a graph convolutional neural network (GCN) is used to extract structural features , and then it is combined with the initial semantic representation to calculate the cosine similarity between node i and node j for constructing the latent edge matrix:

[0092]

[0093] where γ is the similarity threshold, indicates that there is an edge between two nodes, indicates that there is no edge.

[0094] To handle the dynamics of node relationships under different connections, the present invention introduces a trainable weight matrix specific to the hop number t . If there is an edge between node i and node j, that is, , then the edge weight matrix E is updated through the following formula t :

[0095]

[0096] . In addition, a shorter graph distance indicates a stronger correlation between nodes. Therefore, an attention bias based on the relative graph position is introduced during the aggregation process. This bias is integrated into the attention mechanism based on the softmax function, thus enhancing the attention between neighboring evidence nodes. The output representation of the evidence graph G t is denoted as S t and is defined as:

[0097]

[0098] where d ij represents the relative graph distance between evidence nodes i and j, and m G represents the fixed slope of a specific attention head. Under this mechanism, to stabilize the learning process, it can be extended to a multi-head format, and each attention head is calculated independently.

[0099] After integrating the outputs of each head, the final graph representation It will be obtained through a normalization layer:

[0100]

[0101] Subsequently, the final fused evidence representation F is obtained by concatenating multi-hop evidence representations t , and F t is input into a multi-layer perceptron (MLP) to obtain the authenticity prediction label of the original claim.

[0102] In some embodiments of the present invention, the steps of the method further include pre-training the query compressor, and the steps of pre-training the query compressor include distillation and alignment processing;

[0103] During the distillation process, a pre-set large language model is used to process the combined statement of the claim statement and the evidence statement to generate a compressed text, which is used as the learning target of the query compressor;

[0104] Use a pre-set large language model to calculate the semantic matching degree between the compressed text and the original claim statement, and take the compressed text with the highest score as the positive sample, and the remaining compressed texts as negative samples;

[0105] Train the query compressor based on the positive samples and negative samples.

[0106] In some embodiments of the present invention, to alleviate intention drift and construct a more compact fact-checking framework, the present invention uses the powerful inductive ability of a large language model (LLM) to perform enhanced pre-training on the query compressor. As Figure 3 shown, the pre-training process is achieved through two steps: distillation and alignment.

[0107] Distillation: Use the compression ability of the large model to enhance the query compressor. Given a claim c and an evidence set E, prompt the LLM to process the concatenated text of the claim and the evidence to generate a compressed text Q ′ , as the learning target of the query compressor.

[0108] Alignment: Enhance the query compressor through Figure 1 consistency. Given c and E, use the distilled query compressor to generate N compressed query texts, and use the LLM to evaluate the semantic matching degree between these N compressed queries and the original claim c, and take the compressed query with the highest score as the positive sample Q + , and the n compressed queries with the lowest scores as negative samples After that, apply a contrastive learning-based loss L a to align the compressed queries to Q + .

[0109] In some embodiments of the present invention, in the step of training the query compressor based on the positive samples and negative samples, the combined statements corresponding to the positive samples and negative samples are input into the query compressor to obtain the compressed combined statements, and the loss function is calculated based on the compressed combined statements and the positive samples and negative samples, and the query compressor is trained based on the loss function value.

[0110] In some embodiments of the present invention, in the step of calculating the loss function based on the compressed combined statements and the positive samples and negative samples, the loss function is calculated based on the following formula:

[0111]

[0112] where L a represents the value of the loss function, N represents the number of combined statements, n represents the number of negative samples, Q i represents the i-th compressed combined statement, Q + represents the positive sample, Q j - represents the j-th negative sample, f(·) represents the encoder, and τ represents the temperature coefficient.

[0113] In some embodiments of the present invention, the steps of the present invention include the joint training of the overall model. The joint process of the present invention aims to improve the model's verification ability for complex statements by optimizing the processes of evidence retrieval and claim verification. Its training mechanism integrates the retrieval and reasoning modules, and through the joint optimization of multi-hop retrieval and reasoning, ensures the consistency of the evidence chain and the prediction result.

[0114] In each hop of the retrieval process, the generative multi-hop retrieval module recursively retrieves the t-th piece of evidence e t and constructs the updated evidence set E t . The cross-entropy loss L r is used in the retrieval stage to compare the retrieved evidence E t with the true evidence set E gold to ensure the accuracy of the evidence chain. In the claim verification stage, the hierarchical interactive reasoning module obtains the predicted label y t of the original claim based on the current evidence set E * . The claim verification loss L v is calculated through the cross-entropy of the predicted label y * and the true label y to ensure the accuracy of the final verification result.

[0115] Finally, the joint training mechanism combines the retrieval loss L r and the verification loss L v , and the overall loss is L = L r + L v. Through backpropagation through gradients, the parameters of the retrieval module and the reasoning module are updated simultaneously, enabling the retrieval module and the verification module to closely collaborate in the two subtasks, thereby improving the overall performance and consistency of the model. The joint optimization strategy ensures the deep integration of evidence retrieval and reasoning, providing an effective solution for the verification of complex claims.

[0116] For the generative multi-hop retrieval module, the present invention allows the use of any evidence retrieval method that can meet the following functional requirements: being able to gradually retrieve evidence related to the claim, construct a multi-hop retrieval chain, and effectively alleviate the problem of intention drift during the retrieval process. For example, dense retrieval models (such as extended versions of Dense Retriever or BM25) can, through an efficient vectorized matching mechanism, replace the generative multi-hop retriever and achieve the retrieval of multi-hop evidence to a certain extent. However, when dealing with query semantic drift in multi-hop retrieval, such methods often perform less stably than generative methods. In addition, generative models such as T5 or GPT can also be used as alternatives, by generating compact next-hop queries to gradually retrieve relevant evidence and solve the query construction problem in multi-hop tasks. It should be noted, however, that although large-scale generative models have excellent performance, their high computational resources and cost may become a bottleneck in practical applications.

[0117] For the hierarchical interaction reasoning module, the present invention allows the use of any reasoning method that can meet the following functional requirements: being able to capture long-range semantic dependencies between multi-hop evidence, model local and global semantic interactions, and support multi-level reasoning of complex logical relationships. Pure graph neural network (GNN) reasoning can, by constructing a global evidence graph and directly modeling semantic relationships between nodes and edges, achieve interactive reasoning between multi-hop evidence to a certain extent. Global Transformer models can also directly model the global dependencies between the claim and all evidence through the self-attention mechanism, thereby simplifying the reasoning process. Reasoning methods based on memory networks can achieve the logical consistency of the multi-step reasoning process by dynamically storing and accessing the evidence information of multi-hop retrieval. These alternative solutions can meet the core functional requirements of multi-hop reasoning tasks in different scenarios, but may differ in terms of flexibility, performance, or computational cost.

[0118] The present invention proposes a novel method for fact-checking complex claims based on multi-hop retrieval and reasoning, effectively solving the problem of intention drift in multi-hop retrieval and promoting the refined reasoning of multi-hop complex association clues.

[0119] First, the present invention analyzes the deficiencies of existing multi-hop evidence retrieval methods, pointing out that these methods usually rely on simple text splicing strategies. As the number of hops increases, the query length continuously expands, resulting in the dilution of core semantic information, and at the same time, the accumulated context is prone to deviate from the original intention of the statement. Therefore, the present invention adopts a generative multi-hop retrieval strategy based on pre-trained query compression. The query compressor performs semantic compression on the statement and historical retrieval evidence to generate a compact and semantically consistent query. This retrieval strategy can alleviate the problem of intention deviation, and at the same time, combined with the prefix tree constraint generation mechanism, significantly improves the efficiency and accuracy of retrieval. Through iterative retrieval, the present invention further constructs a coherent evidence chain, providing reliable support for complex statement verification.

[0120] Secondly, the present invention recognizes that when existing multi-hop reasoning methods handle multi-hop evidence, it is difficult to simultaneously capture the long-distance semantic dependencies within a sentence and the global associations between multiple pieces of evidence, resulting in the inability to effectively integrate the logical relationships across evidence during the reasoning process. Based on this, the present invention designs a hierarchical interactive reasoning mechanism in the reasoning stage, modeling the associations between evidence through two levels: intra-sentence reasoning and inter-sentence reasoning. This reasoning process can effectively model local and global semantics, capture potential association clues across evidence, thereby realizing refined and multi-level reasoning for complex statements and improving the accuracy and robustness of the fact-checking task.

[0121] Finally, the present invention adopts a joint training mechanism to optimize the multi-hop retrieval and reasoning processes in an end-to-end manner, ensuring that the retrieval results can better serve the reasoning task, while the reasoning feedback can in turn guide the retrieval process in the reverse direction, forming an effective closed-loop optimization. By simultaneously optimizing the retrieval loss and the verification loss, the collaborative work of retrieval and reasoning is realized, gradually constructing a high-quality evidence chain and improving the reasoning accuracy.

[0122] In summary, the present invention proposes an innovative solution. The model consists of a generative multi-hop retriever and a hierarchical interactive reasoner, aiming to achieve a deep integration of evidence retrieval and reasoning. This solution needs to address the following two technical problems: First, during the multi-hop retrieval process, it is easy to have an intention deviation problem, resulting in the retrieval results gradually deviating from the core semantics of the claim. Second, it is difficult to capture long-distance semantic dependencies and complex interactions between multi-hop evidences simultaneously during the reasoning stage, which limits the comprehensive understanding of complex claims. For the first problem, the present invention adopts an autoregressive generation method and introduces a query compressor. By compressing the claim and historical retrieval evidences into a concise query with consistent semantics, the intention deviation is effectively alleviated. Combining with a prefix tree to limit the generation scope significantly improves the retrieval efficiency and accuracy. For the second problem, a recursive memory Transformer is used in the reasoning module for intra-sentence reasoning, so as to capture long-term semantic dependencies. Then, by performing inter-sentence reasoning on the multi-hop evidence subgraph, the interaction between evidence segments is clarified, thereby enhancing the modeling ability of complex associations between multi-hop evidences and achieving a significant improvement in global semantic understanding.

[0123] The beneficial effects of this solution include:

[0124] 1. By introducing a pre-training enhanced query compressor, the combination of the claim and historical evidences is effectively compressed to generate a compact and semantically consistent query, thus alleviating the problems of excessive query length and semantic deviation during the multi-hop retrieval process. At the same time, combining with a constrained generation mechanism (such as a prefix tree) ensures the accuracy and efficiency of the retrieval results.

[0125] 2. Design a hierarchical structure for intra-sentence reasoning and inter-sentence reasoning. In intra-sentence reasoning, a cyclic memory Transformer is used to capture long-distance semantic dependencies. In inter-sentence reasoning, based on latent edge learning and graph neural networks, a multi-hop evidence subgraph is constructed to model complex cross-evidence semantic associations.

[0126] 3. Through joint optimization of the retrieval loss and the verification loss, end-to-end integration of the evidence retrieval and reasoning processes is achieved, enabling the retrieval results to dynamically adapt to the reasoning requirements, and the reasoning feedback further guides the optimization of the retrieval module.

[0127] An embodiment of the present invention also provides a complex claim fact-checking system based on multi-hop retrieval and reasoning. The system includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps achieved by the method described above.

[0128] An embodiment of the present invention also provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps implemented by the foregoing complex claim fact-checking method based on multi-hop retrieval and reasoning are realized. The computer-readable storage medium may be a tangible storage medium, such as a random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium well-known in the technical field.

[0129] Those of ordinary skill in the art should understand that the various exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Specifically, whether to implement in hardware or software depends on the specific application and design constraints of the technical solution. A professional technician can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application-specific integrated circuit (ASIC), appropriate firmware, a plug-in, a functional card, and so on. When implemented in software, the elements of the present invention are programs or code segments used to execute the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted through a data signal carried in a carrier wave on a transmission medium or a communication link.

[0130] It should be clear that the present invention is not limited to the specific configurations and processes described above and illustrated in the figures. For the sake of brevity, the detailed description of known methods is omitted here. In the above embodiments, several specific steps are described and illustrated as examples. However, the method process of the present invention is not limited to the specific steps described and illustrated. Those skilled in the art can make various changes, modifications, and additions, or change the order between steps after understanding the spirit of the present invention.

[0131] In the present invention, the features described and / or illustrated for one embodiment can be used in the same way or in a similar way in one or more other embodiments, and / or combined with the features of other embodiments or replace the features of other embodiments.

[0132] The above are only the preferred embodiments of the present invention and are not used to limit the present invention. For those skilled in the art, various changes and modifications can be made to the embodiments of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for fact-checking complex claims based on multi-hop retrieval and reasoning, characterized in that: The steps of the method include: Obtain the statement to be verified, and construct a set of evidence corresponding to the statement through multiple retrieval rounds; In each search round, the claim statement is combined with the evidence statement in the current evidence set, and is searched from the database by a constraint generation searcher to obtain the retrieved evidence statement, and the evidence statement is added to the evidence set; Combining the claim statement with each evidence statement in the evidence set into a claim-evidence pair, and inputting each of the claim-evidence pairs into a recursive memory model for intra-sentence reasoning to obtain a corresponding claim-evidence pair vector; Constructing a multi-hop reasoning subgraph corresponding to each evidence sentence based on the statement-evidence pair vector, and obtaining an inter-sentence reasoning vector based on all the multi-hop reasoning subgraphs; The inter-sentence reasoning vector is input into a preset MLP model to obtain a verification result of the declaration sentence.

2. The method for checking facts of complex claims based on multi-hop retrieval and reasoning according to claim 1, characterized in that: In the step of combining the declaration statement with the evidence statement in the current evidence set and retrieving the evidence statement from the database by means of a constraint generation searcher: Combining the declaration statement with the evidence statement in the current evidence set, and inputting the combined statement into a preset query compressor to obtain a compressed combined statement; The compressed combined sentence is input into the constraint generation retriever to obtain the retrieved evidence sentence.

3. The method for fact-checking complex claims based on multi-hop retrieval and reasoning according to claim 1, characterized in that: In the step of combining the declaration statement with each evidence statement in the evidence set into a declaration-evidence pair, inputting each of the declaration-evidence pairs into a recursive memory model for intra-sentence reasoning, and obtaining a corresponding declaration-evidence pair vector, each of the declaration-evidence pairs is input into a recursive memory model including multiple transformer layers to obtain a corresponding declaration-evidence pair vector.

4. The method for fact-checking complex claims based on multi-hop retrieval and reasoning according to claim 1, characterized in that: In the evidence set, the evidence statements are sequentially numbered based on the order in which the evidence statements are added, and in the step of constructing a multi-hop reasoning subgraph corresponding to each evidence statement based on the declaration evidence pair vector: In the multi-hop reasoning subgraph corresponding to the evidence statement, a corresponding node of any evidence statement with a pre-order number of the evidence statement is included, and multiple path edges are constructed by sequentially connecting each node based on the number of the evidence statement in the multi-hop reasoning subgraph; Based on the similarity of the statement-evidence pair vectors corresponding to each node, potential connections between nodes are constructed to obtain the final multi-hop reasoning subgraph.

5. The method for checking facts of complex claims based on multi-hop retrieval and reasoning according to claim 4, characterized in that: In the step of constructing potential connections between nodes based on the similarity of the claim-evidence pair vectors corresponding to each node to obtain the final multi-hop reasoning subgraph, the similarity is calculated based on the claim-evidence pair vectors corresponding to the nodes and compared with a preset similarity threshold to determine whether to construct a potential connection between the two nodes.

6. The method for fact-checking complex claims based on multi-hop retrieval and reasoning according to claim 1, characterized in that: In the step of obtaining the inter-sentence reasoning vector based on all multi-hop reasoning subgraphs, Each multi-hop reasoning subgraph is processed through the graph attention layer, and all multi-hop reasoning subgraphs are aggregated to obtain the inter-sentence reasoning vector.

7. The method for checking facts of complex claims based on multi-hop retrieval and reasoning according to claim 2, characterized in that: The method further comprises pre-training the query compressor, wherein the step of pre-training the query compressor comprises distillation and alignment processing; During the distillation process, a pre-set large language model is used to process the combination of declaration sentences and evidence sentences to generate compressed text as the learning target of the query compressor; Use the pre-set large language model to calculate the semantic matching degree between the compressed text and the original declaration sentence, take the compressed text with the highest score as the positive sample, and take the rest of the compressed text as the negative sample; The query compressor is trained based on the positive samples and the negative samples.

8. The method for checking facts of complex claims based on multi-hop retrieval and reasoning according to claim 7, characterized in that: In the step of training the query compressor based on the positive samples and negative samples, the combined sentences corresponding to the positive samples and negative samples are input into the query compressor to obtain compressed combined sentences, a loss function is calculated based on the compressed combined sentences and the positive samples and negative samples, and the query compressor is trained based on the loss function value.

9. The method for checking facts of complex claims based on multi-hop retrieval and reasoning according to claim 8, characterized in that: In the step of calculating the loss function based on the compressed combined sentence and the positive sample and the negative sample, the loss function is calculated based on the following formula: Among them, L a represents the value of the loss function, N represents the number of combined sentences, n represents the number of negative samples, Q i represents the i-th compressed combined statement, Q + represents a positive sample, represents the jth negative sample, f(·) represents the encoder, and τ represents the temperature coefficient.

10. A complex claim fact-checking system based on multi-hop retrieval and reasoning, characterized in that: The system includes a computer device, which includes a processor and a memory. The memory stores computer instructions. The processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the system implements the steps implemented by the method as described in any one of claims 1 to 9.

Images