CN-121809461-B - Fine evidence tracing and declaration verification method based on heterogeneous target fusion

CN121809461BCN 121809461 BCN121809461 BCN 121809461BCN-121809461-B

Abstract

The invention relates to a natural language processing technology, in particular to a subtle real evidence tracing and statement verification method based on heterogeneous target fusion, which comprises the following steps of searching candidate documents related to a statement to be verified and extracting candidate sentences related to the statement to be verified; the method comprises the steps of constructing a concatenated statement-sentence association probability distribution model and a statement-sentence set association probability distribution model, learning a prediction probability distribution between a statement and a single candidate sentence based on the statement-sentence association probability distribution model, screening a plurality of candidate sentences with higher demonstration value to obtain a candidate demonstration set, capturing a semantic synergistic effect between the candidate sentences in the candidate demonstration set based on the statement-sentence set association probability distribution model, calculating a global category probability distribution of the statement to be verified and the candidate demonstration set, adopting heterogeneous targets to jointly optimize parameters of the two models, and realizing fine demonstration tracing and demonstration verification based on the two trained models. The invention can provide visual and accurate viewpoint verification results.

Inventors

ZENG JUAN
WANG YAO
LIU YANAN
Deng Haotian
Chen Polong
LUO WEILIN
WAN HAI

Assignees

中山大学

Dates

Publication Date: 20260508
Application Date: 20260312

Claims (10)

1. A fine evidence tracing and declaration verification method based on heterogeneous target fusion is characterized by comprising the following steps: S1, searching candidate documents related to a statement to be verified from a corpus by document searching by adopting a searching strategy based on semantic analysis; s2, extracting candidate sentences related to the statement to be verified from the candidate documents, and constructing a candidate sentence subset; S3, constructing a correlation probability distribution model, wherein the constructed correlation probability distribution model comprises a feature extraction module and a target classification module which are connected, and the feature extraction module is used for extracting Gao Weiyin layers of representations of input sequences, the target classification module comprises a full-connection layer and a normalized index function which are connected, and the target classification module learns and extracts semantic features in Gao Weiyin layers of representations according to the extracted Gao Weiyin layers of representations, and outputs correlation probability distribution of declarations and sentences or outputs correlation probability distribution of declarations and sentence sets; Setting a cascading target classification architecture by using two associated probability distribution models, and setting a first associated probability distribution model as a statement-sentence associated probability distribution model; S4, learning prediction probability distribution between the declaration and the single candidate sentences based on the declaration-sentence association probability distribution model, and quantifying local confidence coefficient of the single candidate sentences serving as support basis of the declaration to be verified, wherein the local confidence coefficient is used for measuring the demonstration value of the candidate sentences; S5, screening a plurality of candidate sentences with higher evidence value from the candidate sentence set based on the local confidence coefficient, and arranging and combining the plurality of candidate sentences into sentence combinations according to the descending order of the evidence value to be used as the candidate evidence set; s6, adopting heterogeneous target joint optimization, synchronously training and updating parameters of two associated probability distribution models; in the process of training the associated probability distribution model, the loss calculated by the prediction probability distribution and the loss calculated by the global category probability distribution are subjected to weighted fusion to construct a joint loss function; s7, based on the trained statement-sentence association probability distribution model and the statement-sentence set association probability distribution model, fine evidence tracing and sound verification are achieved through a top-down or bottom-up extraction strategy.
2. The fine verification tracing and declaration verification method according to claim 1, wherein step S2 further constructs a sequence to be verified of a declaration-candidate sentence subset, wherein each element in the sequence to be verified includes a declaration, a candidate sentence subset for verifying authenticity of the declaration; the step S4 includes: S41, constructing a declaration-single sentence sequence according to a sequence to be verified of a declaration-candidate sentence subset, wherein the declaration-single sentence sequence is used as an input sequence of a declaration-sentence association probability distribution model and comprises a starting mark, a declaration, a separation mark and a single candidate sentence, the starting mark is used for aggregating semantic features of the declaration-single sentence sequence, and the separation mark is used for definitely defining boundaries of the declaration and the single candidate sentence; S42, inputting the declaration-single sentence sequence into a feature extraction module of the declaration-sentence associated probability distribution model, and outputting a feature vector corresponding to the initial mark after the feature extraction module encodes the declaration-single sentence sequence Feature vector The method comprises the steps of representing complete semantic relation and deep logic interaction between a statement to be verified and a single candidate sentence; s43, feature vector And calculating the prediction probability distribution of the corresponding single candidate sentence relative to the statement to be verified through the full connection layer and the normalized exponential function in the statement-sentence association probability distribution model.
3. The fine verification method according to claim 2, wherein in step S41, for the candidate sentence subset constructed in step S2 Each candidate sentence in the list is independently analyzed, and a statement to be verified is declared From the following components The composition of individual words, noted as A candidate sentence From the following components The composition of individual words, noted as Wherein The format of the "claim-clause" sequence is defined as follows: ; Wherein [ CLS ] represents a start tag and [ SEP ] represents a separation tag.
4. The fine demonstration trace-back and statement verification method of claim 2, wherein the predictive probability distribution calculated in step S43 is: ; Wherein the method comprises the steps of And Is a learnable parameter; representing a normalized exponential function, wherein the normalized exponential function is used for nonlinearly mapping real values in any range output after full-connection layer linear transformation into probability distribution vectors with the sum of elements being 1 so as to represent the local confidence that corresponding single candidate sentences belong to each classification dimension; Predicting probability distribution The method comprises the following steps of supporting probability, refuting probability and irrelevant probability, and is used for accurately describing three-dimensional demonstration values relative to a statement to be verified: the support probability indicates that the candidate sentences can confirm the confidence of the statement to be verified; the refuting probability represents the confidence that the candidate sentences can be verified; irrelevant probability represents the confidence that the candidate sentence does not contain valid verification information of the statement to be verified.
5. The fine demonstration reason and statement verification method of claim 1, wherein step S5 comprises: s51, arranging all candidate sentences in the candidate sentence set in a descending order according to the demonstration value, removing candidate sentences irrelevant to statement, selecting a plurality of candidate sentences with higher confidence level for descending order arrangement and combination to obtain sentence combination, and forming a candidate demonstration set; S52, constructing a sequence of a statement-sentence subset according to the candidate evidence set as an input sequence of a statement-sentence set association probability distribution model, wherein the sequence of the statement-sentence set comprises a start mark, a statement, a separation mark and sentence combinations of the candidate evidence set, the start mark is used for aggregating global semantic features of the sequence of the statement-sentence subset and providing a unified semantic representation basis for the association probability distribution model, and the separation mark is used for definitely defining boundaries of the statement and the sentence combinations; s53, inputting the sequence of the statement-sentence subset into a feature extraction module in the statement-sentence set association probability distribution model, and outputting a feature vector corresponding to a start mark after the feature extraction module encodes the sequence of the statement-sentence subset Feature vector The method comprises the steps of representing complete semantic relation and deep logic interaction between a statement to be verified and a candidate evidence set; S54, feature vector And calculating global probability distribution of sentence combinations in the candidate real evidence set relative to the statement to be verified through a full connection layer and a normalized exponential function in the statement-sentence set association probability distribution model.
6. The fine-verification-tracing and declaration verification method according to claim 5, wherein the "declaration-sentence subset" sequence constructed in step S52 is: Wherein For a sentence combination in the candidate evidence set, In order to make the start-up mark, In order to be stated, Is a separation mark.
7. The fine demonstration tracing and declaration verification method according to claim 5, wherein the global probability distribution calculated in step S54 is: ; Wherein the method comprises the steps of And Is a trainable parameter; the global probability distribution is used for comprehensively describing the overall supporting effect of the candidate real evidence set to be verified, and specifically comprises the following steps: Sentence subset support probability The method comprises the steps of judging whether a current sentence combination forms a complete supporting evidence chain or not; Sentence subset refuting probability The method comprises the steps of judging whether a current sentence combination forms a complete refuting evidence chain or not; probability of information shortage For determining whether the current set of candidate facts still lacks critical information sufficient to infer authenticity.
8. The fine-verification-tracing-and-declaration verification method according to claim 5, wherein step S6 includes: S61, setting an original training data set to comprise a plurality of sample facts, wherein each sample fact comprises a statement to be verified c and a statement-sentence set classification label And a minimum set of sentences for validating or witnessing the statement c to be validated, wherein A value of 0 represents that the minimum sentence set supports a claim to be validated, A value of 1 represents that the minimum sentence set refutes the statement to be verified, Marking the candidate sentences in the candidate sentence set obtained in the step S2 according to the original training data set to obtain true classification labels of declarations and single sentences Wherein A value of 0 represents the single sentence support claim, A value of 1 represents the single sentence disagreement declaration, A value of 2 represents that the single sentence is independent of the declaration; S62, calculating the verification value judgment loss of the single candidate sentence layer as sentence-statement pair association loss based on the prediction probability distribution obtained in the step S4 and the real classification label obtained in the step S61, and summing the sentence-statement pair association loss of all candidate sentences to obtain statement-single sentence association probability distribution total loss ; S63, generating training samples of a statement-sentence set association probability distribution model according to the candidate demonstration sets obtained in the step S51; Firstly initializing a sentence set S 'of a training sample as an empty set, then starting from the empty set, gradually adding candidate sentences in the candidate evidence sets, which are ordered in descending order of the evidence value, into the sentence set, judging whether the current sentence set contains the minimum sentence set in the sample evidence of the original training data set after each addition, and if so, obtaining a training sample, wherein the training sample comprises a statement c to be verified, the sentence set S' and a corresponding statement-sentence set classification label ; S64, calculating global association loss between the statement and the sentence set of the training sample obtained in the step S63 according to the global probability distribution The definition is as follows: ; Wherein, the Is a global probability distribution Is the first of (2) Individual elements, indicating functions The method comprises the steps of identifying the matching situation of a probability distribution prediction result and a statement-sentence set classification label of a training sample in the step S63; s65, associating statement-sentence with probability distribution total loss Loss of association with global And carrying out weighted fusion, constructing a heterogeneous target joint loss function, and synchronously updating parameters of the declaration-sentence association probability distribution model and the declaration-sentence set association probability distribution model.
9. The fine verification method according to claim 8, wherein the step S62 is provided with a third step of The predictive probability distribution of each candidate sentence is Will be at the first Sentence-declaration pair association loss between candidate sentences and declarations The definition is as follows: ; Wherein, the Is a predictive probability distribution Is the first of (2) The number of elements to be added to the composition, Represent the first The support probabilities of the individual candidate sentences are, Represent the first The probability of refuting the individual candidate sentences, Represent the first Irrelevant probabilities of the candidate sentences; to indicate a function if True, the indicator function takes a value of 1 if If false, the indication function takes a value of 0; the sum of all sentence-declaration pair association losses is taken as the total loss of declaration-single sentence association probability distribution: ; Wherein the method comprises the steps of And (3) collecting the total number of candidate sentences in the candidate sentence set obtained in the step S2.
10. The method for tracing and declaration verification according to claim 5, wherein step S7 sequentially executes steps S1, S2, S4 and S5 on declarations to be verified in the test set, and obtains a complete candidate evidence set through step S51; Firstly judging the verification result of the complete candidate real evidence set, then gradually adding sentences into the subset initialized to the empty set according to the real evidence value sequence until the judgment result generated by the subset is consistent with the verification result of the complete candidate real evidence set, and outputting the current minimum subset as a fine real evidence; The bottom-up extraction strategy is to sequentially add sentences of the complete candidate real evidence set to the subsets initialized as empty sets, and the sentence combination obtained by each addition passes through the steps S52-S54 to judge the relation between the statement and the sentence combination, and once the judgment result is confirmation or the certificate pseudo-statement, the search is immediately stopped and the current subset is output as the fine real evidence.

Description

Fine evidence tracing and declaration verification method based on heterogeneous target fusion Technical Field The invention relates to a natural language processing technology, in particular to a precise evidence tracing and declaration verification method based on heterogeneous target fusion. Background In the age of today's increasingly convenient information dissemination, networks are being enriched with vast amounts of unproven information and rumors. Fact verification (Fact Verification) has resulted, which determines the authenticity of sentences, mainly by analyzing a corpus of related sentences. The goal of fact verification is to verify the authenticity of a given sentence by extracting the relevant sentence from a corpus of text. The fact verification system needs to determine whether a sentence is "supported", "refuted" or "under information" and find out the corresponding text evidence. The fact verification has become an extremely critical issue in the current internet management, and is an effective means for coping with network false information. A common fact verification process is to first retrieve a document related to a sentence to be verified (i.e., a target sentence, TARGET SENTENCE), and then extract the most representative supporting sentence or refuted sentence from the retrieval result, so as to analyze and verify the authenticity of the target sentence. Based on such verification procedures, most of the current fact verification studies directly rely on related sentences as the evidence, but ignore contradiction or redundancy problems that may exist in the sentences. More specifically, conventional fact verification is generally divided into three steps, namely (1) searching related candidate documents according to a given sentence in a specific corpus (such as wikipedia), (2) screening out the associated sentence with the strongest association with the target sentence from the search result, and (3) judging the true or false of the target sentence by using the associated sentence, namely, judging the true or false of the target sentence. That is, the conventional method generally uses the related sentence subsets selected by the sentence selection module directly as the evidence, but these sentence subsets may include not only precise evidence, but also sentences that are not necessary for determining the authenticity of the sentence. These sentences, which are independent of the authenticity of the judgment sentence, can negatively affect the understanding and subsequent verification of the demonstration, especially where contradictory information may mislead the verification of the sentence. Therefore, it is necessary to mine fine evidence of support or refute claims to achieve better proof of fact and human understanding. Fine demonstration refers to the smallest collection of sentences of the smallest semantic units that can support or refute declarations. For example, in one example of a mainstream verification dataset, for a statement that "overlay Reed is a movie actor," any sentence in the demonstration may support the statement. The fine demonstration contains only sentences required for verifying the statement, and does not contain redundant or contradictory information. In summary, the existing fact verification method has the following disadvantages: 1. the verification accuracy is insufficient, the existing method is mostly focused on sentence verification, and related sentences extracted by a sentence selection module are directly used as a verification set, but contradiction and redundant information possibly existing in the sentences are not fully considered. 2. The limitation of the joint learning method is that, unlike the conventional three-stage method, sentence selection and sentence verification are simultaneously performed in a joint training learning manner to generate sentences with variable evidence. However, although this method can generate accurate proofs, these proofs may contain sentences that conflict with each other, and do not necessarily improve the overall performance of the proof of fact. In recent years, attempts have been made to identify fine facts using reinforcement learning based algorithms. Although reinforcement learning works well in sentence extraction and verification tasks, this method requires a lot of resources and training time is long. In addition, it has a requirement on the search length, i.e., limits the number of sentences that can acquire evidence from the environment. 3. Performance loss due to task separation-more recently, there has been an interpretable fact verification based on potential representations by refining the logic knowledge in the statements. The method innovatively disassembles the declaration to the phrase level for judgment, but the input evidence set is limited to a fixed number of sentence sets. Because the existing fact verification method usually adopts coarse-granularity document or sentence leve