A few-shot supervised Chinese fact-checking system enhanced by rumor detection data

Abstract

The application discloses a kind of few-sample supervised Chinese fact verification systems of rumor detection data enhancement, utilize the knowledge transfer of cross-task data enhancement, obtain external supervision signal from false news detection task, assist in executing Chinese fact verification task.Specifically, the system includes: cross-task prompt cascade module, aims to given Chinese rumor check task input and the input of Chinese fact verification task is processed into the semantic representation sequence of form uniformity;Efficient parameter fine-tuning module based on language model: based on the semantic representation sequence described above, obtain language model sequence probability distribution through the module;Label decoding module: design the two-step strategy of label retrieval based on mapping table query and label retrieval based on probability sequence sorting, decode language model sequence probability distribution, combine the constructed cross-task label mapping rule table and pre-defined target sequence, query and match out the final Chinese fact verification task label.

Description

Technical Field

[0001] This invention relates to the field of supervised Chinese fact-checking with few samples, and specifically discloses a supervised Chinese fact-checking system with few samples enhanced by rumor detection data. Background Technology

[0002] With the booming development of the internet, information has exploded online. However, the quality of internet information sources varies greatly, and much of the information is false, prompting governments, society, and researchers in related fields to pay close attention to the authenticity of information. Rumor detection tasks have emerged and achieved significant results. However, some rumors cannot be determined by algorithms alone and require concrete evidence. With the rise of artificial intelligence technology, to address these issues, deep learning-based fact-checking tasks aim to determine the veracity of rumors based on potential evidence, expanding applicable scenarios and making internet data authenticity detection results more reliable, attracting significant attention from industry and academia.

[0003] Currently, fact-checking research based on deep learning mainly relies on fully supervised learning, which presents several challenges: it requires large-scale labeled datasets, but labeling is time-consuming and labor-intensive, and it is difficult to guarantee high-quality labeling and accurate consistency; performance is highly dependent on the quality of labeled data, which directly affects model performance. If the labeling is incorrect or inaccurate, the model will learn incorrect patterns or produce misleading results; and the model has poor domain transferability, meaning that while the model may perform well on a specific dataset, its performance may deteriorate on datasets with different types of rumors, requiring relabeling and retraining, which increases the cost and complexity of research and application. Summary of the Invention

[0004] The purpose of this invention is to provide a few-sample supervised Chinese fact-checking system enhanced with rumor detection data. Compared with the previous mainstream Chinese fact-checking systems, it saves time and effort in data annotation, solves the current problem of insufficient data annotation, and gets rid of the dilemma of the model over-reliance on high-quality supervised data for fact-checking.

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

[0006] This invention provides a few-sample supervised Chinese fact-checking system enhanced with rumor detection data, comprising the following main modules: a cross-task prompting cascading module, a language model-based efficient parameter fine-tuning module, and a tag decoding module, as detailed in the attached specification. Figure 1 The details are as follows:

[0007] The cross-task hint cascading module includes task breakdown prefixes and hint combination templates. See the attached manual for details. Figure 2 ;

[0008] This module is an efficient parameter fine-tuning module based on a language model. It includes a pre-trained language model with fixed parameters and a small-scale cascaded optimizable parameter network. (See the attached instruction manual.) Figure 4 ;

[0009] The tag decoding module employs a pipeline-style two-step strategy: tag retrieval based on a mapping table lookup and tag retrieval based on probability sequence sorting. It contains a set of predefined target sequences corresponding to intermediate answers. (N-2, the number of original rumor detection task label categories) and a label mapping rule table across rumor checking and fact-checking tasks;

[0010] Based on the above module structure, the technical solution adopted in this invention and the function of each module are as follows:

[0011] Cross-task prompt cascade module: It is used to merge the inputs of the given Chinese rumor checking task and Chinese fact checking task into a unified semantic representation sequence through pipeline processing of task decomposition prefix and prompt joint template, and serve as the input of the pre-trained language model in the efficient parameter fine-tuning module based on language model;

[0012] The efficient parameter fine-tuning module based on the language model encodes and decodes the semantic representation sequence described above using the pre-trained language model in this module, outputting a probability distribution prediction of the language model sequence. The pre-trained language model parameters are fixed in this module. During the learning process, the output of the language model, i.e., the probability distribution of the language model sequence, is optimized and adjusted by optimizing a small-scale extended optimizable parameter network, thereby improving the prediction accuracy of the final Chinese fact-checking tags.

[0013] The label decoding module further decodes the output of the efficient parameter fine-tuning module, namely the predicted language model sequence probability distribution. Specifically, it uses a pipeline-style two-step strategy: label retrieval based on mapping table lookup and label retrieval based on probability sequence sorting. Combining the designed cross-task label mapping rule table and predefined target sequences, it compares the intermediate answer with the highest fit to the language model probability sequence with the examples in the mapping rule table and finally determines the Chinese fact-checking label.

[0014] Preferably, the two tasks referred to in the aforementioned few-sample supervised Chinese fact-checking system enhanced with rumor detection data—Chinese rumor detection and Chinese fact-checking—are defined as follows:

[0015] Chinese Rumor Detection: A binary classification task in the field of natural language processing, with category labels of rumor (FAKE) and true non-rumor (REAL). The task is to determine whether a given text contains false or misleading information.

[0016] Chinese Fact Check: A three-classification task in the field of natural language processing, with label categories of support, refute, and not enough information. Given the typical input of the Chinese fact check task: the original text pair {Claim, Evidence} serving as the fact statement and the premise evidence, the task is to determine the relationship between the two in terms of label categories.

[0017] Furthermore, the aforementioned few-sample supervised Chinese fact-checking system enhanced with rumor detection data is divided into a model training phase and an inference phase, and the operation process of each of these phases differs to some extent.

[0018] Preferably, in the aforementioned few-sample supervised Chinese fact-checking system enhanced with rumor detection data, the cross-task prompt cascading module constructs a semantic representation sequence with a unified form (corresponding to the appendix to the specification). Figure 3 The specific process is as follows:

[0019] First, design three fixed task decomposition prefixes: those that support the "fact statement", those that deny the "fact statement", and those that are unclear about the "fact statement", such as "We can confirm / we can deny / we cannot confirm or deny [fact statement]".

[0020] Furthermore, by concatenating the task decomposition prefix with the factual statement portion of the Chinese rumor checking task input, three corresponding cross-task semantic composite expressions are obtained.

[0021] Furthermore, a set of prompt joint templates is designed to perform secondary processing on the three cross-task semantic composite expressions mentioned above to obtain the sentence patterns of the original binary semantics for the Chinese rumor detection task, which can then be used as input for the subsequent efficient parameter fine-tuning module based on the language model.

[0022] Furthermore, a prompt joint template is randomly selected from the prompt joint template set designed above. Through the slots set in the template, it is combined with the three constructed cross-task semantic composite representations to obtain a semantic representation sequence with a unified form, which serves as the input to the pre-trained language model in the subsequent efficient parameter fine-tuning module.

[0023] Preferably, the efficient parameter fine-tuning module based on the language model includes the following steps during the model training phase:

[0024] First, through a cross-task prompting cascade module, the raw training data for Chinese rumor detection and Chinese fact-checking supervision are simultaneously processed through a pipeline using task decomposition prefixes and prompting joint templates to construct a semantic representation sequence with a unified form. Then, the supervised training data processed from Chinese rumor detection and Chinese fact-checking are randomly combined and used as input to the pre-trained language model in the efficient parameter fine-tuning module based on the language model.

[0025] Furthermore, the semantic representation sequence constructed above is input into the pre-trained language model of the efficient parameter fine-tuning module to perform vector embedding on the character sequence input with natural language as the carrier: the vocabulary corresponding to the pre-trained language model is mapped to a one-hot vector, and low-dimensional embedding is performed through the vocabulary matrix;

[0026] Furthermore, the low-dimensional embedding vector is passed through the pre-trained language model's encoder-decoder layers and extended optimizable parameter network, and then through the forward propagation of the multi-layer Transformer Block module (multi-head self-attention mechanism in the Block module) and the optimizable parameter network. After the above encoding and decoding process, the semantic vector representation of the language model after encoding and decoding operations is obtained. The physical meaning of the output semantic vector is: the probability distribution prediction of the natural language model sequence output by the language model.

[0027] Furthermore, based on the probability distribution of the output language model sequence and combined with the predefined target sequence, the parameters of the optimizable network are updated by calculating the backpropagation loss value of the three defined loss functions: conditional language model loss, negative sample penalty loss, and sentence length normalization loss, thereby completing the training phase of efficient parameter fine-tuning based on the language model.

[0028] The above describes the training phase of the efficient parameter fine-tuning module. It does not involve the tag decoding module, as this training process aims to train intermediate predictions for a unified semantic representation across tasks, rather than predictions for the final Chinese fact-checking tags.

[0029] Preferably, the efficient parameter fine-tuning module includes the following steps in the prediction inference stage:

[0030] First, a unified semantic representation sequence is constructed through a cross-task prompt cascade module. The semantic representation sequence data of the constructed Chinese fact-checking task is then used as the input of the pre-trained language model in the efficient parameter fine-tuning module based on the language model.

[0031] Furthermore, by using this efficient parameter fine-tuning module based on the language model for inference, the probability distribution of the language model sequence is predicted. At this time, only the semantic vector representation after the language model encoding and decoding operations is calculated, that is, the probability distribution of the language model predicted sequence, and the loss calculation process and the backpropagation parameter update steps are discarded.

[0032] Furthermore, the final Chinese fact-checking task labels are predicted through the label decoding module. Specifically, the semantic vector output representation obtained from the language model-based efficient parameter fine-tuning module, i.e., the output language model sequence probability distribution, is used to calculate the cross-entropy value between the semantic vector and the predefined target sequence, and the target sequence with the smallest cross-entropy is selected as the intermediate answer. Subsequently, the intermediate answers under different task decomposition prefixes are combined and mapped to the matching row in the label mapping rule table, ultimately indexing the Chinese fact-checking task labels.

[0033] Preferably, in the efficient parameter fine-tuning module based on the language model, the pre-trained language model uses the T0-3B Chinese generative language model based on the Transformer architecture, and its operation flow is as follows:

[0034] First, the character sequence input using natural language as the carrier is vector-embedded. The resulting low-dimensional embedding vector is then input into the multi-layer Transformer Block of the generative language model encoder and decoder. Each independent Transformer Block contains the following sub-networks: Multi-head Self-Attention Layer (MHSA), Dropout, Feedforward Fully Connected Layer (FFN), and Residual and Layer Normalization Layer (Add&Norm).

[0035] Furthermore, positional information is encoded using the sine function sin(·) and the cosine function cos(·) to obtain a positional feature vector. The low-dimensional embedding vector is then summed bit by bit with the positional feature vector, as shown in the following formula:

[0036]

[0037]

[0038] Where i represents half of the character subscript, PE (pos,2i) and PE (pos,2i) d represents the position vector of the character with even and odd indices, respectively. model Refers to the hidden layer dimension of the Transformer block;

[0039] Furthermore, a vector output with the same dimension is obtained through a multi-head self-attention layer. Specifically, an optimizable matrix of Q, K, and V is constructed, and the scaled dot product attention Attention(Q,K,V) of each attention head is calculated, as shown in the following formula.

[0040]

[0041] Here, softmax(·) represents the "soft maximization" function;

[0042] Furthermore, multiple single-attention head vectors are concatenated to obtain a multi-head self-attention representation, as shown in the following formula;

[0043] MultiHead(Q,K,V)=Concat(head1,...,head h W O

[0044]

[0045] in

[0046] Furthermore, semantic embedding representations are obtained through a random deactivated network and a feedforward fully connected layer, and then processed through a residual network and layer normalization operations. The residual module avoids the gradient explosion problem, and the layer normalization operation accelerates model convergence by controlling the mean and variance, as shown in the following formula;

[0047] LayerNorm(X+MultiHeadAttention(X))

[0048] LayerNorm(X+FeedForward(X))

[0049] Where LayerNorm(·) refers to the layer normalization operation; (X+MultiHeadAttention(X)) and (X+FeedForward(X)) represent the residual connection networks after passing through the multi-head self-attention network and the feedforward network, respectively;

[0050] Through the above process, involving encoding and decoding through multiple layers of Transformer Blocks, the probability distribution representation D of the output language model sequence is obtained. LM .

[0051] Preferably, for the external small-scale optimizable parameter network in the efficient parameter fine-tuning module based on the language model, by performing external network parameter operations on the feed-forward layers in each multi-head self-attention layer of all Transformer Blocks in the pre-trained language model, excellent learning ability is achieved while keeping the parameters of the pre-trained language model fixed. The operation process is as follows. For the corresponding flowchart, please refer to the appendix of the specification Figure 4 :

[0052] First, initialize three trainable weight matrices Among them, the hyperparameter r is the rank of the low-rank matrix ≤ min(d, k), where d and k correspond to the input and output dimensions of the attached feed-forward layer;

[0053] Furthermore, given the input vector X and the parameters W0 of the pre-attached feed-forward layer, the original feed-forward calculation h0 = W0X is updated to the low-rank modified feed-forward calculation h0 = W0X + (ABC)X;

[0054] During training, by fixing the parameters of the attached feed-forward layer of the pre-trained language model, it is transformed into optimizing the parameters ABC of the external small-scale optimizable parameter network, thereby achieving the learning ability of the language model;

[0055] This efficient parameter fine-tuning module greatly reduces the number of parameters to be trained and provides excellent model fitting ability. During the deployment phase, the trained parameter network can be externally connected to the original language model at any time without additional memory and inference latency.

[0056] Furthermore, the efficient parameter fine-tuning module based on the language model has three types of loss functions: conditional language model loss L LM , negative sample penalty loss L UK , and sentence length normalization loss L LN , and their calculation processes are as follows:

[0057]

[0058] Among them, p(y t |X, y<t) refers to the conditional probability of generating the t-th character given the input X and the characters before the t-th character are known. T refers to the character length of the decoded sequence. The conditional language model loss L LM aims to maximize the fitting degree between the sequence probability distribution of the language model and the correct model target sequence;

[0059]

[0060] Among them, <Let L represent the conditional probability of the t-th character in a negative sample given the input x and the first t-1 characters. The denominator of the above formula represents the sum of the conditional probabilities of all incorrect characters in the negative samples. The negative sample penalty loss L is calculated by averaging the total length of all incorrect characters. UK The aim is to reduce the fit between the probability distribution of the model's output sequence and the incorrect model target sequence;

[0061]

[0062] in, It is the conditional language model loss L LM The reciprocal of the sentence length standardization loss L LN The aim is to reduce the impact of inconsistent output sequence lengths on model loss calculations and to control the length of the output sequence probability distribution to be close to the correct target sequence length.

[0063] L total =L LM +L UK +L LN Furthermore, the total loss of the model is the aforementioned loss L. LM L UK L LN The model is optimized by simply summing the values ​​and then calculating the gradient of the circumscribed parameter network through backpropagation.

[0064] Preferably, the tag decoding module has the following characteristics:

[0065] The input to the tag decoding module is the output of the efficient parameter fine-tuning module mentioned above: the probability distribution prediction of the language model sequence;

[0066] The output of the tag decoding module is the final classification tag for the Chinese fact-checking task.

[0067] The operation process of the tag decoding module is as follows:

[0068] Based on three cross-task semantic composite representations using three task decomposition prefixes, the efficient parameter fine-tuning module outputs the language probability distribution sequence D. LM Given the probability distribution of three output language sequences. The tag decoding module's workflow consists of a pipelined two-step strategy: tag retrieval based on mapping table lookup and tag retrieval based on probability sequence sorting, as detailed below:

[0069] The operation process of tag retrieval based on mapping table lookup is described in the attached instruction manual. Figure 6 The details are as follows:

[0070] Calculate the probability distribution of the language sequence for each output. With predefined target sequence The cross-entropy between (N=2, the number of categories in the Chinese rumor detection task) yields a character-level cross-entropy set;

[0071] By normalizing the sentence length and averaging the sets, we can obtain the probability distribution of each output language sequence. Corresponding average cross-entropy Logit a The formula is as follows:

[0072]

[0073] in, This refers to the entropy value between the above-mentioned language sequence probability distribution and each character in the predefined target sequence; the entire expression represents the mean cross-entropy at the target sequence level.

[0074] Furthermore, for each predicted language sequence probability distribution Select the predefined target sequence a with the smallest average cross-entropy calculation result. k As an intermediate answer;

[0075] Based on the intermediate answer combinations of three predefined target sequences corresponding to three cross-task semantic composite representations, a matching query is performed from a pre-defined mapping table (the full name of the mapping table is the Cross-Task Tag Mapping Rule Table, see attached manual). Figure 5 This yields the final label for the Chinese fact-checking task.

[0076] The operation process of the tag retrieval based on probability sequence sorting is as follows:

[0077] Since the mapping table of the tag retrieval method based on mapping table query does not completely include all combinations, a retrieval method based on probability sequence sorting is used for cases that do not appear in the preset mapping table.

[0078] First, we obtain the corresponding intermediate answer arrangement using the target labels of the pre-defined mapping table;

[0079] Then, the cross-entropy between each intermediate answer and the corresponding predefined target sequence of all intermediate answer permutations is calculated, and the sums are averaged to obtain the average cross-entropy for each label;

[0080] Finally, the label with the smallest average cross-entropy is selected as the final label for the Chinese fact-checking task by sorting by cross-entropy.

[0081] The present invention provides a few-sample supervised Chinese fact-checking system enhanced with rumor detection data, the main advantages and beneficial effects of which are reflected in the following aspects:

[0082] 1. Reduce annotation costs and improve efficiency: Traditional fact-checking requires a large amount of manually annotated data, which is not only time-consuming and labor-intensive, but also makes it difficult to ensure the consistency and accuracy of annotation. This invention uses data from rumor detection tasks as an enhancement method, reducing reliance on large-scale annotated data and lowering annotation costs and labor intensity;

[0083] 2. Enhancing the model's generality and transferability: In existing methods, model performance is often highly dependent on specific labeled datasets, resulting in poor performance on datasets with different rumor types, requiring re-labeling and retraining. This invention improves the model's transferability and generality by introducing external rumor detection data and enhancing it with external knowledge.

[0084] 3. Improved performance in few-shot learning: Traditional methods may struggle to achieve ideal performance when labeled data is limited. This invention enhances the model's learning ability in few-shot environments by incorporating rumor detection data, enabling it to maintain good performance even in scenarios with limited data resources.

[0085] 4. Exploring new methods for enhancing external knowledge: This invention not only innovates in model architecture and data augmentation, but also explores methods for enhancing fact-checking by utilizing external knowledge from heterogeneous tasks, providing new perspectives and methodologies for future research.

[0086] This invention provides a solution based on external knowledge enhancement for heterogeneous tasks, especially a method for enhancing Chinese rumor detection data, which is similar to Chinese fact-checking tasks. This is a research topic with great practical significance and represents an important development opportunity for the future research field of Chinese rumor detection.

[0087] Of course, product instances implementing this invention in any scenario do not necessarily need to achieve all the advantages and benefits described above at the same time. Attached Figure Description

[0088] Figure 1 This is a structural block diagram of the present invention;

[0089] Figure 2 A schematic diagram of the task decomposition prefix and prompt joint template in the cross-task prompt cascading module of the system;

[0090] Figure 3 A schematic diagram illustrating the construction process of three cross-task semantic composite expressions in the cross-task prompt cascading module of the system;

[0091] Figure 4 This is a schematic diagram of the small-scale external optimizable parameter network and the feedforward network layer dependency architecture in the pre-trained language model in the efficient parameter fine-tuning module based on the language model of the system.

[0092] Figure 5 This refers to the cross-task label mapping rule table in the label decoding module of the system.

[0093] Figure 6 This is a schematic diagram illustrating the tag decoding process of the system's tag decoding module for the final Chinese fact-checking task. Detailed Implementation

[0094] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the embodiments described below are merely some embodiments of the present invention, and not embodiments compatible with all application scenarios. Referring to the embodiments demonstrated in this invention, all other embodiments obtained by those skilled in the art without any inventive step are also within the protection scope of this invention.

[0095] This invention provides a few-sample supervised Chinese fact-checking system enhanced with rumor detection data, as shown in the appendix to the specification. Figure 1 As shown, it includes the following modules:

[0096] 1. The cross-task prompt cascade module takes the input of the Chinese rumor checking task and the Chinese fact-checking task as input, and merges them into a semantic representation sequence with a unified form through the pipeline processing of the above task decomposition prefix and prompt joint template. This sequence serves as the output of the cross-task prompt cascade module and also as the input of the efficient parameter fine-tuning module.

[0097] 2. An efficient parameter fine-tuning module based on a language model, which includes a pre-trained language model with fixed parameters and a small-scale extended optimizable parameter network, aims to predict the probability distribution of the corresponding language model sequence for the above-mentioned unified semantic representation sequence;

[0098] 3. The label decoding module aims to further decode the probability distribution of the predicted sequences from the language model. This module includes a cross-task label mapping rule table and predefined target sequences. Specifically, through a pipeline-style two-step strategy—label retrieval based on the mapping table and label retrieval based on probability sequence sorting—the intermediate answer with the highest fit to the probability sequence of the language model is retrieved, thereby determining the final label for the Chinese fact-checking task.

[0099] First, given a Chinese fact-checking task as input to the system, the cross-task prompting cascade module performs pipelined processing based on task decomposition prefixes and prompt joint templates. (Refer to the attached manual for details.) Figure 2 .

[0100] Input: The original text pair {Claim, Evidence} serving as the factual statement and the premise evidence;

[0101] Output: A transformed, uniformly semantic representation sequence. This module contains three task decomposition prefixes and nine cue joint templates.

[0102] Refer to the instruction manual. Figure 2 Included with instruction manual Figure 3 The construction process of the formally unified semantic representation sequence in the cross-task prompting cascading module of this system is as follows:

[0103] Given three fixed task decomposition prefixes P1, P2, and P3, and three randomly selected templates T1, T4, and T7 from the prompt joint template, three corresponding cross-task semantic composite expressions are obtained by concatenating the original content of the factual statement with the task decomposition prefixes, as shown in the appendix to the instruction manual. Figure 3 As shown, by concatenating the original content of the fact statement "Scientists have discovered the existence of..." on Earth and the task decomposition prefix P1: "[New Hypothesis] = Regarding [Original Hypothesis], it is true.", we obtain the cross-task semantic compound statement "Regarding scientists' discovery of... on Earth, it is true".

[0104] Subsequently, the three cross-task semantic composite representations mentioned above undergo secondary processing, as shown in the appendix to the instruction manual. Figure 3 As shown, the randomly selected prompt template T4, "Assuming [premise], can we infer [new hypothesis]? Yes or no?", is combined with a cross-task semantic compound expression through slot insertion to obtain the sentence pattern for the Chinese rumor detection binary classification task: "Assuming the scientific community has not yet..., can we infer about scientists...? Yes or no?", which is called a semantic representation sequence with unified form, and serves as the input for the subsequent language model-based efficient parameter fine-tuning module.

[0105] The efficient parameter fine-tuning module based on the language model receives input from the cross-task prompting cascade module, a sequence of semantic representations in a uniform form, and completes the following encoding and decoding steps:

[0106] 1. Given the semantic representation sequence constructed above, input it into the pre-trained language model in the efficient parameter fine-tuning module to perform vector embedding on the natural language character sequence; map it to a one-hot vector using the vocabulary corresponding to the pre-trained language model, and perform low-dimensional embedding vector through the vocabulary matrix;

[0107] 2. Low-dimensional embedding vectors are obtained through the encoder-decoder layers and external optimizable parameter network of the pre-trained language model, and through multi-layer Transformer Block modules and optimizable parameter network, to obtain semantic vector representations that characterize the probability distribution of sequence predictions of the language model;

[0108] 3. This step must be completed during the model training phase, but is skipped during the prediction phase: Based on the probability distribution of the predicted sequence of the above language model, combined with the predefined model target sequence, calculate the three types of loss functions: conditional language model loss, negative sample penalty loss, and sentence length normalization loss. Use the backpropagation algorithm of loss values ​​to optimize the parameters of the external small-scale network and complete the training of the parameter fine-tuning module.

[0109] 4. This step is required during the model prediction phase; it is skipped during the training phase. See the attached manual. Figure 6 As shown, the semantic vector representation output obtained through the label decoding module is the probability distribution prediction of the language model sequence. Through a pipelined two-step label retrieval based on mapping table query and label retrieval based on probability sequence sorting, combined with the designed cross-task label mapping rule table and predefined target sequence, the final label of the Chinese fact-checking task is predicted.

[0110] Preferably, the computation process of the small-scale extended optimizable parameter network in the efficient parameter fine-tuning module based on the language model is as follows:

[0111] The corresponding instruction manual is attached. Figure 6 Given an input vector X, pre-attached feedforward layer parameters W0, and three trainable weight matrices... Here, the hyperparameter r is the rank of the low-rank matrix ≤ min(d,k), where d and k correspond to the input and output dimensions of the attached feedforward layer. The original feedforward network outputs h0 as h0 = W0X; in contrast, the feedforward network after the low-rank attached feedforward layer is calculated as h0 = W0X + ABCX. During training, the parameters of the attached feedforward layer of the pre-trained language model are fixed, and the parameters ABC of the circumscribed small-scale optimizable parameter network are optimized to achieve model learning.

[0112] Preferably, the efficient parameter fine-tuning module based on the language model includes three main types of loss: conditional language model loss L... LM Negative sample penalty loss L UK and sentence length standardization loss L LM The total loss of the model is the loss L mentioned above. LM L UK L LN The model is then optimized by calculating the gradient of the circumscribed parameter network through backpropagation, which is a simple summation of the two parameters.

[0113] Preferably, the input to the tag decoding module is the language model probability distribution sequence output by the efficient parameter fine-tuning module based on the language model, and the output is the final Chinese fact-checking task tag classification. The specific process of this module corresponds to the appendix to the instruction manual. Figure 6 :

[0114] Given three efficient parameter fine-tuning modules, output a sequence of language probability distributions. Corresponding to the three cross-task semantic composite representations of the three task decomposition prefixes, a pipeline-style two-step strategy is implemented: tag retrieval based on mapping table query and tag retrieval based on probability sequence sorting;

[0115] First, the tag retrieval method based on mapping table queries includes the following steps: calculating the probability distribution sequence of each output language. With predefined target sequence The cross-entropy between (N=2, the number of categories in the Chinese rumor detection task) is used to obtain a character-level cross-entropy set; by normalizing the sentence length, the average of the set is used to obtain the probability distribution sequence of each output language. Corresponding average cross-entropy Logit a For each output language probability distribution sequence Select the average cross-entropy calculation result Logit a Minimum predefined target sequence a k As intermediate answers; based on the combination of intermediate answers of three predefined target sequences corresponding to three cross-task semantic composite expressions, a query and matching is performed from the preset cross-task label mapping rule table to obtain the final label of the Chinese fact-checking task.

[0116] Since the preset cross-task tag mapping rule table of the tag retrieval method based on mapping table query does not completely include all combinations, a retrieval method based on probability sequence sorting is adopted for cases that do not appear in the preset cross-task tag mapping rule table.

[0117] The retrieval method based on probability sequence sorting, as shown in the appendix to the instruction manual. Figure 6 As shown, the process includes the following steps: obtaining the corresponding intermediate answer arrangement through the target labels in the preset cross-task label mapping rule table; calculating the cross-entropy between each intermediate answer in the intermediate answer arrangement and the corresponding predefined target sequence, and summing and averaging them to obtain the average cross-entropy of each label; finally, sorting by cross-entropy and selecting the label with the smallest average cross-entropy as the final label for the Chinese fact-checking task.

[0118] Overall, compared with previous mainstream research in the field of Chinese fact-checking, this system saves time and effort in data annotation, solves the current problem of insufficient data annotation, and avoids the dilemma of models over-relying on high-quality supervisory data for fact-checking. Moreover, in a supervisory environment with a small amount of fact-checking data, the fact-checking performance based on this system is still competitive to a certain extent.

[0119] To reiterate, the above embodiments are merely some examples of the present invention, and not embodiments compatible with all application scenarios. All other embodiments obtained by those skilled in the art without any inventive step, referring to the embodiments demonstrated in this invention, are also within the scope of protection of this invention.

Claims

1. A few-shot supervised Chinese fact-checking system enhanced by rumor detection data, characterized in that, Includes the following modules: The cross-task prompt cascade module includes task decomposition prefixes and prompt joint templates. Based on the inputs of the Chinese rumor detection task and the Chinese fact-checking task, it constructs a semantic representation sequence with a unified form. Given the inputs of the Chinese rumor detection task and the Chinese fact-checking task, the module processes the task decomposition prefixes and prompt joint templates in a pipeline manner and merges them into a semantic representation sequence with a unified form, which is the output of the cross-task prompt cascade module. The efficient parameter fine-tuning module based on the language model includes a pre-trained language model with fixed parameters and a small-scale extended optimizable parameter network. Based on the semantic representation sequence with a unified form, it predicts the probability distribution of the corresponding language model sequence through encoding and decoding operations. A tag decoding module includes a set of predefined target sequences and a tag mapping rule table across the rumor checking and fact checking tasks, N=2, for the number of original rumor detection task label categories; a decoding operation is performed on the predicted language model sequence probability distribution, that is, through the design of the cross-task label mapping rule table and the predefined target sequence, through the pipeline two-step strategy: label retrieval based on mapping table query and label retrieval based on probability sequence sorting, the intermediate answer with the highest fitting degree with the language model probability sequence is matched, so as to determine the final label of the Chinese fact checking task. The unified semantic representation sequence in the cross-task prompt cascading module is constructed as follows: Three fixed task decomposition prefixes are designed: three prompt prefixes that support "fact statement", deny "fact statement", and are unclear about "fact statement". By concatenating Chinese rumors, the task input is checked against the fact statement part of the task decomposition prefix, and the corresponding three cross-task semantic compound expressions are obtained respectively. Design a set of prompt joint templates. These templates are designed to perform secondary processing on three cross-task semantic composite expressions to obtain the sentence patterns of the original binary semantics for the Chinese rumor detection task, which can then be used as input for the subsequent efficient parameter fine-tuning module based on the language model. A prompt joint template is randomly selected from the set of designed prompt joint templates. Through the slots set in the template, it is combined with the three constructed cross-task semantic composite representations to obtain a semantic representation sequence with a unified form. This sequence is used as the input of the pre-trained language model in the efficient parameter fine-tuning module based on the language model.

2. The few-sample supervised Chinese fact-checking system enhanced with rumor detection data according to claim 1, characterized in that, The efficient parameter fine-tuning module based on language models in the model training phase is as follows: The original training data for Chinese rumor detection and Chinese fact-checking supervision are used to construct a unified semantic representation sequence through a cross-task prompting cascade module. All semantic representation sequences constructed from the original training data are randomly combined and used as input to the pre-trained language model in the efficient parameter fine-tuning module based on the language model. The efficient parameter fine-tuning module based on the language model optimizes the performance of the language model in predicting the probability distribution of sequences through a loss function, including the following steps: The randomly combined semantic representation sequence data is input into the pre-trained language model of this efficient parameter fine-tuning module to embed the character sequence with natural language as the carrier into a vector; the vocabulary corresponding to the language model is used to map to a one-hot vector, and a low-dimensional embedding vector is obtained through the vocabulary matrix; The low-dimensional embedding vector is passed through the encoding and decoding layers and the external optimizable parameter network of the pre-trained language model. Through the forward propagation of the multi-layer Transformer Block module and the optimizable parameter network, the semantic vector representation after the encoding and decoding operation of the language model is obtained. The physical meaning of the semantic vector is: the probability distribution of the predicted sequence of the language model. Based on the probability distribution of the predicted sequence from the pre-trained language model, and combined with the predefined target sequence, three types of loss functions are calculated: conditional language model loss, negative sample penalty loss, and sentence length normalization loss. The parameters of the external small-scale network are optimized using the backpropagation algorithm of the loss value, thus completing the training of the pre-trained language model of the efficient parameter fine-tuning module.

3. The few-sample supervised Chinese fact-checking system enhanced with rumor detection data according to claim 1, characterized in that, The efficient parameter fine-tuning module based on language models in the model prediction process uses the following method: By constructing a unified semantic representation sequence through a cross-task prompt cascade module, the semantic representation sequence data of the constructed Chinese fact-checking task is used as the input of the pre-trained language model in the efficient parameter fine-tuning module based on the language model. By using an efficient parameter fine-tuning module based on the language model, the probability distribution of the language model sequence is predicted. At this time, only the semantic vector representation after the language model encoding and decoding operations is calculated, that is, the probability distribution of the language model predicted sequence, and the loss calculation process and the backpropagation parameter update steps are discarded. The tag decoding module represents the output semantic vector, i.e., the probability distribution of the language model prediction sequence. Through a pipelined two-step tag retrieval based on mapping table lookup and tag retrieval based on probability sequence sorting, the final tag for the Chinese fact-checking task is predicted.

4. The few-sample supervised Chinese fact-checking system enhanced with rumor detection data according to claim 1, characterized in that, The small-scale extended optimizable parameter network in the efficient parameter fine-tuning module based on the language model includes: By performing extrinsic network parameter operations on the feedforward layer in each multi-head self-attention layer of all Transformer Blocks in the pre-trained language model, and optimizing only the extrinsic network parameters, efficient model parameter fine-tuning is achieved. Initialize three trainable weight matrices A∈ ^(d×r), B∈ ^(r×r), C∈ ^(r×k) where the hyperparameter r is the rank of the low-rank matrix ≤ min(d, k), and d and k correspond to the input and output dimensions of the feedforward layer to which it is attached; Given an input vector X and pre-attached feedforward layer parameters Original feedforward calculation Updated to low-rank modified feedforward computation During training, the parameters of the dependent feedforward layer of the pre-trained language model are fixed, and the parameters of the circumscribed small-scale optimizable parameter network are optimized. ABC This enables the learning of language models.

5. A few-sample supervised Chinese fact-checking system enhanced with rumor detection data according to claim 1, characterized in that, Three types of losses in an efficient parameter fine-tuning module based on a language model: Conditional language model loss Negative sample penalty loss and sentence length standardization loss The definition is as follows: in, Given an input X and the characters preceding the t-th character, the conditional probability of generating the t-th character is given; T represents the character length of the decoded sequence; and the conditional language model loss is... The aim is to maximize the fit between the language model sequence probability distribution and the correct model target sequence; in, Let represent the conditional probability of the t-th character in a negative sample given the input x and the preceding t-1 characters. The denominator of the above formula represents the sum of the conditional probabilities of all incorrect characters in the negative samples. The negative sample penalty loss is calculated by averaging the total length of all incorrect characters. The aim is to reduce the fit between the probability distribution of the model's output sequence and the incorrect model target sequence; in, It is the conditional language model loss. The reciprocal of the sum, i.e., the conditional probability of the language model, is the sentence length normalization loss. The aim is to reduce the impact of inconsistent output sequence lengths on model loss calculations and to control the length of the output sequence probability distribution to the correct target sequence length. The total loss of the model is set as the loss. The model is optimized by simply summing the values ​​and then calculating the gradient of the circumscribed parameter network through backpropagation.

6. The few-sample supervised Chinese fact-checking system enhanced with rumor detection data according to claim 1, characterized in that, The tag decoding module includes: The input to the label decoding module is the semantic vector representation of the efficient parameter fine-tuning module, which is the probability distribution of the predicted sequence by the language model. The tag decoding module outputs the final tag classification for the Chinese fact-checking task. The probability distribution of the three language model prediction sequences output by the efficient parameter fine-tuning module. This corresponds to three cross-task semantic composite representations based on task decomposition prefixes, and executes a pipeline-style two-step strategy: tag retrieval based on mapping table lookup and tag retrieval based on probability sequence sorting. The specific steps are as follows: First, the tag retrieval method based on mapping table lookup involves the following steps: Calculate the language model probability distribution sequence for each input. With predefined target sequence The cross-entropy between them is used to obtain a character-level cross-entropy set; By normalizing the sentence length and averaging the sets, we can obtain the probability distribution sequence of each output language. Corresponding average cross-entropy The formula is as follows: in, It's an abbreviation for Cross-Entropy. It refers to the entropy value between each character of the output language model probability distribution sequence and the predefined target sequence, and the entire expression represents the mean cross-entropy at the target sequence level; For each output language model probability distribution sequence Select the average cross-entropy calculation result Minimal predefined target sequence As an intermediate answer; Based on the intermediate answer combinations of three predefined target sequences corresponding to three cross-task semantic composite expressions, the final label classification of the Chinese fact-checking task is obtained by querying the pre-set cross-task label mapping rule table for matching.

7. A few-sample supervised Chinese fact-checking system enhanced with rumor detection data according to claim 6, characterized in that, Since the preset cross-task tag mapping rule table of the tag retrieval method based on mapping table query does not completely include all combinations, for cases that do not appear in the preset cross-task tag mapping rule table, a retrieval method based on probability sequence sorting is adopted. The specific steps are as follows: First, by using the target tags in the pre-defined cross-task tag mapping rule table, the corresponding intermediate answer arrangement is obtained; Secondly, the cross-entropy between each intermediate answer in all the intermediate answer permutations and the corresponding predefined target sequence is calculated, and the average of the cross-entropy results is obtained to get the average cross-entropy for each label; Finally, the label with the smallest average cross-entropy is selected as the final label classification for the Chinese fact-checking task by sorting by average cross-entropy.

Images