True and false judgment method, system and equipment for cross-domain declaration and storage medium

Abstract

The invention belongs to the field of field generalization, and particularly relates to a cross-field declaration true and false judgment method, system and device and a storage medium, a training declaration and evidence are aligned, low-level features and high-level features are obtained through an encoder, a feature matrix is formed, the encoder is trained through the feature matrix, and a true and false judgment result is obtained. Obtaining expert network domain specific features through the expert network, training the gating network, obtaining aggregation features in combination with the first weight and the second weight, obtaining coupling features according to the aggregation features and the advanced features, training a classifier to obtain an optimized classifier, and inputting the coupling features into the optimized classifier to obtain the optimal classifier. And the declarated label category probability is obtained. By decoupling the domain invariant features and the domain specific features in the coarse-grained and fine-grained reasoning units, the model can accurately perform cross-domain declaration verification, and the accuracy of a declaration judgment result is improved.

Description

Technical Field

[0001] The present invention belongs to the field of domain generalization, and specifically relates to a method, system, device and storage medium for determining the truth or falsity of cross-domain statements. Background Art

[0002] With the rapid growth in the scale and volume of false information, leveraging artificial intelligence (AI) to effectively distinguish true information from misleading content has become imperative. To address this challenge, the field of natural language processing has developed fact-verification systems, which assess the authenticity of ambiguous claims by collecting and analyzing sufficient textual evidence. Fact-verification systems analyze the evidence to assess the authenticity of claims and ultimately output one of three verdicts: support, refutation, or insufficient evidence to verify.

[0003] Current fact-verification research focuses on building complex neural network architectures to model the semantic relationships between claims and evidence, primarily including methods based on pre-trained language models and graph neural network techniques. These methods are typically designed for in-domain scenarios, assuming that training and test data come from the same source and satisfy the independent and identically distributed (IID) condition. However, this assumption is often difficult to maintain in practice. Furthermore, the high cost of manual annotation and data privacy issues make it impractical to collect labeled data for all potential domains to train high-performance models. Currently, research on domain-generalizable fact-verification systems that can effectively operate in unknown target domains remains limited. These methods can be broadly categorized into two categories: data manipulation-based methods and representation learning-based methods. The former employs specific rules to enhance the diversity of target domain samples or generate pseudo target samples for model training. The latter utilizes unlabeled target data to pre-train language models or explicitly aligns the representations of source and unlabeled target data through distance-based optimization methods. Furthermore, some research explores the use of instruction-fine-tuned large language models to generate auxiliary information to alleviate the challenges of few-shot fact-verification. Despite some progress, these methods still have significant limitations. Specifically, they typically require a large amount of unlabeled target domain data for language model pre-training, or a small amount of labeled target samples to construct pseudo samples. For methods based on large language models, achieving ideal results often requires providing examples from the target domain for contextual learning. However, in most practical applications, obtaining labeled or unlabeled data in the target domain is extremely difficult, either impossible or extremely challenging.

[0004] Existing studies have explored the problem of domain generalization in text classification. These methods learn domain-invariant representations through distance-based distribution adaptation or learn domain-specific features by generating domain-related attribute cues. However, these methods face challenges when applied to fact verification tasks due to the complex and contradictory logical relationships inherent in claim-evidence pairs. When part of the argument does not match the evidence, the claim may be refuted by the evidence even if other elements are supported. Therefore, directly optimizing the representation distribution differences of claim-evidence pairs across different domains may obscure the crucial fine-grained logical relationships that are key domain-invariant features in claim verification, thereby leading to incorrect judgments about the claims. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method, system, device and storage medium for judging the truth or falsity of cross-domain statements. By decoupling the domain-invariant features and domain-specific features in coarse-grained and fine-grained reasoning units, it is possible to accurately perform cross-domain statement verification and improve the accuracy of statement judgment results.

[0006] A method for determining the truth or falsity of a cross-domain statement, comprising:

[0007] Obtaining training statements and evidence pre-set in a database, the training statements and evidence being from different fields;

[0008] Dividing the training statement into a plurality of training statement blocks, aligning the training statement blocks with evidence to obtain block alignment pairs, and aligning the training statement with evidence to obtain argument-evidence pairs;

[0009] Inputting the block alignment pairs and argument-evidence pairs into an encoder respectively to obtain low-level block alignment pair features and low-level argument-evidence pair features, projecting the low-level block alignment pair features and the low-level argument-evidence pair features to obtain training high-level features, the training high-level features including block alignment pair high-level features and argument-evidence pair high-level features, and combining several block alignment pair high-level features in the same domain into a training feature matrix of argument-evidence pairs;

[0010] Obtain an expert network, process the low-level block alignment features and low-level argument evidence features through the expert network, and obtain domain-specific features of expert networks in different fields;

[0011] Get the gated network;

[0012] Obtaining a first weight and a second weight according to the low-level block alignment pair features, the low-level argument evidence pair features, and the gating network;

[0013] Obtaining training aggregate features based on the first weight, the second weight, and domain-specific features of expert networks in different fields;

[0014] Obtaining a training coupled feature based on the training high-level feature and the training aggregate feature;

[0015] Inputting the training coupled features into a classifier to obtain training statement label category probabilities;

[0016] Obtaining a classifier loss based on the training declared label category probabilities and the true categories, and adjusting the classifier weights based on the classifier loss to obtain an optimized classifier;

[0017] According to the training feature matrix, an encoder loss is calculated, an encoder weight is adjusted according to the encoder loss to obtain an optimized encoder, and actual claims and evidence are input into the optimized encoder to obtain actual low-level block alignment pair features, actual low-level argument-evidence pair features, and actual high-level features;

[0018] The actual expert network domain-specific features are obtained according to the actual low-level block alignment pair features and the actual low-level argument evidence pair features. The actual aggregation features are obtained according to the actual expert network domain-specific features, the first weight and the second weight. The actual coupling features are obtained according to the actual aggregation features and the actual high-level features. The actual coupling features are input into the optimized classifier to obtain the declared label category probability, and the maximum label category probability is selected as the judgment result.

[0019] Specifically, a claim is a proposition, opinion, or assertion that requires verification or proof. A claim itself does not provide information supporting its truthfulness, but rather relies on other evidence to establish its credibility.

[0020] Evidence is specific information, data, facts, or materials used to support or refute a claim. It provides an objective basis to help determine the truthfulness or validity of a claim.

[0021] A cross-domain claim means that the evidence used to judge the truth or falsity of the claim and the claim usually do not belong to the same field, that is, the truth or falsity of the claim is judged by evidence from other fields.

[0022] Optionally, the block alignment pairs and argument-evidence pairs are input into an encoder respectively to obtain low-level block alignment pair features and low-level argument-evidence pair features, and the low-level block alignment pair features and low-level argument-evidence pair features are projected to obtain training high-level features including:

[0023] The block alignment pairs and argument evidence pairs are input into the encoder respectively to obtain low-level block alignment pair features and low-level argument evidence pair features, which are expressed as:

[0024] X ce ,X al =Encoder(x ce ,x al );

[0025] x al =Concat({a1,…,a l});

[0026] Among them, Encoder is the encoder, Concat means splicing, a l For the declaration block, x al is a block alignment pair, x ce For the argument evidence pair, X ce is the low-level block alignment feature, X al Evidence for low-level arguments features;

[0027] And project the low-level block alignment features and low-level argument evidence features to obtain the training high-level features, which are expressed as:

[0028] z ce ,z al =f p (X ce ,X al );

[0029] Among them, z ce ,z al To train high-level features, f p Represents a projection.

[0030] Optionally, calculating an encoder loss according to the training feature matrix, and adjusting an encoder weight according to the encoder loss to obtain an optimized encoder includes:

[0031] The covariance matrix is ​​calculated based on the training feature matrix, and the calculation method is:

[0032]

[0033] in, is the covariance matrix of argument-evidence pairs in domain i, Is from the field The characteristic matrix of the argument evidence pair, T is the transposed moment, n i is the number of samples in domain i, 1 represents a column vector whose elements are all 1;

[0034] According to the covariance matrix of different fields, the first alignment loss of the argument-evidence pair is calculated and expressed as:

[0035]

[0036] in, represents the Frobenius norm of the matrix, is the alignment loss, is the covariance matrix of argument-evidence pairs in domain j, K is the number of domains;

[0037] Calculate the second alignment loss for the block alignment pair;

[0038] According to the high-level features of different fields, the first contrast loss of argument evidence to high-level features is calculated and expressed as:

[0039]

[0040]

[0041] Among them, A(i)≡I\{i} represents the set of all samples except sample i, P(i)≡{p∈A(i):y p =y i} represents the index set of all positive samples relative to sample i, |P(i)| represents the number of positive samples, the symbol represents the inner product, and τ is a scalar parameter. is the argument evidence for the positive sample to the high-level features, is the argument evidence pair high-level features of sample i, is the feature of the ath sample in the sample set A(i);

[0042] Calculate the second contrast loss of block alignment on high-level features;

[0043] Summing the first alignment loss, the second alignment loss, the first contrast loss, and the second contrast loss to obtain an encoder loss;

[0044] The encoder weights are adjusted according to the encoder loss to obtain an optimized encoder.

[0045] Optionally, the low-level block alignment features and low-level argument evidence features are processed by the expert network to obtain the domain-specific features of the expert network in different fields as follows:

[0046]

[0047] in, is the domain-specific feature of the expert network generated by the k-th expert network, x al is a block alignment pair, x ce For the argument evidence pair, is an expert network, and K represents the number of fields.

[0048] Optionally, obtain the gating network including:

[0049] Get the preset gating network;

[0050] According to the low-level block alignment pair features, the low-level argument evidence pair features and the preset gating network, a first training weight and a second training weight are obtained, which are expressed as:

[0051] w ce ,w al =Softmax(g a (x ce ,x al ));

[0052] Among them, w ce is the first training weight, w al is the second training weight, g a is the preset gating network, x al is a block alignment pair, x ce Evidence for the argument;

[0053] Obtaining a weight loss value according to the first training weight, the second training weight, and a weight loss formula;

[0054] The weight loss formula is:

[0055]

[0056] in, is the weight loss value, K is the number of fields, d i represents the domain label of the i-th sample, I is the indicator function, and denotes the weights assigned to the kth domain expert for the claim-evidence pair and the block alignment pair, respectively, is the first weight of the k-th domain expert, is the second weight of the k-th domain expert;

[0057] The network weights of the preset gating network are adjusted according to the weight loss value to obtain the gating network.

[0058] Optionally, according to the actual aggregated features and the actual high-level features, an actual coupling feature is obtained and expressed as:

[0059] x h =f h (Concat(z ce ,z al ,h ce ,h al ,Attention(z ce ,z al ,h ce ,h al )));

[0060] Among them, Concat means splicing, zce ,z al represents the actual high-level features, h ce ,h al represents the actual aggregated features, attention represents the self-attention operation applied to features from different channels, and f h (·) is the feature aggregation function based on the multi-layer perceptron MLP, x h is the actual coupling characteristic.

[0061] Optionally, the coupling feature is input into a classifier to obtain a training statement label category probability expression as:

[0062] P(y)=Softmax(f c (ReLU(x h )))

[0063] Among them, f c (·) represents the classifier function, ReLU represents the nonlinear activation function, and x h is the actual coupling characteristic.

[0064] A cross-domain statement truth judgment system, comprising:

[0065] A first acquisition module is used to acquire a training statement and evidence preset in a database, wherein the training statement and evidence come from different fields;

[0066] an alignment module, configured to divide the training statement into a plurality of training statement blocks, align the training statement blocks with evidence to obtain block alignment pairs, and align the training statement with evidence to obtain argument-evidence pairs;

[0067] a first feature conversion module, configured to input the block alignment pairs and argument-evidence pairs into an encoder respectively to obtain low-level block alignment pair features and low-level argument-evidence pair features, project the low-level block alignment pair features and the low-level argument-evidence pair features to obtain training high-level features, the training high-level features including block alignment pair high-level features and argument-evidence pair high-level features, and combine a plurality of block alignment pair high-level features in the same domain into a training feature matrix of argument-evidence pairs;

[0068] a first training module, configured to calculate an encoder loss based on the training feature matrix, adjust an encoder weight based on the encoder loss to obtain an optimized encoder, input actual claims and evidence into the optimized encoder to obtain actual low-level block alignment pair features, actual low-level argument-evidence pair features, and actual high-level features;

[0069] The second feature conversion module is used to obtain the expert network, which processes the low-level block alignment features and low-level argument evidence features through the expert network to obtain the domain-specific features of the expert network in different fields;

[0070] A second acquisition module is used to acquire a gated network;

[0071] a third acquisition module, configured to obtain a first weight and a second weight according to the low-level block alignment pair features, the low-level argument evidence pair features, and the gating network;

[0072] a first calculation module, configured to obtain a training aggregate feature based on the first weight, the second weight, and domain-specific features of expert networks in different fields;

[0073] A second calculation module is used to obtain a training coupling feature based on the training high-level features and the training aggregate features;

[0074] A third calculation module is used to input the training coupling features into a classifier to obtain a training statement label category probability;

[0075] A second training module is configured to obtain a classifier loss based on the training declared label category probabilities and the true categories, and adjust the classifier weights based on the classifier loss to obtain an optimized classifier;

[0076] The output module is used to obtain the actual aggregation feature based on the actual low-level block alignment pair feature, the actual low-level argument evidence pair feature and the gating network, obtain the actual coupling feature based on the actual aggregation feature and the actual high-level feature, input the actual coupling feature into the optimized classifier, obtain the declared label category probability, and select the largest label category probability as the judgment result.

[0077] A terminal device includes a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, a true or false judgment method of a cross-domain statement is adopted.

[0078] A computer-readable storage medium stores a computer program. When the computer program is loaded and executed by a processor, a true or false judgment method of a cross-domain statement is adopted.

[0079] The beneficial effects of the present invention are:

[0080] By decoupling the domain-invariant features and domain-specific features in coarse-grained and fine-grained reasoning units, the model can accurately perform cross-domain claim verification. In order to capture local logical relationships, we adopt a reasoning unit decoupling mechanism based on a large language model (LLM) to extract fine-grained fragment alignment pairs. Subsequently, a domain-invariant representation learning module based on contrast enhancement and a domain-specific representation learning module based on an expert network are used to optimize the feature distribution of the reasoning unit while learning domain-related knowledge. Compared to the prior art, this application can accurately perform cross-domain claim verification by decoupling the domain-invariant features and domain-specific features in coarse-grained and fine-grained reasoning units, without the need to label the data, thereby improving the accuracy of the claim judgment results.

[0081] 2. Extensive experiments on multiple datasets demonstrate that MIUFD outperforms competing baselines in reducing domain variance in feature distributions. Furthermore, our model exhibits enhanced robustness when the number of source domains and examples decreases. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 Schematic diagram of a flow chart of a method for determining the authenticity of a cross-domain statement of the present invention;

[0083] Figure 2 Schematic diagram of statistical information of data sets in different fields of the present invention;

[0084] Figure 3 Schematic diagram of the performance scores of different models of the present invention on four data sets;

[0085] Figure 4 Schematic diagram of the results of the ablation experiment of the present invention;

[0086] Figure 5 Schematic diagram of the A-distance performance between the source domain and the target domain for different data sets of the present invention;

[0087] Figure 6 This is a T-SNE visualization diagram of the source domain and target domain feature representation after model training of the present invention. Figure 6 (a) is a T-SNE visualization diagram of the baseline model RoBERTa after training. Figure 6 (b) is a T-SNE visualization diagram of the baseline model Deepcoral after training. Figure 6 (c) is a T-SNE visualization diagram of the baseline model Mixup after training. Figure 6 (d) T-SNE visualization diagram of the selected baseline model MiuFD after training;

[0088] Figure 7This is a schematic diagram showing the performance comparison of the models trained with two source domains in the present invention. Figure 7 (a) is a schematic diagram of the performance comparison of the model when the data set is scientific. Figure 7 (b) is a schematic diagram of the model performance comparison when the dataset is healthy;

[0089] Figure 8 Schematic diagram showing the performance comparison of the models trained with different numbers of source training samples in the present invention. Figure 8 (a) is a schematic diagram showing the performance comparison of models trained with different numbers of source training samples when the dataset is Crisis Encyclopedia. Figure 8 (b) is a schematic diagram comparing the performance of models trained with different numbers of source training samples when the dataset is Wikipedia. DETAILED DESCRIPTION

[0090] A method for determining the truth or falsehood of cross-domain statements, such as Figure 1 As shown, the present invention includes:

[0091] S1. Obtain training statements and evidence preset in the database. The training statements and evidence come from different fields.

[0092] S2. Divide the training statement into several training statement blocks, align the training statement blocks with the evidence to obtain block alignment pairs, and align the training statement with the evidence to obtain argument-evidence pairs;

[0093] Specifically, in order to construct reasoning units of different granularities for generalized feature learning, we first regard the original claim-evidence pair as a coarse-grained reasoning unit, which includes global logical relations. In order to generate fine-grained reasoning units containing local logical information, a large language model (LLM) tuned by instructions is used to decouple the reasoning units in the claim and evidence. Specifically, we input the claim c and the evidence sentence E = {e1,…,en} into the LLM, and the LLM outputs a series of "block-aligned pairs", denoted as A = {a1,…,a l This process consists of two phases: claim chunking and evidence alignment, both of which are performed by a command-tuned LLM. In the first phase, the claim is segmented into multiple chunks, each representing a coherent unit of text within the claim that can be independently verified against relevant evidence. The chunking process is guided by prompting the LLM with the directive, "\textit{Split the following text into smaller chunks that can be independently verified}." To ensure that the LLM generates output in the expected format, constrained decoding is employed to prevent the model from introducing new words, skipping words, or inappropriately modifying information.

[0094] A claim is a statement, assertion, or opinion that needs to be verified. It is usually a factual assertion made by someone that may be true or false. It can be an explicit statement of fact (such as "a certain data / event is X") or an implicit subjective judgment (such as "Policy X leads to an economic recession"). The claim must be verifiable (i.e., its truth or falsity can be proven through evidence).

[0095] For example, drinking eight glasses of water a day can speed up weight loss.”

[0096] Social Event: "A country's unemployment rate reached a historical high in 2023."

[0097] Scientific controversy: “Global warming is primarily caused by human activities.

[0098] Evidence is objective facts, data, research results or authoritative information used to support or refute a statement and must be reliable (such as coming from authoritative institutions, peer-reviewed research, etc.).

[0099] For example:

[0100] Evidence 1: Research from Harvard Medical School shows that there is no direct causal relationship between drinking water and metabolic rate (authoritative research).

[0101] Evidence 2: Clinical trials show that increasing water intake has no statistically significant effect on weight loss (experimental data).

[0102] Verify that "a certain country's unemployment rate has reached a new high."

[0103] Evidence 1: Unemployment rate data for 2023 released by the country’s statistics bureau (official data).

[0104] Evidence 2: Annual employment report of the International Labor Organization (ILO) (a third-party authoritative report).

[0105] Support the claim that “global warming is caused by human activities”.

[0106] Evidence 1: The IPCC (United Nations Panel on Climate Change) report points out that industrial carbon emissions are the main cause (scientific consensus).

[0107] Evidence 2: Data on the correlation between atmospheric carbon dioxide concentration and industrialization process (observational data).

[0108] Cross-domain claims are claims that need to be verified and the evidence to verify the truth of the claim is not from the same domain.

[0109] In the second stage, we align the parts of the evidence sentence that are relevant to the claim block with the claim block. LLM achieves this alignment by taking the claim, evidence, and the previously generated claim block as input. The prompt template structure for this stage is as follows: \textit{"c,E, aligns the following claim expression with the relevant substring in the evidence text: <chunk-1> …… <chunk-n>To mitigate hallucination effects and ensure that LLM does not align chunks with tokens that come only from claims, we impose constraints on the model requiring it to generate alignments using only tokens present in the evidence, excluding any claim-specific tokens.

[0110] Based on these strategies, we are able to infer logical relations between local arguments of claims and evidence, which are important domain-invariant knowledge that generalizes to unseen domains. Moreover, the original claim-evidence pairs capture global relational information as well as domain-specific details, both of which contribute to learning domain-invariant and domain-specific features, thereby enhancing the final verdict prediction.

[0111] S3. Input the block alignment pairs and argument-evidence pairs into the encoder respectively to obtain low-level block alignment pair features and low-level argument-evidence pair features, and project the low-level block alignment pair features and low-level argument-evidence pair features to obtain training high-level features. The training high-level features include block alignment pair high-level features and argument-evidence pair high-level features. Several block alignment pair high-level features in the same domain are combined into a training feature matrix of argument-evidence pairs.

[0112] The block alignment pairs and argument evidence pairs are input into the encoder respectively to obtain low-level block alignment pair features and low-level argument evidence pair features, and the low-level block alignment pair features and low-level argument evidence pair features are projected to obtain the training high-level features including:

[0113] The block alignment pairs and argument evidence pairs are input into the encoder respectively to obtain low-level block alignment pair features and low-level argument evidence pair features, which are expressed as:

[0114] X ce ,X al =Encoder(x ce ,x al );

[0115] x al =Concat({a1,…,a l });

[0116] Among them, Encoder is the encoder, Concat means splicing, a l For the declaration block, x al is a block alignment pair, x ce For the argument evidence pair, X ce is the low-level block alignment feature, X al are low-level argument-evidence pair features, which are believed to reflect the characteristics of coarse-grained and fine-grained reasoning units.

[0117] To learn domain-invariant representations for the inference unit, we feed these low-level features into a projection layer to generate high-level features.

[0118] And project the low-level block alignment features and low-level argument evidence features to obtain the training high-level features, which are expressed as:

[0119] z ce ,z al =f p (X ce ,X al );

[0120] Among them, z ce ,z al To train high-level features, f p Represents a projection.

[0121] S4. Obtain an expert network, and process the low-level block alignment features and low-level argument evidence features through the expert network to obtain domain-specific features of expert networks in different fields;

[0122] Specifically, samples from unseen domains may show similarities with one or more seen domains, allowing us to leverage an ensemble of expert networks to adaptively extract domain-specific features based on the information data. This approach enables the model to trigger relevant parameters to make decision predictions.

[0123] Specifically, given the low-level representations of claim-evidence pairs and block alignment pairs, we introduce K expert networks. These representations are processed by the expert networks as follows:

[0124] The expert network processes the low-level block alignment features and low-level argument evidence features, and obtains the domain-specific features of the expert network as follows:

[0125]

[0126] in, is the domain-specific feature of the expert network generated by the k-th expert network, x al is a block alignment pair, x ce For the argument evidence pair, is an expert network, and K represents the number of fields.

[0127] represents the kth expert network, which is used to process the claim-evidence pair and the block alignment pair. Each expert network consists of a single linear layer. Next, we design a gating network g a , used to generate feature selection weights for different expert networks.

[0128] S5, obtaining the gated network;

[0129] S6. Obtaining a first weight and a second weight according to the low-level block alignment pair features, the low-level argument evidence pair features, and the gating network;

[0130] According to the low-level block alignment pair features and the low-level argument evidence pair features and the gating network, the first weight and the second weight are obtained including:

[0131] Get the preset gating network;

[0132] Specifically, the preset gating network is a gating network that has not been trained yet, and its parameters are randomly set to initial values.

[0133] According to the low-level block alignment features, the low-level argument evidence features and the preset gating network, the first training weight and the second training weight are obtained, which are expressed as:

[0134] w ce ,w al =Softmax(g a (x ce ,x al ));

[0135] in, is the first training weight, is the second training weight, g a is the preset gating network, x al is a block alignment pair, x ce Evidence for the argument;

[0136] Specifically, g a Is a projection head:

[0137] Obtaining a weight loss value according to the first training weight, the second training weight, and the weight loss formula;

[0138] Specifically, in order to optimize the gating network and generate appropriate feature selection weights, we introduce a domain-specific loss function that uses domain labels as supervisory signals, as follows:

[0139] The weight loss formula is:

[0140]

[0141] in, is the weight loss value, K is the number of fields, d i represents the domain label of the i-th sample, I is the indicator function, and denotes the weights assigned to the kth domain expert for the claim-evidence pair and the block alignment pair, respectively, is the first weight of the k-th domain expert, is the second weight of the k-th domain expert;

[0142] The network weights of the preset gating network are adjusted according to the weight loss value to obtain the gating network.

[0143] After training, the domain-specific features from each expert network can be adaptively fused to cope with test samples from unseen domains, thereby facilitating effective decision prediction.

[0144] S7. Obtaining a training aggregate feature based on the first weight, the second weight, and the domain-specific feature of the expert network;

[0145] Specifically, the aggregation process is as follows:

[0146] After the gating network is trained, domain-specific representations of claim-evidence pairs and block alignment pairs are obtained by aggregating the features produced by the corresponding expert networks:

[0147]

[0148] Among them, H ce ,H al To train the aggregated features, is the first weight, is the second weight, and Domain-specific features of expert networks.

[0149] S8. Obtaining training coupled features based on the training high-level features and the training aggregated features;

[0150] Specifically, a heterogeneous feature fusion technique is used to combine the concatenated features with the attention-based interaction features to generate coupled features, with the goal of generating a comprehensive feature to facilitate accurate decision prediction.

[0151] According to the training high-level features and training aggregate features, the training coupled features are expressed as:

[0152] x h =f h (Concat(z ce ,z al ,h ce ,h al ,Attention(z ce ,z al ,h ce ,h al )));

[0153] Among them, Concat means splicing, z ce ,z al Indicates training high-level features, h ce ,h al Represents the training aggregate feature, attention represents the self-attention operation applied to features from different channels, f h (·) is the feature aggregation function based on the multi-layer perceptron MLP, x h To train coupling features.

[0154] S9. Input the coupling features into the classifier to obtain the training statement label category probability;

[0155] Input the coupled features into the classifier and obtain the training statement label category probability as:

[0156] P(y)=Softmax(f c (ReLU(x h )))

[0157] Among them, f c (·) represents the classifier function, ReLU represents the nonlinear activation function, and x h is a coupling feature.

[0158] S10, obtaining a classifier loss based on the training declared label category probability and the true category, and adjusting the classifier weight according to the classifier loss to obtain an optimized classifier;

[0159] Specifically, in model training, we use negative log-likelihood loss, the formula is as follows:

[0160]

[0161] Among them, P(y i,true ) represents the true label category y of the i-th sample i,true Finally, we combine the training objectives of domain-invariant feature learning, domain-specific feature learning, and decision prediction objectives to form an overall loss function, which is expressed as follows:

[0162]

[0163] Among them, λ1 and λ2 are hyperparameters used to control the relative contributions of domain-invariant loss and domain-specific loss, and overall is the overall loss sum.

[0164] S11. Calculate the encoder loss based on the training feature matrix, adjust the encoder weight according to the encoder loss to obtain an optimized encoder, input the actual claim and evidence into the optimized encoder, and obtain the actual low-level block alignment pair features, the actual low-level argument-evidence pair features, and the actual high-level features;

[0165] Specifically, a joint optimization strategy is adopted to learn a unified distribution of domain-invariant features across different domains and categories of samples. Specifically, by combining domain alignment with label-aware contrastive learning, the difference in feature distribution between different domains is reduced while the distance between samples between categories is increased. Specifically, given a batch I of training data from K different source domains, we use the alignment objective to align any two domains. and The second-order statistics (i.e., correlation) of the features between them are aligned.

[0166] According to the training feature matrix, the encoder loss is calculated, and the encoder weight is adjusted according to the encoder loss. The optimized encoder includes:

[0167] The covariance matrix is ​​calculated based on the training feature matrix and is calculated as follows:

[0168]

[0169] in, is the covariance matrix of argument-evidence pairs in domain i, Is from the field The characteristic matrix of the argument evidence pair, T is the transposed moment, n i is the number of samples in domain i, 1 represents a column vector whose elements are all 1;

[0170] According to the covariance matrix of different fields, the first alignment loss of the argument-evidence pair is calculated and expressed as:

[0171]

[0172] in, represents the Frobenius norm of the matrix, is the alignment loss, is the covariance matrix of argument-evidence pairs in domain j, K is the number of domains;

[0173] Calculate the second alignment loss for the block alignment pair;

[0174] Specifically, the calculation method of the second alignment loss of the block alignment pair is exactly the same as the calculation method of the first alignment loss.

[0175] According to the high-level features of different fields, the first contrast loss of argument evidence to high-level features is calculated and expressed as:

[0176]

[0177]

[0178] Among them, A(i)≡I\{i} represents the set of all samples except sample i, P(i)≡{p∈A(i):y p =y i } represents the index set of all positive samples relative to sample i, |P(i)| represents the number of positive samples, the symbol represents the inner product, and τ is a scalar parameter. is the argument evidence for the positive sample to the high-level features, is the argument evidence pair high-level features of sample i, is the feature of the ath sample in the sample set A(i);

[0179] Calculate the second contrast loss of block alignment on high-level features;

[0180] Specifically, in addition to reducing the representation differences between samples from different domains, we also introduce a contrast-based objective to optimize the distance between samples in the embedding space through label information. Similarly, the calculation method of the second contrast loss is exactly the same as that of the first contrast loss.

[0181] Summing the first alignment loss, the second alignment loss, the first contrast loss, and the second contrast loss to obtain the encoder loss;

[0182] Specifically, by jointly minimizing these losses, our model effectively reduces the distribution differences between samples from different domains while increasing the separation between samples from different categories. This ultimately generates a unified feature distribution, such as Figure 2 shown.

[0183] The encoder weights are adjusted according to the encoder loss to obtain the optimized encoder.

[0184] Specifically, the encoder weights are changed by back propagation to obtain an optimized encoder.

[0185] S12. According to the actual low-level block alignment pair features, the actual low-level argument evidence pair features and the gating network, the actual aggregation features are obtained. According to the actual aggregation features and the actual high-level features, the actual coupling features are obtained. The actual coupling features are input into the optimized classifier to obtain the declared label category probability, and the largest label category probability is selected as the judgment result.

[0186] Specifically, the label categories of a statement include three types: support, refutation, and insufficient evidence. Based on the output probability of the classifier for each statement label, the one with the highest probability is selected as the judgment result.

[0187] Experiments have shown that:

[0188] Experimental evaluation is conducted on four benchmark datasets from different fields, which are designed for fact verification tasks: FEVER (Wikipedia), Climate-FEVER (climate), Sci-Fact (science), and PubHealth (health). The sample format in each dataset is a claim-evidence-label triple, where the claim is labeled as "supported (S)", "refuted (R), or "not enough info (N)". The statistical information of these datasets is summarized as follows Figure 2 shown.

[0189] FEVER: The FEVER dataset contains crowdsourced statements collected from Wikipedia articles and represents a general domain dataset.

[0190] Climate-FEVER: The Climate-FEVER dataset includes artificial and real-world statements related to climate change collected from various online platforms.

[0191] Sci-Fact: The Sci-Fact dataset contains scientifically accurate statements, each supported by sentence-level rationale provided by experts in basic science, clinical medicine, and other relevant fields.

[0192] PubHealth: The PubHealth dataset contains statements extracted from various fact-checking websites and news media covering a wide range of health-related topics, including biomedical issues, government healthcare policies, and public health issues.

[0193] A leave-one-domain-out evaluation is performed on the above datasets. In the experiment, each dataset is designated as the target domain, while the remaining datasets are used as source domains for model training.

[0194] In this application, we use Label Accuracy (Acc.) and Macro-F1 (Macro-F1) as the primary evaluation metrics. Label Accuracy measures the overall proportion of correct predictions made by the model. It provides a direct assessment of the model's overall performance across all classes. Macro-F1, on the other hand, assesses the model's performance across all classes, treating each class equally, regardless of its frequency in the dataset. This metric is particularly valuable when working with imbalanced datasets, as it ensures that the performance of the minority class is given the same weight as the performance of the majority class.

[0195] Model Summary:

[0196] The proposed MIUFD model (the model adopted in this application) is compared with several competitive baselines to verify its performance on the generalized fact verification task. The baseline models are divided into two groups based on their methodology.

[0197] The first group contains mixed-domain methods that combine samples from multiple source domains and use empirical risk minimization (ERM) as the optimization objective. The following methods are considered:

[0198] A pre-trained model based on BERT, optimized to directly combine claims and evidence as input for claim label classification.

[0199] A bidirectional model based on the Gated Recurrent Unit (GRU), designed for sequential text processing. We use RoBERTa as the backbone model to generate basic token embeddings. RoBERTa is a widely used text classification model that uses convolutional layers to extract text features. RoBERTa is used as the backbone to generate token embeddings.

[0200] Models based on mixture of experts combine domain-specific features through a gating mechanism to make decision predictions.

[0201] The second group contains domain generalization methods designed for image and text classification. They include the following methods:

[0202] A feature alignment-based method that uses a kernel-based metric to minimize the maximum mean difference (MMD) between the source and target domains.

[0203] A fake news detection model based on adversarial neural networks that uses gradient reversal layers to learn domain-invariant features.

[0204] A feature alignment-based method reduces domain shift by aligning the second-order statistics of the feature distributions of the source and target domains.

[0205] A data augmentation-based method generates synthetic training data by linearly interpolating a pair of training samples.

[0206] An example-based hint learning algorithm activates relevant parameters in a language model by generating domain-specific hints.

[0207] Experimental setup:

[0208] Experiments are conducted on the above datasets using a leave-one-out-domain setting. Specifically, each dataset serves as the target domain, while the remaining datasets serve as the source domain for model training. In these experiments, the trained models are directly evaluated on target domain samples without any domain adaptation. In terms of model configuration, we use RoBERTa-base as a shared feature encoder for extracting low-level reasoning unit features. The maximum truncation length for claim-evidence pairs and block alignment is set to 150. We set the number of expert networks kk to 3. During training, we set the batch size to 48. The training objective of the model is to minimize the overall loss, using the Adam optimizer with an initial learning rate of 1e-5 and a weight decay of 1e-4. For the loss trade-off hyperparameters, we select the values ​​of λ1 and λ2 from the set {0.01, 0.05, 0.1, 1, 5} and select the model with the best performance on the joint source domain test set for the target domain data through grid search.

[0209] To test the effectiveness of our model, we focus on the following research questions to guide our experiments.

[0210] Does MIUFD improve claim verification performance in unseen target domains compared to baseline models? How does each component of MIUFD contribute to the overall performance? How does the domain shift of features learned by MIUFD differ from that of features learned by baseline models?

[0211] How does the robustness of MIUFD compare to baseline models? How do hyperparameters affect the performance of MIUDF?.

[0212] Experimental analysis:

[0213] To address the question of whether MIUFD improves the claim verification performance in unseen target domains compared to the baseline model, the overall performance of the baseline model and our proposed method is evaluated on four datasets from different domains. The results are summarized in Figure 3 middle. Figure 3 The best baseline model and the best performing model are underlined and bolded, respectively.

[0214] from Figure 3 As can be seen in Figure 2, models based on domain generalization generally outperform traditional claim verification methods on multiple datasets. For example, Deepcoral improves RoBERTa by 10.9% in F1-score and 8.58% in accuracy. This observation suggests that simply minimizing the empirical risk on the source domain does not significantly improve the generalization ability of the model. Among the baseline models based on domain generalization, Deepcoral performs best on most datasets. This finding suggests that domain alignment effectively alleviates the distribution differences between different domains, enabling the model to learn domain-invariant features. In addition, Mixup shows relatively strong performance in some models, surpassing Deepcoral by 0.25% and 2.13% in F1-score on the Climate-FEVER and Sci-Fact datasets, respectively. This suggests that generating synthetic training examples also helps the model learn more generalizable features.

[0215] Next, we focus on the performance comparison between the proposed model and baseline models. Overall, MIUFD consistently outperforms all baseline models on all datasets, achieving an average F1-score and accuracy improvement of 7.82% and 6.33%, respectively, compared to the best-performing baseline model, Deepcoral.

[0216] These improvements validate the superiority of the approach used in this application. The results demonstrate that the ability to simultaneously learn domain-invariant and domain-specific representations complement each other and improve model performance. Furthermore, the observed improvement can be attributed to the decoupling strategy for reasoning units, which extracts fine-grained logical relational features and facilitates the model's learning of domain-invariant discriminative reasoning knowledge in previously unseen test domains.

[0217] Ablation experiment:

[0218] To address the question of how each component of MIUFD contributes to the overall performance, we conduct ablation experiments to evaluate the impact of each component on the model performance. In these experiments, we systematically remove or replace key modules in our proposed model and Figure 4 The corresponding results are shown in . Specifically, Figure 4 As shown in the figure, we evaluated the effects of removing two key components: "without DI feature encoding" and "without DS feature encoding." These two experiments involve omitting the relevant inference unit features and loss functions. The "without inference unit decoupling" experiment uses only the original claim-evidence pairs without block alignment for claim verification. The "without feature coupling layer" experiment replaces the heterogeneous coupling layer with a direct feature concatenation. Furthermore, "without alignment loss" and "without contrastive loss" remove the corresponding losses in the domain-invariant representation learning module.

[0219] Figure 4 The results presented in

[15] clearly show that removing or replacing any component or loss function leads to a significant performance drop on most datasets. These findings highlight the effectiveness of the proposed module. In particular, excluding the domain-invariant feature encoding module leads to the most significant performance drop, with an average reduction of 11.82% in F1-score and 12.19% in accuracy. This highlights the importance of learning domain-invariant features in alleviating the domain shift problem. In addition, removing the alignment loss and contrastive loss leads to a reduction of 6.66% and 4.48% in F1-score, respectively. These findings suggest that the model benefits from the combination of feature alignment and label-aware contrastive optimization, with the former reducing inter-domain differences and the latter enhancing the separation of features from different categories.

[0220] Removing the domain-specific feature encoder, in addition to the domain-invariant features, results in an average drop of 7.87% in F1-score and 8.13% in accuracy, respectively. This highlights the effectiveness of incorporating domain-specific knowledge in decision prediction, which helps map unseen target domain samples to the relevant source domain and activate the corresponding expert network based on lexical similarity. Furthermore, we examined the impact of removing the LLM-based reasoning unit decoupling module. As expected, this variant experienced a 9.49% drop in average F1-score, marking the second largest performance drop. This suggests that constructing fine-grained reasoning units and modeling the relationship between them is crucial for extracting both domain-invariant and domain-specific features, ultimately improving the model's performance in out-of-domain settings. Finally, replacing the heterogeneous feature coupling layer with direct feature concatenation results in a 5.76% drop in F1-score. This suggests that the attention mechanism in the heterogeneous coupling layer promotes more effective interactions between multi-channel features, helping to generate more generalizable reasoning features.

[0221] Domain difference analysis:

[0222] To answer the question of how the domain shift of features learned by MIUFD differs from that of features learned by the baseline model, this application studies the domain differences of features learned by our model and the baseline model.

[0223] We first conduct a quantitative analysis, drawing on previous studies, by computing the proxy A-distance between the feature distributions of the source and target domains. We compare our model with three baseline models: RoBERTa, DeepCORAL, and Mixup. In particular, when computing the AA-distance, we sample the same number of samples from each source domain and ensure that the total number of samples in the source domain matches the number of samples in the target domain. Figure 5 As shown in Figure 2, on most datasets, RoBERTa's AA-distance is significantly higher than other models, with an average of 1.66. This may be because traditional models rely on empirical risk minimization and fail to effectively reduce the distribution shift between samples from different fields.

[0224] In contrast, our model MIUFD significantly reduces the A-distance by 0.7 on average. This improvement is mainly attributed to the domain-invariant feature encoder in our model, which effectively narrows the domain gap.

[0225] To further verify the effectiveness of the model proposed in this application, we conducted a qualitative analysis by visualizing the feature representation using t-SNE. In this experiment, we trained the model on source domains including Climate, Sci-Fact, and PubHealth, and used FEVER as the target domain. The visualization results are shown in Figure 2. Figure 6 shown.

[0226] from Figure 6 As can be seen, the baseline models DeepCORAL and Mixup (which are designed for domain generalization) produce more uniform feature representations than the traditional RoBERTa model. In RoBERTa, features are clearly clustered across domains, indicating large differences between the domains. In contrast, the features generated by our model have a smaller distance between the target and source domains, demonstrating a better ability to reduce the difference in feature distribution between domains than DeepCORAL and Mixup. This advantage is attributed not only to the domain-invariant encoder, but also to the domain-specific encoder, which utilizes an expert network to map target domain features to the relevant source domain.

[0227] Furthermore, our model reduces the distance between samples of the same type and enhances the separation between samples of different categories, where data points of the same shape are clustered together. This advantage further supports our hypothesis that the label-based contrast mechanism enables the model to make more accurate classification decisions by establishing clearer feature boundaries.

[0228] To answer the question of how hyperparameters affect the performance of MIUDF?, we evaluate the performance of our proposed model and the baseline model by reducing the number of source domains and examples for training. First, we evaluate the robustness of the model to the number of source domains. We train the model using two source domains and report the macro F1 scores on two target datasets (FEVER and Climate-FEVER). The results are shown in Figure 2. Figure 7 As shown in Figure 3, the average performance of MIUFD and the baseline model on the target dataset drops significantly compared to the model trained using three source domains. This drop in performance suggests that more source domains help the model extract more generalizable features. In addition, learning from multiple domain samples helps alleviate the bias introduced by some source domain data that have large distribution differences from the target domain. When comparing the performance of our model with the baseline model, we found that MIUFD consistently outperformed the baseline model on both target datasets under different source domain settings. This advantage may be due to our model's ability to integrate domain-invariant features and domain-specific features, thereby reducing the negative impact when there are fewer source domains. In addition, the fine-grained reasoning unit in the model can more accurately capture the relationship between evidence and arguments, thereby generating features that contribute to the final judgment prediction.

[0229] Next, we tested the robustness of the model to the number of source domain samples. We tested the performance of our model and the baseline model on two target datasets (Sci-Fact and PubHealth) by changing the data ratio (5%, 10%, 20%, 40%, 80%). The results are shown in Figure 2. Figure 8 As shown in , the performance of most models gradually improves with the increase of the number of examples. This is consistent with our expectation that more training data can improve the generalization ability of the model. When comparing the performance of our model with the baseline model, we found that our model consistently outperforms the baseline model under different target datasets and data ratios. For example, Figure 8 As shown in Figure 2, on the Sci-Fact dataset, when the data ratio is 5%, our method improves the F1 score by 13.14% and 14.53% compared to PADA and Deepcoral, respectively. These improvements further emphasize the robustness of our model when processing different amounts of training data, which is attributed to the model's ability to combine diverse and discriminative feature representations for label classification.

[0230] A cross-domain statement truth judgment system, comprising:

[0231] A first acquisition module is used to acquire a training statement and evidence preset in a database, wherein the training statement and evidence come from different fields;

[0232] an alignment module, configured to divide the training statement into a plurality of training statement blocks, align the training statement blocks with evidence to obtain block alignment pairs, and align the training statement with evidence to obtain argument-evidence pairs;

[0233] a first feature conversion module, configured to input the block alignment pairs and argument-evidence pairs into an encoder respectively to obtain low-level block alignment pair features and low-level argument-evidence pair features, project the low-level block alignment pair features and the low-level argument-evidence pair features to obtain training high-level features, the training high-level features including block alignment pair high-level features and argument-evidence pair high-level features, and combine a plurality of block alignment pair high-level features in the same domain into a training feature matrix of argument-evidence pairs;

[0234] The second feature conversion module is used to obtain the expert network, which processes the low-level block alignment features and low-level argument evidence features through the expert network to obtain the domain-specific features of the expert network in different fields;

[0235] A second acquisition module is used to acquire a gated network;

[0236] a third acquisition module, configured to obtain a first weight and a second weight according to the low-level block alignment pair features, the low-level argument evidence pair features, and the gating network;

[0237] a first calculation module, configured to obtain a training aggregate feature based on the first weight, the second weight, and domain-specific features of expert networks in different fields;

[0238] A second calculation module is used to obtain a training coupling feature based on the training high-level features and the training aggregate features;

[0239] A third calculation module is used to input the training coupling features into a classifier to obtain a training statement label category probability;

[0240] A second training module is configured to obtain a classifier loss based on the training declared label category probabilities and the true categories, and adjust the classifier weights based on the classifier loss to obtain an optimized classifier;

[0241] a first training module, configured to calculate an encoder loss based on the training feature matrix, adjust an encoder weight based on the encoder loss to obtain an optimized encoder, input actual claims and evidence into the optimized encoder to obtain actual low-level block alignment pair features, actual low-level argument-evidence pair features, and actual high-level features;

[0242] The output module is used to obtain the actual expert network domain-specific features based on the actual low-level block alignment features and the actual low-level argument evidence pair features, obtain the actual aggregate features based on the actual expert network domain-specific features, the first weight and the second weight, obtain the actual coupling features based on the actual aggregate features and the actual high-level features, input the actual coupling features into the optimized classifier, obtain the declared label category probability, and select the largest label category probability as the judgment result.

[0243] An embodiment of the present application also discloses a terminal device, including a memory and a processor. The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, a cross-domain declaration truth judgment method is adopted.

[0244] Among them, the terminal device can be a computer device such as a desktop computer, a laptop computer or a cloud server, and the terminal device includes but is not limited to a processor and a memory. For example, the terminal device can also include input and output devices, network access devices and buses, etc.

[0245] Among them, the processor can adopt a central processing unit (CPU). Of course, according to actual usage, other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. can also be adopted. The general-purpose processor can adopt a microprocessor or any conventional processor, etc., and this application does not impose any restrictions on this.

[0246] Among them, the memory can be an internal storage unit of the terminal device, such as the hard disk or memory of the terminal device, or it can be an external storage device of the terminal device, such as a plug-in hard disk, smart memory card (SMC), secure digital card (SD) or flash memory card (FC) equipped on the terminal device, etc., and the memory can also be a combination of the internal storage unit and the external storage device of the terminal device. The memory is used to store computer programs and other programs and data required by the terminal device. The memory can also be used to temporarily store data that has been output or is to be output. This application does not impose any restrictions on this.

[0247] Among them, through this terminal device, the method for judging the authenticity of the cross-domain declaration in the above embodiment is stored in the memory of the terminal device, and is loaded and executed on the processor of the terminal device for easy use.

[0248] An embodiment of the present application further discloses a computer-readable storage medium, and the computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, the method for determining the truth or falsehood of a cross-domain statement in the above embodiment is adopted.

[0249] Among them, the computer program can be stored in a computer-readable medium, the computer program includes computer program code, the computer program code can be in the form of source code, object code, executable file or certain middleware, etc. The computer-readable medium includes any entity or device that can carry computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signal, telecommunication signal and software distribution medium, etc. It should be noted that computer-readable medium includes but is not limited to the above-mentioned components.

[0250] Among them, through this computer-readable storage medium, the method for judging the truth or falsehood of the cross-domain statement in the above embodiment is stored in the computer-readable storage medium, and is loaded and executed on the processor to facilitate the storage and application of the above method.

[0251] Those skilled in the art should understand that the discussion of any of the above embodiments is merely illustrative and is not intended to imply that the scope of protection of the present application is limited to these examples. In line with the present application, the technical features in the above embodiments or different embodiments may be combined, the steps may be implemented in any order, and there are many other variations of different aspects of one or more embodiments of the present application as described above, which are not provided in detail for the sake of simplicity.

[0252] The one or more embodiments of this application are intended to encompass all such substitutions, modifications, and variations that fall within the broad scope of this application. Therefore, any omissions, modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of one or more embodiments of this application should be included in the scope of protection of this application. < / chunk-1>

Claims

1. A method for determining the truth or falsity of a cross-domain statement, characterized by: include: Obtaining training statements and evidence pre-set in a database, the training statements and evidence being from different fields; Dividing the training statement into a plurality of training statement blocks, aligning the training statement blocks with evidence to obtain block alignment pairs, and aligning the training statement with evidence to obtain argument-evidence pairs; Inputting the block alignment pairs and argument-evidence pairs into an encoder respectively to obtain low-level block alignment pair features and low-level argument-evidence pair features, projecting the low-level block alignment pair features and the low-level argument-evidence pair features to obtain training high-level features, the training high-level features including block alignment pair high-level features and argument-evidence pair high-level features, and combining several block alignment pair high-level features in the same domain into a training feature matrix of argument-evidence pairs; Obtain an expert network, process the low-level block alignment features and low-level argument evidence features through the expert network, and obtain domain-specific features of expert networks in different fields; Get the gated network; Obtaining a first weight and a second weight according to the low-level block alignment pair features, the low-level argument evidence pair features, and the gating network; Obtaining training aggregate features based on the first weight, the second weight, and domain-specific features of expert networks in different fields; Obtaining a training coupled feature based on the training high-level feature and the training aggregate feature; Inputting the training coupled features into a classifier to obtain training statement label category probabilities; Obtaining a classifier loss based on the training declared label category probabilities and the true categories, and adjusting the classifier weights based on the classifier loss to obtain an optimized classifier; According to the training feature matrix, an encoder loss is calculated, an encoder weight is adjusted according to the encoder loss to obtain an optimized encoder, and actual claims and evidence are input into the optimized encoder to obtain actual low-level block alignment pair features, actual low-level argument-evidence pair features, and actual high-level features; The actual expert network domain-specific features are obtained according to the actual low-level block alignment pair features and the actual low-level argument evidence pair features. The actual aggregation features are obtained according to the actual expert network domain-specific features, the first weight and the second weight. The actual coupling features are obtained according to the actual aggregation features and the actual high-level features. The actual coupling features are input into the optimized classifier to obtain the declared label category probability, and the maximum label category probability is selected as the judgment result.

2. The method for determining the authenticity of a cross-domain statement as claimed in claim 1, wherein: The block alignment pairs and argument evidence pairs are input into the encoder respectively to obtain low-level block alignment pair features and low-level argument evidence pair features, and the low-level block alignment pair features and low-level argument evidence pair features are projected to obtain training high-level features including: The block alignment pairs and argument evidence pairs are input into the encoder respectively to obtain low-level block alignment pair features and low-level argument evidence pair features, which are expressed as: X ce ,X al =Encoder(x ce ,x al ); x al =Concat({a1,…,a l }); Among them, Encoder is the encoder, Concat means splicing, a l For the training statement block, x al is a block alignment pair, x ce For the argument evidence pair, X ce is the low-level block alignment feature, X al Evidence for low-level arguments features; And project the low-level block alignment features and low-level argument evidence features to obtain the training high-level features, which are expressed as: z ce ,z al =f p (X ce ,X al ); Among them, z ce ,z al To train high-level features, f p Represents a projection.

3. The method for determining the authenticity of a cross-domain statement as claimed in claim 1, wherein: The calculating the encoder loss according to the training feature matrix, and adjusting the encoder weight according to the encoder loss to obtain the optimized encoder includes: The covariance matrix is ​​calculated based on the training feature matrix, and the calculation method is: in, is the covariance matrix of argument-evidence pairs in domain i, Is from the field The characteristic matrix of the argument evidence pair, T is the transposed moment, n i is the number of samples in domain i, 1 represents a column vector whose elements are all 1; According to the covariance matrix of different fields, the first alignment loss of the argument-evidence pair is calculated and expressed as: in, represents the Frobenius norm of the matrix, is the alignment loss, is the covariance matrix of argument-evidence pairs in domain j, K is the number of domains; Calculate the second alignment loss for the block alignment pair; According to the high-level features of different fields, the first contrast loss of argument evidence to high-level features is calculated and expressed as: Among them, A(i)≡I\{i} represents the set of all samples except sample i, P(i)≡{p∈A(i):y p =y i } represents the index set of all positive samples relative to sample i, |P(i)| represents the number of positive samples, the symbol represents the inner product, τ is a scalar parameter, is the argument evidence for the positive sample to the high-level features, is the argument evidence pair high-level features of sample i, is the feature of the ath sample in the sample set A(i); Calculate the second contrast loss of block alignment on high-level features; Summing the first alignment loss, the second alignment loss, the first contrast loss, and the second contrast loss to obtain an encoder loss; The encoder weights are adjusted according to the encoder loss to obtain an optimized encoder.

4. The method for determining the authenticity of a cross-domain statement as claimed in claim 1, wherein: By processing the low-level block alignment features and low-level argument evidence features through the expert network, the domain-specific features of the expert network in different fields are expressed as: in, is the domain-specific feature of the expert network generated by the k-th expert network, x al is a block alignment pair, x ce For the argument evidence pair, is an expert network, and K represents the number of fields.

5. The method for determining the authenticity of a cross-domain statement as claimed in claim 1, wherein: The acquisition gating network includes: Get the preset gating network; According to the low-level block alignment pair features, the low-level argument evidence pair features and the preset gating network, a first training weight and a second training weight are obtained, which are expressed as: w ce ,w al =Softmax(g a (x ce ,x al )); Among them, w ce is the first training weight, w al is the second training weight, g a is the preset gating network, x al is a block alignment pair, x ce Evidence for the argument; Obtaining a weight loss value according to the first training weight, the second training weight, and a weight loss formula; The weight loss formula is: in, is the weight loss value, K is the number of fields, d i represents the domain label of the i-th sample, I is the indicator function, and denotes the weights assigned to the kth domain expert for the claim-evidence pair and the block alignment pair, respectively, is the first weight of the k-th domain expert, is the second weight of the k-th domain expert; The network weights of the preset gating network are adjusted according to the weight loss value to obtain the gating network.

6. The method for determining the authenticity of a cross-domain statement as claimed in claim 1, wherein: According to the actual aggregation features and the actual high-level features, the actual coupling features are expressed as: x h =f h (Concat(z ce ,z al ,h ce ,h al ,Attention(z ce ,z al ,h ce ,h al ))); Among them, Concat means splicing, z ce ,z al represents the actual high-level features, h ce ,h al represents the actual aggregated features, attention represents the self-attention operation applied to features from different channels, and f h (·) is the feature aggregation function based on the multi-layer perceptron MLP, x h is the actual coupling characteristic.

7. The method for determining the authenticity of a cross-domain statement as claimed in claim 1, wherein: The coupling feature is input into the classifier to obtain the training statement label category probability expression as: P(y)=Softmax(f c (ReLU(x h ))) Among them, f c (·) represents the classifier function, ReLU represents the nonlinear activation function, and x h is the actual coupling characteristic.

8. A system for determining the truth or falsity of cross-domain statements, characterized by: include: A first acquisition module is used to acquire a training statement and evidence preset in a database, wherein the training statement and evidence come from different fields; an alignment module, configured to divide the training statement into a plurality of training statement blocks, align the training statement blocks with evidence to obtain block alignment pairs, and align the training statement with evidence to obtain argument-evidence pairs; a first feature conversion module, configured to input the block alignment pairs and argument-evidence pairs into an encoder respectively to obtain low-level block alignment pair features and low-level argument-evidence pair features, project the low-level block alignment pair features and the low-level argument-evidence pair features to obtain training high-level features, wherein the training high-level features include block alignment pair high-level features and argument-evidence pair high-level features, and combine multiple block alignment pair high-level features in the same domain into a training feature matrix of argument-evidence pairs; The second feature conversion module is used to obtain the expert network, which processes the low-level block alignment features and low-level argument evidence features through the expert network to obtain the domain-specific features of the expert network in different fields; A second acquisition module is used to acquire a gated network; a third acquisition module, configured to obtain a first weight and a second weight according to the low-level block alignment pair features, the low-level argument evidence pair features, and the gating network; a first calculation module, configured to obtain a training aggregate feature based on the first weight, the second weight, and domain-specific features of expert networks in different fields; A second calculation module is used to obtain a training coupling feature based on the training high-level features and the training aggregate features; A third calculation module is used to input the training coupling features into a classifier to obtain the training statement label category probability; A second training module is configured to obtain a classifier loss based on the training declared label category probabilities and the true categories, and adjust the classifier weights based on the classifier loss to obtain an optimized classifier; a first training module, configured to calculate an encoder loss based on the training feature matrix, adjust an encoder weight based on the encoder loss to obtain an optimized encoder, input actual claims and evidence into the optimized encoder to obtain actual low-level block alignment pair features, actual low-level argument-evidence pair features, and actual high-level features; The output module is used to obtain the actual expert network domain-specific features based on the actual low-level block alignment features and the actual low-level argument evidence pair features, obtain the actual aggregate features based on the actual expert network domain-specific features, the first weight and the second weight, obtain the actual coupling features based on the actual aggregate features and the actual high-level features, input the actual coupling features into the optimized classifier, obtain the declared label category probability, and select the largest label category probability as the judgment result.

9. A terminal device comprising a memory and a processor, characterized in that: The memory stores a computer program that can be run on the processor. When the processor loads and executes the computer program, the method according to any one of claims 1 to 7 is adopted.

10. A computer-readable storage medium storing a computer program, wherein: When the computer program is loaded and executed by a processor, the method according to any one of claims 1 to 7 is adopted.

Images