A fine empirical traceability and statement verification method based on heterogeneous target fusion

Abstract

The present application relates to natural language processing technology, and is a kind of fine empirical traceability and declaration verification method based on heterogeneous target fusion, comprising the following steps: retrieving candidate documents related to the declaration to be verified, extracting candidate sentences related to the declaration to be verified;Cascade is built to the declaration-sentence association probability distribution model, declaration-sentence set association probability distribution model;Based on the declaration-sentence association probability distribution model, learn the prediction probability distribution between the declaration and the single candidate sentence;Filtering a number of candidate sentences with higher empirical value to obtain a candidate empirical set;Based on the declaration-sentence set association probability distribution model, capture the semantic synergy effect between the candidate sentences in the candidate empirical set, calculate the global class probability distribution of the declaration to be verified and the candidate empirical set;Two models are optimized by using heterogeneous target;Based on the two trained models, fine empirical traceability and declaration verification are realized.The present application can provide intuitive and accurate opinion verification results.

Description

Technical Field

[0001] This invention relates to natural language processing technology, specifically a refined empirical causal attribution and declaration verification method based on heterogeneous target fusion. Background Technology

[0002] In today's era of increasingly convenient information dissemination, the internet is rife with unverified information and rumors. Fact verification has emerged to address this, primarily determining the veracity of statements by analyzing relevant sentence corpora. The goal of fact verification is to verify the veracity of a given statement by extracting relevant sentences from a text corpus. A fact verification system needs to determine whether a statement is "supported," "refuted," or "insufficiently informed," and identify corresponding textual evidence. Fact verification has become a crucial issue in current internet governance and an effective means of combating online misinformation.

[0003] A common fact-verification process involves first retrieving documents related to the statement to be verified (i.e., the target sentence), then extracting the most representative supporting or rebuttal sentences from the search results to analyze and verify the truthfulness of the target sentence. Based on this process, most current fact-verification research relies directly on relevant sentences as empirical evidence, neglecting potential contradictions or redundancy in these sentences.

[0004] More specifically, traditional fact verification typically involves three steps: (1) searching for relevant candidate documents in a specific corpus (such as Wikipedia) based on a given statement; (2) selecting the most relevant sentences from the search results; and (3) using the relevant sentences to determine the truth or falsehood of the target statement, i.e., determining whether it is "true" or "false". In other words, traditional methods usually directly use the set of relevant sentences selected by the sentence selection module as empirical evidence. However, these sentence sets may contain not only precise empirical evidence but also sentences that are not necessary for determining the truth or falsehood of the statement. These sentences, which are irrelevant to determining the truth or falsehood of the statement, can negatively impact the understanding of the empirical evidence and subsequent verification, especially since contradictory information may mislead the verification of the statement. Therefore, it is necessary to mine fine-grained empirical evidence that supports or refutes the statement in order to achieve better fact verification and human understanding. Fine-grained empirical evidence refers to the smallest set of sentences that can support or refute the statement, which is the smallest semantic unit. For example, in one example of a mainstream verification dataset, any sentence in the empirical evidence can support the statement "Oliver Reed is a movie actor". Detailed empirical evidence contains only the sentences necessary to verify the statements, and does not include any extraneous or contradictory information.

[0005] In summary, existing fact-verification methods have the following shortcomings:

[0006] 1. Insufficient empirical accuracy: Most existing methods focus on sentence verification, directly using the relevant sentences extracted by the sentence selection module as the empirical set, without fully considering the contradictions and redundant information that may exist in these sentences.

[0007] 2. Limitations of Joint Learning Methods: Unlike traditional three-stage methods, joint training simultaneously performs sentence selection and verification to generate sentences with variable empirical evidence. However, while this method can generate accurate empirical evidence, these evidences may contain conflicting sentences, and may not necessarily improve the overall performance of fact verification. In recent years, some researchers have attempted to use reinforcement learning-based algorithms to identify fine-grained empirical evidence. Although reinforcement learning performs well in sentence extraction and verification tasks, this method requires significant resources and has a long training time. Furthermore, it is subject to search length limitations, i.e., it restricts the number of sentences from the environment that can be used to obtain empirical evidence.

[0008] 3. Performance loss due to task separation: Recently, some researchers have extracted logical knowledge from statements and performed interpretable fact verification based on latent representations. This method innovatively decomposes statements to the phrase level for judgment, but its input empirical set is limited to a fixed set of sentences.

[0009] Because existing fact-verification methods often employ coarse-grained document or sentence-level screening, directly judging the truth or falsehood of the extracted text, this may not only introduce a large amount of information that is irrelevant to the statement or conflicting, leading to a waste of computational resources, but also make it difficult to provide accurate and interpretable empirical support, failing to ensure both the accuracy of verification and efficiency and traceability. Summary of the Invention

[0010] To address the challenge of quickly and accurately locating key empirical evidence and making explanatory judgments on claims when faced with massive amounts of text in existing fact-checking methods, this invention proposes a refined empirical causal analysis and claim verification method based on heterogeneous target fusion. By effectively extracting and combining refined empirical evidence from the text, it provides users with more intuitive and accurate results for verifying viewpoints. This method can not only quickly determine the truth or falsehood of claims but also provide minimal and necessary empirical evidence, making the decision-making process more transparent and efficient.

[0011] The technical solution adopted in this invention is: a refined empirical causal attribution and assertion verification method based on heterogeneous target fusion, comprising the following steps:

[0012] S1. A semantic parsing-based retrieval strategy is adopted to retrieve candidate documents related to the statement to be verified from the corpus through document retrieval;

[0013] S2. Extract candidate sentences related to the statement to be verified from the candidate documents and construct a candidate sentence set;

[0014] S3. Construct an association probability distribution model. The constructed association probability distribution model includes a connected feature extraction module and a target classification module. The feature extraction module is used to extract the high-dimensional hidden layer representation of the input sequence. The target classification module includes a connected fully connected layer and a normalized exponential function. Based on the extracted high-dimensional hidden layer representation, the target classification module learns and extracts the semantic features in the high-dimensional hidden layer representation, and outputs the association probability distribution between the statement and the sentence, or the association probability distribution between the statement and the sentence set.

[0015] A cascaded target classification architecture is set up using two association probability distribution models. The first association probability distribution model is set as a declaration-sentence association probability distribution model; the second association probability distribution model is set as a declaration-sentence set association probability distribution model.

[0016] S4. Based on the declaration-sentence association probability distribution model, learn the predicted probability distribution between the declaration and a single candidate sentence, quantify the local confidence of a single candidate sentence as the supporting basis for the declaration to be verified, and use it to measure the empirical value of the candidate sentence.

[0017] S5. Based on local confidence, select several candidate sentences with high empirical value from the candidate sentence set, and arrange the candidate sentences in descending order of empirical value to form a sentence combination as the candidate empirical set; based on the declaration-sentence set association probability distribution model, capture the semantic synergy effect between multiple candidate sentences in the candidate empirical set, and calculate the global category probability distribution of the declaration to be verified and the candidate empirical set.

[0018] S6. Adopt heterogeneous objective joint optimization to simultaneously train and update the parameters of two related probability distribution models;

[0019] During the training of the correlation probability distribution model, the loss calculated from the predicted probability distribution and the loss calculated from the global category probability distribution are weighted and fused to construct a joint loss function. The parameters of the two correlation probability distribution models are updated synchronously through the backpropagation algorithm to optimize the multi-level discrimination capability for inputs of different granularities and to achieve an automatic balance between local empirical accuracy and global judgment robustness.

[0020] S7. Based on the trained declaration-sentence association probability distribution model and declaration-sentence set association probability distribution model, refined empirical causal attribution and declaration verification are achieved through top-down or bottom-up extraction strategies.

[0021] Compared with existing technologies, the refined empirical causal attribution and statement verification method based on heterogeneous target fusion proposed in this invention innovatively introduces a multi-target fusion learning method based on the statement-sentence association probability distribution and the statement-sentence set association probability distribution: The statement-sentence association probability distribution model efficiently locates candidate statements most closely related to the statement from massive amounts of text, eliminating irrelevant or disruptive information; the statement-sentence set association probability distribution model performs comprehensive reasoning on the filtered candidate statements, outputting a minimum and necessary empirical set, and determining the truth or falsity of the statement. Compared with existing technologies, this invention not only reduces redundant interference while ensuring verification accuracy, but also provides more interpretable verification results, achieving rapid processing of massive amounts of text and accurate presentation of key empirical evidence, significantly improving verification efficiency and interpretability. Attached Figure Description

[0022] Figure 1 This is an overall flowchart of the detailed empirical causal attribution and declaration verification method in this embodiment of the invention;

[0023] Figure 2 This is a schematic diagram of the structure of the correlation probability distribution model constructed in the embodiment of the present invention. Detailed Implementation

[0024] This invention uses RoBERTa (Robustly Optimized BERT Pretraining Approach) as the core of feature extraction. By integrating dual classification objectives of "statement-sentence" and "statement-sentence set," it constructs a probability distribution model capable of modeling the association between local details and global context. During model training, a joint loss function is used to collaboratively optimize heterogeneous tasks, enabling the model to jointly represent sentence-level and set-level category distributions. Based on this, the invention deploys a target classification architecture with two cascaded association probability distribution models and designs two parsing strategies. These strategies aim to achieve sequential calculation of the statement-sentence association probability distribution and the statement-sentence set association probability distribution from the converged model after training, accurately extracting detailed empirical evidence and realizing statement verification. Detailed empirical evidence refers to the minimum set of sentences that supports or refutes a given statement. This minimum set of sentences satisfies the following condition: if any sentence in the minimum set is removed, the remaining sentences in the minimum set will no longer support or refute the given statement.

[0025] The present invention will be further described in detail below with reference to the embodiments and accompanying drawings, but the embodiments of the present invention are not limited thereto.

[0026] Example

[0027] This embodiment provides a refined empirical causation and claim verification method based on heterogeneous target fusion, and describes in detail the complete implementation process of claim verification and refined empirical causation method; and specifically uses a typical fact verification task as an example to demonstrate how to use the technical solution of this embodiment to perform refined empirical extraction and authenticity verification of the claim "someone is an actor born in a certain place".

[0028] See Figure 1 The detailed empirical causal ablation and assertion verification method in this embodiment specifically includes the following steps:

[0029] S1. A semantic parsing-based retrieval strategy is adopted to retrieve candidate documents related to the statement to be verified from the corpus through document retrieval.

[0030] First, key phrases in the statement to be verified are extracted through constituent syntactic analysis to construct multidimensional query conditions. Then, a broad search is performed in an authoritative encyclopedia database using an online interface to obtain multiple candidate documents related to the statement to be verified, which are then used as search results. Finally, a document credibility assessment mechanism is introduced to select high-quality documents from the search results, forming a candidate document set with an empirical basis.

[0031] The system received the input statement "someone is an actor born in a certain place" and retrieved several relevant candidate documents from the Wikipedia corpus through document retrieval.

[0032] S2. Extract candidate sentences related to the statement to be verified from the candidate documents and construct a set of candidate sentences; further construct the statement-candidate sentence set to be verified sequence.

[0033] The candidate documents undergo fine-grained processing by transforming them into complete semantic units using text segmentation techniques. A filtering and normalization mechanism is established to remove interfering segments with abnormal lengths, missing semantic meanings, and extremely low relevance from the semantic units, resulting in candidate sentences relevant to the statement to be verified. Each candidate sentence is then subjected to text normalization to obtain a candidate sentence set. Furthermore, a verification sequence of declarations and candidate sentence sets is constructed, and an index mapping between candidate sentences and source documents is established to provide standardized data input for subsequent refined modeling. In the verification sequence, each element includes a declaration and a set of candidate sentences used to verify the truth or falsity of the declaration.

[0034] This step constructs a candidate sentence set from the retrieved candidate documents as follows:

[0035] "Someone was born on a certain day in a certain place": Directly provide the place of birth information;

[0036] "Someone is a famous actor from a certain place": confirms their actor status;

[0037] "Someone has won multiple film awards": This is related to an actor's identity, but not directly related to verifying their place of birth.

[0038] "Someone has acted in many classic movies": This is also related to an actor's identity, but not directly related to verifying their place of birth.

[0039] S3. Construct a correlation probability distribution model, such as Figure 2 As shown, the constructed association probability distribution model includes a connected feature extraction module and a target classification module. The feature extraction module uses a pre-trained language model to extract the high-dimensional hidden layer representation of the input sequence. The target classification module includes a connected fully connected layer and a normalized exponential function (Softmax). Based on the extracted high-dimensional hidden layer representation, the target classification module learns and extracts semantic features from the high-dimensional hidden layer representation, outputting the association probability distribution between declarations and sentences, or the association probability distribution between declarations and sentence sets. In this embodiment, the input sequence includes a "declaration-single sentence" sequence and a "declaration-sentence set" sequence.

[0040] Furthermore, a cascaded target classification architecture is constructed using two association probability distribution models. The first association probability distribution model is set as a declaration-sentence association probability distribution model, with its input sequence being a "declaration-single sentence" sequence; the second association probability distribution model is set as a declaration-sentence set association probability distribution model, with its input sequence being a "declaration-sentence set" sequence. The feature extraction modules (i.e., pre-trained language models) in both association probability distribution models have the same model parameters, i.e., parameter sharing; however, the parameters of the fully connected layers in the two association probability distribution models are different, i.e., parameter non-sharing.

[0041] The statement-sentence association probability distribution model aims to learn and extract local semantic interaction features between "statements and individual candidate sentences" to achieve empirical category identification at the single-sentence level. The statement-sentence set association probability distribution model aims to learn and extract global semantic collaborative features between "statements and sentence sets". Thus, through these two cascaded association probability distribution models, the system is provided with multi-granularity semantic representation and category identification capabilities, ranging from local single-point evidence to a global logical chain.

[0042] Furthermore, this embodiment uses a pre-trained language model based on RoBERTa to construct an association probability distribution model, which contains 350 million parameters and can effectively handle mixed Chinese and English text.

[0043] S4. Calculate the statement-sentence association probability distribution: Based on the statement-sentence association probability distribution model set in step S3, learn the predicted probability distribution between the statement and a single candidate sentence, quantify the local confidence of a single candidate sentence as the supporting basis for the statement to be verified, and use it to measure the empirical value of the candidate sentence.

[0044] The specific process of this step includes the following sub-steps:

[0045] S41. Based on the sequence to be verified in the statement-candidate sentence set, construct a "statement-single sentence" sequence as the input sequence for the statement-sentence association probability distribution model; the "statement-single sentence" sequence includes a start tag, statement, separator tag and a single candidate sentence.

[0046] Specifically, the candidate sentence set constructed in step S2 Each candidate sentence (or simply sentence) in the dataset is analyzed independently, resulting in a statement to be verified. Depend on Composed of 1 word, denoted as A candidate sentence Depend on Composed of 1 word, denoted as ,in The format of the "declaration-single sentence" sequence is defined as follows:

[0047] ;

[0048] [CLS] represents the start marker, used to aggregate the semantic features of the "declaration-sentence" sequence; [SEP] represents the separator, which acts as a delimiter to clearly define the boundary between the declaration and the individual candidate sentence, two independent text units.

[0049] S42. Input the "declaration-single sentence" sequence constructed in step S41 into the feature extraction module of the declaration-sentence association probability distribution model. After encoding the "declaration-single sentence" sequence, the feature extraction module outputs a 768-dimensional feature vector corresponding to the starting marker [CLS], and denoted as... .

[0050] Feature vector It characterizes the complete semantic relationship and deep logical interaction between the statement to be verified and a single candidate sentence, serving as direct input for subsequent tasks.

[0051] S43. The feature vector extracted in step S42 is further processed through a fully connected layer and a normalized exponential function in the declaration-sentence association probability distribution model to calculate the predicted probability distribution of the corresponding individual candidate sentence relative to the declaration to be verified. :

[0052] ;

[0053] in and These are learnable parameters. ; Represents the space of real numbers; The normalization exponential function is used to nonlinearly map the real values ​​of any range of the output after the linear transformation of the fully connected layer into a probability distribution vector whose elements sum to 1, so as to characterize the local confidence of the corresponding individual candidate sentence belonging to each classification dimension.

[0054] The predicted probability distribution These include the probabilities of support, rebuttal, and irrelevance, used to precisely characterize the empirical value relative to the claim to be verified in three dimensions:

[0055] Support probability represents the confidence level that a candidate sentence can confirm the statement to be verified;

[0056] The probability of refutation represents the confidence level that a candidate sentence can disprove the statement to be verified.

[0057] The irrelevance probability represents the confidence that a candidate sentence does not contain valid verification information for the statement to be verified.

[0058] This step involves the independent analysis of each statement-sentence pair. The system then enters the statement-sentence pair processing stage, where the core task is to independently evaluate the semantic relationship between each candidate sentence and the statement to be verified. For each candidate sentence, the model is constructed in the form of " Someone is an actor born in a certain place. Candidate sentence content The input sequence is subjected to deep semantic analysis by an encoder based on a RoBERTa-based pre-trained language model.

[0059] When processing the first candidate sentence, the model calculated a predicted probability distribution of [0.85, 0.05, 0.10], where 0.85 represents the probability of supporting the statement, 0.05 represents the probability of refuting the statement, and 0.10 represents the probability of being irrelevant to the statement, indicating that this candidate sentence has a strong supporting relationship with the statement to be verified. The predicted probability distribution of the second candidate sentence is [0.75, 0.15, 0.10], showing a strong supporting tendency. However, the predicted probability distributions of the third and fourth candidate sentences are [0.20, 0.05, 0.75] and [0.25, 0.10, 0.65], respectively, indicating that these two candidate sentences have a low correlation with the statement to be verified.

[0060] S5. Calculate the probability distribution of the statement-sentence set association: Based on the local confidence obtained in step S4, select several candidate sentences with high empirical value from the candidate sentence set, and arrange these candidate sentences in descending order of empirical value to form a sentence combination as the candidate empirical set; based on the statement-sentence set association probability distribution model set in step S3, capture the semantic synergy effect between multiple candidate sentences in the candidate empirical set, and calculate the global category probability distribution of the statement to be verified and the candidate empirical set.

[0061] In this embodiment, step S5 specifically includes the following steps:

[0062] S51. Sort all candidate sentences in the candidate sentence set in descending order of empirical value, remove candidate sentences that are irrelevant to the statement, and select several candidate sentences with high confidence based on the sequence length limitation of the pre-trained language model, and combine them in descending order to obtain sentence combinations that constitute the candidate empirical set. .

[0063] Specifically, the empirical value of the candidate matrix is ​​the maximum of the support and rebuttal probabilities output in step S43. For each candidate empirical set, there is only one combination of candidate sentences.

[0064] S52. Based on the candidate empirical set, construct a "declaration-sentence set" sequence as the input sequence for the declaration-sentence set association probability distribution model. The "declaration-sentence set" sequence includes the start marker [CLS], declaration c, separator marker [SEP], and a combination of sentences from the candidate empirical set.

[0065] Specifically, the constructed "declaration-sentence set" sequence is as follows: ,in The start tag is used to aggregate the global semantic features of the "declaration-sentence set" sequence, providing a unified semantic representation basis for the association probability distribution model; the separator tag is used to clearly define the boundary between the declaration and the sentence combination.

[0066] S53. Input the "declaration-sentence set" sequence constructed in step S52 into the feature extraction module of the declaration-sentence set association probability distribution model. After encoding the "declaration-sentence set" sequence, the feature extraction module outputs a 768-dimensional feature vector corresponding to the starting label, denoted as... .

[0067] Feature vector It is used to characterize the complete semantic relationship and deep logical interaction between the statement to be verified and the candidate empirical set, serving as direct input for subsequent tasks.

[0068] S54. The feature vector extracted in step S53 By using the fully connected layer and normalized exponential function in the declaration-sentence set association probability distribution model, the global probability distribution of sentence combinations relative to the declarations to be verified in the candidate empirical set is calculated. :

[0069] ;

[0070] in and These are trainable parameters. The global probability distribution comprehensively characterizes the overall support effect of the candidate empirical set on the claims to be verified. Specifically, the global probability distribution includes:

[0071] Sentence set support probability This is used to determine whether the current sentence combination constitutes a complete supporting chain of evidence.

[0072] Sentence set refutation probability This is used to determine whether the current sentence combination constitutes a complete chain of rebuttal evidence.

[0073] Probability of insufficient information It is used to determine whether the current candidate empirical set still lacks key information sufficient to infer its truth or falsity.

[0074] This step involves the candidate sentence selection and ranking mechanism. Based on the category probability of a declaration-single sentence, the system implements an intelligent candidate sentence selection strategy. By calculating the maximum value of the support probability and rebuttal probability for each candidate sentence, the candidate sentence with the highest discriminative value is identified.

[0075] In this embodiment, the maximum probability value of the first candidate sentence is 0.85, the second candidate sentence is 0.75, the third candidate sentence is 0.20, and the fourth candidate sentence is 0.25. Based on this ranking, the system selects the first two candidate sentences to form a sentence combination, effectively filtering out low-relevance interference information. This selection process demonstrates the efficiency advantage of this embodiment in processing a large number of candidate sentences, automatically focusing on the most relevant empirical content.

[0076] Step S5 is the judgment of the declaration-sentence set. Upon entering this step, the system begins to evaluate the overall support effect of the combination of candidate sentences on the declaration to be verified. In this step, the declaration-sentence set association probability distribution model receives statements in the form of " Someone is an actor born in a certain place. A person was born on a certain day in a certain month of a certain year in a certain place. This person is a famous actor from a certain place. The “declaration-sentence set” sequence is used as the input sequence.

[0077] Using the same encoder and classifier structure, the declaration-sentence set association probability distribution model calculates the comprehensive probability distribution of the sentence sequence [0.92, 0.03, 0.05]. This result shows that the combination of two or more candidate sentences with high empirical value increases the support probability of the statement to be verified from 0.85 for a single candidate sentence to 0.92, proving the effectiveness of multi-candidate sentence collaborative verification; at the same time, the rebuttal probability decreases to 0.03 and the uncertainty probability decreases to 0.05, indicating that the empirical verification of the combination of multiple candidate sentences significantly improves the reliability of the judgment.

[0078] S6. Adopt heterogeneous objective joint optimization, synchronously train and update the parameters of two related probability distribution models, and balance local empirical accuracy and global judgment robustness.

[0079] During the training of the correlation probability distribution model, the loss calculated in step S4 (the predicted probability distribution between a statement and a single candidate sentence) and the loss calculated in step S5 (the global category probability distribution between the statement to be verified and the candidate empirical set) are weighted and fused to construct a joint loss function. The parameters of the two correlation probability distribution models are then updated synchronously using the backpropagation algorithm. This optimizes the multi-level discrimination capability for inputs of different granularities and achieves an automatic balance between local empirical accuracy and global judgment robustness. Through this collaborative constraint of heterogeneous objectives, the model's sensitivity to fine-grained empirical evidence is enhanced, enabling it to accurately identify single-point evidence while grasping the overall logic. The specific process is as follows:

[0080] S61. Suppose the original training dataset includes several empirical samples. Each empirical sample includes: a statement to be verified, and a statement-sentence set classification label. And the minimal set of sentences used to confirm or disprove the statement c to be verified, wherein A value of 0 indicates that the minimum set of sentences supports the declaration to be verified. A value of 1 represents the minimum set of sentences that refutes the statement to be verified. A value of 2 indicates insufficient information in the minimum sentence set; based on the original training dataset, candidate sentences in the candidate sentence set obtained in step S2 are labeled to obtain the true classification labels for the statement and individual sentences. ,in A value of 0 indicates that the single sentence supports the declaration. A value of 1 represents the rebuttal statement of that single sentence. A value of 2 indicates that the single sentence is irrelevant to the statement.

[0081] S62. Based on the predicted probability distribution obtained in step S4 and the true classification label obtained in step S61, calculate the empirical value judgment loss at the level of a single candidate sentence as the sentence-statement pair association loss. Sum the sentence-statement pair association losses of all candidate sentences to obtain the total loss of the statement-single sentence association probability distribution.

[0082] This step involves training the model using a targeted loss function. The core logic is to calculate the deviation between the predicted probability distribution and the true classification label (such as cross-entropy loss).

[0083] Let the first The predicted probability distribution of the candidate sentences is as follows: , will the The sentence-statement pair association loss between candidate sentences and statements. Defined as:

[0084] ;

[0085] in, It is a predicted probability distribution The One element, Indicates the first The support probability of each candidate sentence Indicates the first The probability of rebuttal for each candidate sentence. Indicates the first The irrelevance probability of each candidate sentence; For indicator functions, if If true, the indicator function takes the value 1. If false, the indicator function takes the value 0.

[0086] The sum of all sentence-statement pair association losses is taken as the total loss of the statement-single sentence association probability distribution:

[0087] ;

[0088] in The total number of candidate sentences in the candidate sentence set obtained in step S2.

[0089] The joint training mechanism in this step enables the statement-sentence association probability distribution model to accurately distinguish which candidate sentences contain key evidence and which are distracting information lacking empirical value. The losses of all candidate sentence-statement pairs are summed to form a global optimization objective. Through this joint training mechanism, the system can achieve optimal empirical extraction accuracy at the sentence level in complex multi-sentence scenarios, providing a scientific basis for subsequent combined verification.

[0090] S63. Based on the candidate empirical set obtained in step S51, generate training samples for the declaration-sentence set association probability distribution model.

[0091] First, initialize the sentence set S' of the training samples as an empty set. Then, starting from the empty set, select the candidate empirical set... Candidate sentences, sorted in descending order of empirical value, are progressively added to the sentence set. After each addition, it is determined whether the current sentence set contains the minimum set of sentences from the empirical samples in the original training dataset. If it does, a training sample is obtained, which specifically includes: the statement to be verified c, the sentence set S', and the corresponding statement-sentence set classification labels. , where 0 represents that the sentence set S' supports the statement to be verified, 1 represents that the sentence set S' refutes the statement to be verified, and 2 represents that the sentence set S' lacks sufficient information.

[0092] S64. Based on the global probability distribution, calculate the global association loss between the statement and the sentence set of the training samples obtained in step S63.

[0093] To train the statement-sentence set association probability distribution model to accurately identify sentence combinations that constitute a "complete and concise" empirical chain, and to ensure that the system has high-precision collaborative verification capabilities in complex multi-sentence environments, the global probability distribution calculated in step S54 is used. The global association loss between the statement and the set of sentences in the training samples. Defined as:

[0094] ;

[0095] in, It is a global probability distribution The One element, indicating function Used to identify the matching between the probability distribution prediction results and the declaration-sentence set classification labels of the training samples in step S63.

[0096] S65. Construct a joint loss function for heterogeneous objectives and simultaneously update the parameters of the declaration-sentence association probability distribution model and the declaration-sentence set association probability distribution model.

[0097] Specifically, the total loss of the declaration-single sentence association probability distribution calculated in step S62 is... The global association loss calculated in step S64 Weighted fusion is performed to construct a joint loss function for heterogeneous objectives (i.e., a joint optimization objective function). The specific formula is defined as follows:

[0098] ;

[0099] in, The preset loss weight hyperparameter is used to dynamically adjust the weight ratio of the local single-sentence empirical extraction task and the global sentence combination verification task in the overall training of the association probability distribution model. The value range is (0,1).

[0100] Based on the aforementioned joint loss function for heterogeneous objectives, a gradient descent algorithm is used for backpropagation. During backpropagation, the error gradient passes through the fully connected layers of both the declaration-sentence association probability distribution model and the declaration-sentence set association probability distribution model, and finally converges in the pre-trained language model at the bottom layer of the two association probability distribution models. This allows the parameters of the pre-trained language model and the weights of the two fully connected layers to be iterated and updated synchronously. Through this collaborative joint training of heterogeneous objectives, a unified network architecture is forced to learn more generalizable multi-granular feature representations: utilizing both the total loss function of the association probability distribution and... The constraint model accurately selects valid empirical evidence and eliminates interference noise at the micro-level single-point dimension, while also utilizing global correlation loss. The constraint model captures the logical loops, mutual corroboration and complementary relationships between multiple empirical evidences at the macro level, ultimately enabling the system to automatically achieve a perfect balance between local empirical sensitivity and global judgment robustness in complex multi-sentence scenarios.

[0101] S7. Based on the trained declaration-sentence association probability distribution model and declaration-sentence set association probability distribution model, refined empirical causal attribution and declaration verification are achieved through top-down or bottom-up extraction strategies.

[0102] After the joint training of the two correlation probability distribution models is completed, this embodiment sequentially executes steps S1, S2, S4, and S5 on the statements to be verified in the test set, and obtains a complete candidate empirical set through step S51; then, refined empirical evidence (i.e., the minimum sufficient set of sentences supporting or refuting the statement, without redundant or contradictory information) is extracted from the complete candidate empirical set through the following two complementary extraction strategies:

[0103] Extraction Strategy 1: Adopt a top-down logical consistency extraction method. First, determine the verification result of the complete candidate empirical set. Then, add sentences to the subset that is initialized as an empty set step by step according to the empirical value sorting, until the judgment result generated by the subset is consistent with the verification result of the complete candidate empirical set. Output the current smallest subset as the refined empirical evidence.

[0104] The extraction strategy first uses all sentences from the complete candidate empirical set to verify the statements, and then gradually extracts refined empirical evidence. Specifically, the relationships between the statements and sentence combinations in the complete candidate empirical set are determined by executing steps S52-S54. If the determination result is that the statements are unrelated, an empty set is returned directly; otherwise, each sentence in the complete candidate empirical set is traversed and added to a subset initialized to an empty set. If the current subset yields the same determination result as the complete candidate empirical set through steps S52-S54, the current subset is returned as refined empirical evidence. This strategy ensures that the found subset can produce the same determination result as the complete candidate empirical set, while minimizing the number of sentences to achieve the goal of accuracy.

[0105] Extraction Strategy 2: Adopt a bottom-up minimum sufficiency construction method, add sentences of the complete candidate empirical set to the subset initialized as an empty set in sequence. Each time a sentence combination is added, the relationship between the statement and the sentence combination is judged through steps S52-S54. Once the judgment result is to confirm or disprove the statement, the search is stopped immediately and the current subset is output as refined empirical evidence.

[0106] Strategy 2 involves progressively constructing refined empirical evidence by considering sentences from the complete candidate empirical set in sequence, returning immediately once a condition is met. Specifically, starting with a subset initialized as an empty set, each sentence in the complete candidate empirical set is checked sequentially, and the current sentence is added to the subset. The updated sentence combination in the current subset is then checked to see if it can be confirmed or disproven through steps S52-S54. If it can, the current subset and the result are returned immediately; otherwise, the next sentence is added, continuing until all sentences in the complete candidate empirical set have been traversed. This strategy is more efficient, as the search stops once the smallest set of sentences sufficient to support or refute the claim is found.

[0107] During the testing phase, the system employs Strategy 2 for declaration verification and fine-grained empirical causation. This strategy starts with an empty set and gradually adds sentences sorted by relevance until the model outputs a clear judgment result.

[0108] In the specific execution process, the system first adds the first sentence, "Someone was born on a certain day in a certain month of a certain year in a certain place," to the empirical set, and then calls the trained model to make a judgment. The model judges the first sentence as "supporting," with a support probability of 0.85, exceeding the preset judgment threshold. Since a clear support judgment has been obtained, the system immediately terminates the search process and outputs this sentence as refined empirical evidence. This process fully embodies the core concept of refined empirical extraction: finding the smallest set of sentences that support or refute the statement.

[0109] Through the complete execution process of this embodiment, the system ultimately outputs the predicted label "Support" and refined empirical evidence {a person, born on a certain date in a certain place}. Compared to the redundant empirical set that traditional methods may extract, this embodiment successfully eliminates unnecessary sentences, achieving accuracy in the empirical evidence.

[0110] This embodiment fully demonstrates the technical advantages of the cascaded heterogeneous target fusion framework. By minimizing the combined loss function, the model can simultaneously optimize both single-sentence judgment capabilities and multi-sentence collaborative analysis capabilities. The combination of a loss value of 0.451 for the declaration-single-sentence association probability distribution and a loss value of 0.083 for global association ensures that the model achieves good performance at both levels.

[0111] To verify the beneficial effects of the method of this invention, experiments were conducted to compare this invention with two other methods that also perform joint learning. The experiments used the general dataset FEVER for fact-checking, and the experimental results are shown in Table 1.

[0112] Table 1. Experimental Comparison Results

[0113]

[0114] Table 1 shows the performance comparison data between the method of the present invention and the prior art. The specific evaluation indicators used are as follows:

[0115] FEVER score: a metric for measuring the accuracy of claim verification, which is based on the premise that the predicted empirical evidence must cover at least one true empirical evidence.

[0116] Label accuracy: measures the accuracy of the classification results of the claim validation, without considering whether the predicted empirical results are correct;

[0117] Empirical precision: Predicts the macro-precision of empirical evidence for both supporting and refuting claims;

[0118] Empirical recall: Predict the macro-recall rate of empirical evidence for supporting and refuting claims;

[0119] F1 score: Predicts the empirical F1 score for both supporting and rebuttal statements.

[0120] Experimental results show that this invention achieves a breakthrough in the accuracy of empirical extraction while maintaining high claim verification efficiency: the empirical accuracy reaches 90.83%, significantly outperforming the comparison techniques. This indicates that the cascaded heterogeneous target fusion framework proposed in this invention, by jointly modeling heterogeneous targets of claims and single sentences, and claims and sentence sets, can more effectively identify fine-grained empirical evidence and avoid interference from irrelevant information. The empirical F1 score of this invention (81.42%) is significantly higher than that of Method 1 (50.59%) and Method 2 (64.98%). Although the label accuracy fluctuates slightly, in the FEVER score, a comprehensive indicator reflecting the verification results and empirical quality, this invention far surpasses Method 1 and is in the same tier as Method 2. This invention significantly improves the quality of empirical extraction without sacrificing core verification capabilities; this high-precision fine-grained empirical evidence has important practical application value for explaining software behavior and improving system transparency.

[0121] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A refined empirical causal attribution and assertion verification method based on heterogeneous target fusion, characterized in that, Includes the following steps: S1. A semantic parsing-based retrieval strategy is adopted to retrieve candidate documents related to the statement to be verified from the corpus through document retrieval; S2. Extract candidate sentences related to the statement to be verified from the candidate documents and construct a candidate sentence set; S3. Construct an association probability distribution model, which includes interconnected feature extraction modules and target classification modules; The feature extraction module is used to extract the high-dimensional hidden layer representation of the input sequence; the target classification module includes a connected fully connected layer and a normalized exponential function. Based on the extracted high-dimensional hidden layer representation, the target classification module learns and extracts the semantic features in the high-dimensional hidden layer representation, and outputs the probability distribution of the association between the statement and the sentence, or the probability distribution of the association between the statement and the sentence set. A cascaded target classification architecture is set up using two association probability distribution models. The first association probability distribution model is set as a declaration-sentence association probability distribution model; the second association probability distribution model is set as a declaration-sentence set association probability distribution model. S4. Based on the declaration-sentence association probability distribution model, learn the predicted probability distribution between the declaration and a single candidate sentence, quantify the local confidence of a single candidate sentence as the supporting basis for the declaration to be verified, and use it to measure the empirical value of the candidate sentence. S5. Based on local confidence, select several candidate sentences with high empirical value from the candidate sentence set, and arrange the candidate sentences in descending order of empirical value to form a sentence combination as the candidate empirical set; based on the declaration-sentence set association probability distribution model, capture the semantic synergy effect between multiple candidate sentences in the candidate empirical set, and calculate the global category probability distribution of the declaration to be verified and the candidate empirical set. S6. Adopt heterogeneous objective joint optimization to simultaneously train and update the parameters of two related probability distribution models; During the training of the correlation probability distribution model, the loss calculated from the predicted probability distribution and the loss calculated from the global category probability distribution are weighted and fused to construct a joint loss function. The parameters of the two correlation probability distribution models are updated synchronously through the backpropagation algorithm to optimize the multi-level discrimination capability for inputs of different granularities and to achieve an automatic balance between local empirical accuracy and global judgment robustness. S7. Based on the trained declaration-sentence association probability distribution model and declaration-sentence set association probability distribution model, refined empirical causal attribution and declaration verification are achieved through top-down or bottom-up extraction strategies.

2. The refined empirical causal abduction and assertion verification method according to claim 1, characterized in that, Step S2 further constructs the sequence to be verified from the declaration-candidate sentence set; In the sequence to be verified, each element includes a statement and a set of candidate sentences for verifying the truth or falsity of the statement; Step S4 includes: S41. Based on the sequence to be verified in the statement-candidate sentence set, construct a "statement-single sentence" sequence as the input sequence for the statement-sentence association probability distribution model. The "statement-single sentence" sequence includes a start tag, a statement, a separator tag, and a single candidate sentence. The start tag is used to aggregate the semantic features of the "statement-single sentence" sequence, and the separator tag is used to clearly define the boundary between the statement and the single candidate sentence. S42. Input the "declaration-single sentence" sequence into the feature extraction module of the declaration-sentence association probability distribution model. After encoding the "declaration-single sentence" sequence, the feature extraction module outputs the feature vector corresponding to the starting marker. eigenvectors Used to characterize the complete semantic relationship and deep logical interaction between the statement to be verified and a single candidate sentence; S43, The feature vector By using the fully connected layer and normalized exponential function in the declaration-sentence association probability distribution model, the predicted probability distribution of a single candidate sentence relative to the declaration to be verified is calculated.

3. The refined empirical causal abduction and assertion verification method according to claim 2, characterized in that, In step S41, the candidate sentence set constructed in step S2 is... Each candidate sentence in the text is analyzed independently, and a statement to be verified is generated. Depend on Composed of 1 word, denoted as A candidate sentence Depend on Composed of 1 word, denoted as ,in The format of the "declaration-single statement" sequence is defined as follows: ； Where [CLS] represents the start marker; [SEP] represents the separator marker.

4. The refined empirical causal abduction and assertion verification method according to claim 2, characterized in that, The predicted probability distribution calculated in step S43 is as follows: ； in and These are learnable parameters; The normalization exponential function is used to nonlinearly map the real values ​​of any range of the output after the linear transformation of the fully connected layer into a probability distribution vector whose elements sum to 1, so as to characterize the local confidence of the corresponding individual candidate sentence belonging to each classification dimension. Predicted probability distribution These include the probabilities of support, rebuttal, and irrelevance, used to precisely characterize the empirical value relative to the claim to be verified in three dimensions: Support probability represents the confidence level that a candidate sentence can confirm the statement to be verified; The probability of refutation represents the confidence level that a candidate sentence can disprove the statement to be verified. The irrelevance probability represents the confidence that a candidate sentence does not contain valid verification information for the statement to be verified.

5. The refined empirical causal abduction and assertion verification method according to claim 1, characterized in that, Step S5 includes: S51. Sort all candidate sentences in the candidate sentence set in descending order of empirical value, remove candidate sentences that are irrelevant to the statement, select a number of candidate sentences with high confidence and arrange them in descending order to obtain sentence combinations, which constitute the candidate empirical set. S52. Based on the candidate empirical set, construct a "declaration-sentence set" sequence as the input sequence for the declaration-sentence set association probability distribution model. The "declaration-sentence set" sequence includes a start tag, declarations, delimiters, and sentence combinations from the candidate empirical set. The start tag is used to aggregate the global semantic features of the "declaration-sentence set" sequence, providing a unified semantic representation basis for the association probability distribution model. The delimiters are used to clearly define the boundaries between declarations and sentence combinations. S53. Input the "declaration-sentence set" sequence into the feature extraction module of the declaration-sentence set association probability distribution model. After encoding the "declaration-sentence set" sequence, the feature extraction module outputs the feature vector corresponding to the starting marker. ; Eigenvector Used to characterize the complete semantic relationship and deep logical interaction between the statement to be verified and the candidate empirical set; S54, The feature vector By using the fully connected layer and normalized exponential function in the declaration-sentence set association probability distribution model, the global probability distribution of sentence combinations in the candidate empirical set relative to the declarations to be verified is calculated.

6. The refined empirical causal abduction and assertion verification method according to claim 5, characterized in that, The "declaration-sentence set" sequence constructed in step S52 is as follows: ,in For the combination of sentences in the candidate empirical set, As the starting marker, For the purpose of this statement, Used as a separator.

7. The refined empirical causal abduction and assertion verification method according to claim 5, characterized in that, The global probability distribution calculated in step S54 is as follows: ； in and These are trainable parameters; The global probability distribution is used to comprehensively characterize the overall support effect of the candidate empirical set on the claims to be verified. The global probability distribution specifically includes: Sentence set support probability This is used to determine whether the current sentence combination constitutes a complete supporting chain of evidence. Sentence set refutation probability This is used to determine whether the current sentence combination constitutes a complete chain of rebuttal evidence. Probability of insufficient information It is used to determine whether the current candidate empirical set still lacks key information sufficient to infer its truth or falsity.

8. The refined empirical causal abduction and assertion verification method according to claim 5, characterized in that, Step S6 includes: S61. Suppose the original training dataset includes several empirical samples. Each empirical sample includes: a statement to be verified, and a statement-sentence set classification label. And the minimal set of sentences used to confirm or disprove the statement c to be verified, wherein A value of 0 indicates that the minimum set of sentences supports the declaration to be verified. A value of 1 represents the minimum set of sentences that refutes the statement to be verified. A value of 2 indicates insufficient information in the minimum sentence set; based on the original training dataset, candidate sentences in the candidate sentence set obtained in step S2 are labeled to obtain the true classification labels for the statement and individual sentences. ,in A value of 0 indicates that the single sentence supports the declaration. A value of 1 represents the rebuttal statement of that single sentence. A value of 2 indicates that the single sentence is irrelevant to the declaration; S62. Based on the predicted probability distribution obtained in step S4 and the true classification labels obtained in step S61, calculate the empirical value judgment loss at the level of a single candidate sentence as the sentence-statement pair association loss. Summate the sentence-statement pair association losses for all candidate sentences to obtain the total loss of the statement-single sentence association probability distribution. ; S63. Based on the candidate empirical set obtained in step S51, generate training samples for the declaration-sentence set association probability distribution model. First, the sentence set S' of the training samples is initialized as an empty set. Then, starting from the empty set, candidate sentences from the candidate empirical set, sorted in descending order of empirical value, are gradually added to the sentence set. After each addition, it is determined whether the current sentence set contains the minimum sentence set of the sample empirical data from the original training dataset. If it does, a training sample is obtained. The training sample includes the statement to be verified c, the sentence set S', and the corresponding statement-sentence set classification label. ; S64. Based on the global probability distribution, calculate the global association loss between the statement and the sentence set of the training samples obtained in step S63; then calculate the global association loss. Defined as: ； in, It is a global probability distribution The One element, indicating function Used to identify the matching between the probability distribution prediction results and the declaration-sentence set classification labels of the training samples in step S63; S65. Associate the total loss of the probability distribution of the declaration-single sentence. Global correlation loss We perform weighted fusion to construct a joint loss function for heterogeneous objectives, and simultaneously update the parameters of the declaration-sentence association probability distribution model and the declaration-sentence set association probability distribution model.

9. The refined empirical causal abduction and assertion verification method according to claim 8, characterized in that, In step S62, the first... The predicted probability distribution of the candidate sentences is as follows: , will the The sentence-statement pair association loss between candidate sentences and statements. Defined as: ； in, It is a predicted probability distribution The One element, Indicates the first The support probability of each candidate sentence Indicates the first The probability of rebuttal for each candidate sentence. Indicates the first The irrelevance probability of each candidate sentence; For indicator functions, if If true, the indicator function takes the value 1. If false, the indicator function takes the value of 0; The sum of all sentence-statement pair association losses is taken as the total loss of the statement-single sentence association probability distribution: ； in The total number of candidate sentences in the candidate sentence set obtained in step S2.

10. The refined empirical causal abduction and assertion verification method according to claim 5, characterized in that, Step S7 executes steps S1, S2, S4 and S5 sequentially on the statements to be verified in the test set, and obtains a complete candidate empirical set through step S51; then, refined empirical evidence is extracted from the complete candidate empirical set through a top-down or bottom-up extraction strategy. The top-down extraction strategy is as follows: First, determine the verification result of the complete candidate empirical set. Then, add sentences to the subset that is initialized as an empty set step by step according to the empirical value sorting, until the judgment result generated by the subset is consistent with the verification result of the complete candidate empirical set. Output the current smallest subset as the refined empirical evidence. The bottom-up extraction strategy is as follows: sentences of the complete candidate empirical set are added sequentially to the subset initialized as an empty set. Each time a sentence combination is added, the relationship between the statement and the sentence combination is determined through steps S52-S54. Once the determination result is a confirmed or disproven statement, the search is stopped immediately and the current subset is output as refined empirical evidence.

Images