Method, device, equipment and storage for distinguishing true or false information based on causal intervention
Through the information authenticity and false identification method based on causal intervention, using partial training and counterfactual prediction, combined with the multi-head self-attention mechanism and Bert encoder, the error rate and low accuracy rate of true and false identification of information authenticity and false identification in the existing technology are solved. The problem of this problem has achieved more efficient identification of the authenticity of information and improved the security and stability of information dissemination.
Patent Information
- Application Number
- CN202411344624.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-25
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2044-09-25
AI Technical Summary
The existing methods for identifying authenticity and false information have problems such as high misjudgment rate, low accuracy rate and time-consuming and labor-intensive information. It is especially difficult to achieve automatic authenticity and false identification in big data environments, and machine learning-based methods have insufficient causal reasoning capabilities.
Using a method of authenticity and false information identification based on causal intervention, the data set with labels is constructed to carry out partial training, counterfactual prediction and unbiased training, and the text embedding module, evidence interaction module and fusion module are used, combined with the Bert encoder and The multi-head self-attention mechanism enables more accurate and robust reasoning.
It effectively improves the accuracy of identifying authenticity of information, reduces public risks and manpower and material consumption, and improves the security and stability of information dissemination.
Smart Images

Figure CN118885617B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of natural language processing technology, and in particular to a method, apparatus, device and memory for distinguishing the authenticity of information based on causal intervention. Background Art
[0002] The rapid development of digital media and the rise and widespread use of social networking platforms in recent years have not only greatly increased information accessibility but also accelerated the spread of false information. Consequently, the impact of false information is growing, sometimes even extending into the real world and threatening public safety. Automatically verifying the authenticity of information is a valuable problem, reducing the time and effort required to detect and prevent the spread of false information. Existing methods for verifying the authenticity of information suffer from high false positive rates. Advanced technologies are urgently needed to continuously optimize existing models to reduce false positive rates and improve recall.
[0003] Traditional methods for verifying information authenticity involve using data provided by news collection and distribution systems to analyze news reports for possible falsities and inaccuracies. Authenticity is then verified manually. However, this manual analysis-based news verification model suffers from low accuracy and is time-consuming and labor-intensive. Furthermore, due to its low efficiency and cumbersome procedures, this method presents difficulties in automatically verifying authenticity in large data sets.
[0004] In recent years, machine learning-based methods have been widely used to improve the efficiency of information verification and enhance information filtering. These include natural language processing, which analyzes the linguistic features, grammatical structure, and semantic content of text to identify potential false information. Another approach involves deep learning models, particularly convolutional neural networks and recurrent neural networks, which are used to analyze news text and multimedia content. However, machine learning-based methods for verifying information authenticity still suffer from weak reasoning about causal relationships, resulting in low accuracy in determining the authenticity of news based on evidence. Summary of the Invention
[0005] Based on this, it is necessary to provide a method, device, equipment and storage for information authenticity identification based on causal intervention, which can accurately identify the authenticity of news or intelligence through relevant evidence reasoning, in response to the above technical problems.
[0006] A method for distinguishing true from false information based on causal intervention, the method comprising:
[0007] According to the specific task orientation of news or information, text data in specific fields are collected to build data sets.
[0008] Based on the news, we collect evidence that can prove its authenticity and record the source to build a labeled training dataset.
[0009] A pre-built information authenticity identification model based on causal intervention is trained using a labeled training data set to obtain a trained information authenticity identification model; the training process includes biased training, counterfactual prediction and unbiased training; biased training is used to take news and evidence pairs as input, and obtain biased prediction results through the interaction and feature fusion of news and evidence; counterfactual prediction is used to block evidence through causal intervention and output the model's counterfactual prediction results; unbiased training is used to obtain unbiased prediction results based on biased prediction results and counterfactual prediction results.
[0010] The news data to be identified is input into the trained information authenticity identification model to obtain biased prediction results containing news and evidence.
[0011] According to the average evidence representation of the dataset, the counterfactual prediction results of the news to be identified are given through causal intervention.
[0012] According to the biased prediction results containing news and evidence and the counterfactual prediction results of the news to be identified, the authenticity identification results of the news data to be identified are obtained.
[0013] In one embodiment, the biased training process includes: taking news and evidence pairs as input, passing through the evidence interaction module and the fusion module, and obtaining the biased prediction result containing news and evidence as follows:
[0014] ;
[0015] in, represents a biased prediction result containing news and evidence, represents the fusion function, Indicates news, Indicates evidence.
[0016] In one embodiment, the counterfactual prediction process includes inferring possible outcomes by assuming that certain variables do not work, and outputting the model's counterfactual prediction result as follows:
[0017] ;
[0018] in, represents the counterfactual prediction result, Indicates news, represents the average evidence representation on the dataset, represents the fusion function, Represents the normalized linear classification layer function.
[0019] In one embodiment, the unbiased training process includes: optimizing model parameters based on biased prediction results and counterfactual prediction results with minimizing cross entropy loss as the training goal to obtain unbiased prediction results.
[0020] In one embodiment, the information truth and falsehood identification model is a causal inference model based on causal intervention.
[0021] In one embodiment, the information authenticity identification model based on causal intervention includes: a text embedding module, an evidence interaction module and a fusion module; the text embedding module is used to adopt a Bert encoder to utilize the world knowledge and understanding ability of the pre-trained language model; the evidence interaction module is used to adopt a multi-layer transformer structure with a multi-head self-attention mechanism, dynamically adjust the attention allocation according to the weights of different evidence, and enhance the correlation between evidence through cross-layer information transmission, so as to achieve more accurate and robust reasoning in a complex environment; the fusion module is used to add residual connections and adaptive activation functions on the basis of a multi-layer fully connected feedforward neural network to enhance the expressive power of the model and the stability of gradient propagation.
[0022] In one embodiment, based on the biased prediction result including news and evidence and the counterfactual prediction result of the news to be identified, a result of identifying the authenticity of the news data to be identified is obtained, including:
[0023] The debiasing factor is determined according to actual work requirements; the value range of the debiasing factor is 0.1 to 0.9.
[0024] According to the biased prediction results containing news and evidence, the counterfactual prediction results of the news to be identified and the debiasing factor, the authenticity identification results of the news data to be identified are obtained.
[0025] A device for distinguishing true from false information based on causal intervention, the device comprising:
[0026] The module for building an annotated training dataset is used to collect text data in a specific field to build a dataset based on the specific task orientation of news or information; it also collects evidence that can prove the authenticity of news and records the source to build an annotated training dataset.
[0027] The information truth-and-false identification model training module is used to train a pre-built information truth-and-false identification model based on causal intervention using a labeled training data set to obtain a trained information truth-and-false identification model; the training process includes biased training, counterfactual prediction, and unbiased training; biased training is used to take news and evidence pairs as input, and obtain biased prediction results through the interaction and feature fusion of news and evidence; counterfactual prediction is used to block evidence through causal intervention and output the model's counterfactual prediction results; unbiased training is used to obtain unbiased prediction results based on biased prediction results and counterfactual prediction results.
[0028] The news data information authenticity identification module is used to input the news data to be identified into a trained information authenticity identification model to obtain biased prediction results containing news and evidence; based on the average evidence representation of the data set, a counterfactual prediction result of the news to be identified is given through causal intervention; based on the biased prediction result containing news and evidence and the counterfactual prediction result of the news to be identified, the information authenticity identification result of the news data to be identified is obtained.
[0029] A computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:
[0030] According to the specific task orientation of news or information, text data in specific fields are collected to build data sets.
[0031] Based on the news, we collect evidence that can prove its authenticity and record the source to build a labeled training dataset.
[0032] A pre-built information authenticity identification model based on causal intervention is trained using a labeled training data set to obtain a trained information authenticity identification model; the training process includes biased training, counterfactual prediction and unbiased training; biased training is used to take news and evidence pairs as input, and obtain biased prediction results through the interaction and feature fusion of news and evidence; counterfactual prediction is used to block evidence through causal intervention and output the model's counterfactual prediction results; unbiased training is used to obtain unbiased prediction results based on biased prediction results and counterfactual prediction results.
[0033] The news data to be identified is input into the trained information authenticity identification model to obtain biased prediction results containing news and evidence.
[0034] According to the average evidence representation of the dataset, the counterfactual prediction results of the news to be identified are given through causal intervention.
[0035] According to the biased prediction results containing news and evidence and the counterfactual prediction results of the news to be identified, the authenticity identification results of the news data to be identified are obtained.
[0036] A computer readable memory having a computer program stored thereon, wherein when the computer program is executed by a processor, the following steps are implemented:
[0037] According to the specific task orientation of news or information, text data in specific fields are collected to build data sets.
[0038] Based on the news, we collect evidence that can prove its authenticity and record the source to build a labeled training dataset.
[0039] A pre-built information authenticity identification model based on causal intervention is trained using a labeled training data set to obtain a trained information authenticity identification model; the training process includes biased training, counterfactual prediction and unbiased training; biased training is used to take news and evidence pairs as input, and obtain biased prediction results through the interaction and feature fusion of news and evidence; counterfactual prediction is used to block evidence through causal intervention and output the model's counterfactual prediction results; unbiased training is used to obtain unbiased prediction results based on biased prediction results and counterfactual prediction results.
[0040] The news data to be identified is input into the trained information authenticity identification model to obtain biased prediction results containing news and evidence.
[0041] According to the average evidence representation of the dataset, the counterfactual prediction results of the news to be identified are given through causal intervention.
[0042] According to the biased prediction results containing news and evidence and the counterfactual prediction results of the news to be identified, the authenticity identification results of the news data to be identified are obtained.
[0043] The aforementioned causal intervention-based information authenticity verification method, device, equipment, and storage device accurately verify the authenticity of news or intelligence through reasoning based on relevant evidence. This method fully utilizes relevant news data, removes sample selection bias, and implements evidence-based reasoning and authenticity verification. Through the causal intervention algorithm, it effectively improves the model's reasoning capabilities and the accuracy of authenticity verification. This reduces public risks and human and material resource consumption in the information sector while further enhancing the security and stability of information dissemination. BRIEF DESCRIPTION OF THE DRAWINGS
[0044] Figure 1 1 is a flow chart of a method for identifying the authenticity of information based on causal intervention in one embodiment;
[0045] Figure 2 A basic architecture diagram of a method for identifying the authenticity of information based on causal intervention in another embodiment;
[0046] Figure 32. It is a structural block diagram of an information authenticity identification device based on causal intervention in one embodiment;
[0047] Figure 4 FIG. 1 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION
[0048] In order to make the purpose, technical solutions and advantages of this application more clearly understood, the present application is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.
[0049] In one embodiment, Figure 1 As shown, a method for distinguishing the authenticity of information based on causal intervention is provided, which includes the following steps:
[0050] Step 100: Based on the specific task orientation of news or information, collect text data in a specific field to build a dataset.
[0051] Specifically, we specify the field of truth and falsehood identification and collect relevant news to build a data set.
[0052] Step 102: Collect evidence that can prove the authenticity of the news and record the source to build a labeled training data set.
[0053] Specifically, we collect relevant evidence that can verify the authenticity of news and record the source, and annotate the authenticity of the news to form news-evidence pairs and construct a labeled training data set.
[0054] Step 104: Use a labeled training data set to train the pre-built information authenticity identification model based on causal intervention to obtain a trained information authenticity identification model; wherein, the training process includes biased training, counterfactual prediction and unbiased training; biased training is used to take news and evidence pairs as input, and obtain biased prediction results through the interaction and feature fusion of news and evidence; counterfactual prediction is used to block evidence through causal intervention, specifically, using the average evidence representation on the data set as the input of evidence together with the news to input the model, and output the model's counterfactual prediction results; unbiased training is used to obtain unbiased prediction results based on the biased prediction results and the counterfactual prediction results.
[0055] Specifically, the basic architecture of the information truth and falsehood identification model based on causal intervention is as follows: Figure 2 shown.
[0056] The information authenticity identification model based on causal intervention consists of three parts: a text embedding module, an evidence interaction module, and a fusion module. General information authenticity identification models can be divided into these three components, but the specific structure of each module varies. Therefore, theoretically, this method is a universal method for improving the accuracy of authenticity identification. Specifically, in this method, the text embedding module uses a Bert encoder to leverage the world knowledge and understanding capabilities of a pre-trained language model. The evidence interaction module is a multi-layer transformer structure that uses a multi-head self-attention mechanism. In particular, this module also introduces a dynamic weight allocation strategy and cross-layer information fusion technology. This module can dynamically adjust attention allocation based on the weight of different evidence and enhance the correlation between evidence through cross-layer information transfer, thereby achieving more accurate and robust reasoning in complex environments. The fusion module adds residual connections and adaptive activation functions to the multi-layer fully connected feedforward neural network to enhance the model's expressive power and gradient propagation stability.
[0057] Causally intervene in the true-false identification algorithm and train and determine the parameters of the neural network that can perform true-false identification based on evidence reasoning.
[0058] Step 106: Input the news data to be identified into the trained information authenticity identification model to obtain a biased prediction result containing news and evidence.
[0059] Specifically, the news data to be tested and related evidence are input into the trained information truth and falsehood identification model to give a biased prediction result combining the news and evidence.
[0060] Each piece of news has 1 to 10 pieces of evidence, and the specific website or source of the evidence is given. If the news is false, it is marked as 0, and if the news is true, it is marked as 1.
[0061] Step 108: Based on the average evidence representation of the data set, a counterfactual prediction result of the news to be identified is given through causal intervention.
[0062] Step 110: Based on the biased prediction results containing news and evidence and the counterfactual prediction results of the news to be identified, obtain the authenticity identification results of the news data to be identified.
[0063] The aforementioned causal intervention-based information authenticity verification method accurately verifies news or intelligence through reasoning based on relevant evidence. This method fully utilizes relevant news data, removes sample selection bias, and implements evidence-based reasoning and authenticity verification. Through the causal intervention algorithm, it effectively improves the model's reasoning capabilities and the accuracy of authenticity verification. This reduces public risks and human and material resource consumption in the information sector while further enhancing the security and stability of information dissemination.
[0064] In one embodiment, the biased training process includes: taking news and evidence pairs as input, passing through the evidence interaction module and the fusion module, and obtaining the biased prediction result containing news and evidence as follows:
[0065] ;
[0066] in, represents a biased prediction result containing news and evidence, represents the fusion function, Indicates news, Indicates evidence.
[0067] Specifically, during the biased training process: the news and evidence At the same time, as the input of the model, the evidence interaction module outputs the probability distribution of true and false categories, that is, the biased prediction results containing news and evidence. , represents the fusion function. At this point, the model estimates the total causal effect of news and evidence on true and false identification.
[0068] In one embodiment, the counterfactual prediction process includes inferring possible outcomes by assuming that certain variables do not work, and outputting the model's counterfactual prediction result as follows:
[0069] ;
[0070] in, represents the counterfactual prediction result, Indicates news, represents the average evidence representation on the dataset, represents the fusion function, Represents the normalized linear classification layer function.
[0071] Specifically, in the counterfactual prediction process: after biased training, the model, in addition to learning reasoning, also inadvertently memorizes and relies on some false correlations for prediction. Counterfactual prediction infers possible results by assuming that some variables do not play a role, that is, it can distinguish how much the remaining variables contribute to the results. By blocking evidence through causal intervention, the model's counterfactual prediction result can be output as , is the average evidence representation on the dataset.
[0072] In one embodiment, the unbiased training process includes: optimizing model parameters based on biased prediction results and counterfactual prediction results with minimizing cross entropy loss as the training goal to obtain unbiased prediction results.
[0073] Specifically, in the unbiased training process: combining biased training and counterfactual prediction to obtain unbiased prediction results, in order to optimize the trainable parameters of the model , using cross entropy loss as the training objective and minimizing it.
[0074] In one embodiment, the information truth and falsehood identification model is a causal inference model based on causal intervention.
[0075] Specifically, the causal inference algorithm uses causal intervention to obtain the model's counterfactual prediction, which is then subtracted from the model's base prediction to remove model bias and selection bias. Specifically, the steps involve first inputting news and evidence into the truth-and-fake identification model to obtain the model's base prediction. Then, the news and blocked data are fed into the truth-and-fake identification model to obtain the model's counterfactual prediction. Finally, the base prediction and counterfactual prediction are combined using element-wise subtraction to obtain the model's unbiased prediction.
[0076] In one embodiment, the information authenticity identification model based on causal intervention includes: a text embedding module, an evidence interaction module and a fusion module; the text embedding module is used to adopt a Bert encoder to utilize the world knowledge and understanding ability of the pre-trained language model; the evidence interaction module is used to adopt a multi-layer transformer structure with a multi-head self-attention mechanism, dynamically adjust the attention allocation according to the weights of different evidence, and enhance the correlation between evidence through cross-layer information transmission, so as to achieve more accurate and robust reasoning in a complex environment; the fusion module is used to add residual connections and adaptive activation functions on the basis of a multi-layer fully connected feedforward neural network to enhance the expressive power of the model and the stability of gradient propagation.
[0077] In one embodiment, step 110 includes: determining a debiasing factor according to actual work requirements; wherein the debiasing factor has a value range of 0.1 to 0.9; and obtaining an information authenticity identification result for the news data to be identified based on the biased prediction results containing news and evidence, the counterfactual prediction results of the news to be identified, and the debiasing factor.
[0078] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. In addition, Figure 1At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.
[0079] In one embodiment, Figure 3 As shown, a device for distinguishing true or false information based on causal intervention is provided, comprising: a labeled training data set construction module, an information true or false distinction model training module, and a news data information true or false distinction module, wherein:
[0080] The module for building an annotated training dataset is used to collect text data in a specific field to build a dataset based on the specific task orientation of news or information; it also collects evidence that can prove the authenticity of news and records the source to build an annotated training dataset.
[0081] The information truth-and-false identification model training module is used to train a pre-built information truth-and-false identification model based on causal intervention using a labeled training data set to obtain a trained information truth-and-false identification model; the training process includes biased training, counterfactual prediction, and unbiased training; biased training is used to take news and evidence pairs as input, and obtain biased prediction results through the interaction and feature fusion of news and evidence; counterfactual prediction is used to block evidence through causal intervention and output the model's counterfactual prediction results; unbiased training is used to obtain unbiased prediction results based on biased prediction results and counterfactual prediction results.
[0082] The news data information authenticity identification module is used to input the news data to be identified into a trained information authenticity identification model to obtain biased prediction results containing news and evidence; based on the average evidence representation of the data set, a counterfactual prediction result of the news to be identified is given through causal intervention; based on the biased prediction result containing news and evidence and the counterfactual prediction result of the news to be identified, the information authenticity identification result of the news data to be identified is obtained.
[0083] In one embodiment, the biased training process in the information authenticity identification model training module includes: taking news and evidence pairs as input, and passing through the evidence interaction module and the fusion module to obtain the biased prediction result containing news and evidence as follows:
[0084] ;
[0085] in, represents a biased prediction result containing news and evidence, represents the fusion function, Indicates news, Indicates evidence.
[0086] In one embodiment, the counterfactual prediction process in the information authenticity identification model training module includes: inferring possible results by assuming that certain variables do not work, and outputting the model's counterfactual prediction result as follows:
[0087] ;
[0088] in, represents the counterfactual prediction result, Indicates news, represents the average evidence representation on the dataset, represents the fusion function, Represents the normalized linear classification layer function.
[0089] In one embodiment, the unbiased training process in the information authenticity identification model training module includes: optimizing the model parameters based on the biased prediction results and the counterfactual prediction results, taking the minimization of cross entropy loss as the training goal, and obtaining the unbiased prediction results.
[0090] In one embodiment, the information truth and falsehood identification model is a causal inference model based on causal intervention.
[0091] In one embodiment, the number of relevant evidences for each news item is 1 to 10, and the specific website or source of the evidence is given.
[0092] In one embodiment, the information authenticity identification module for news data is also used to determine a debiasing factor based on actual work requirements; the debiasing factor has a value range of 0.1 to 0.9; based on the biased prediction results containing news and evidence, the counterfactual prediction results of the news to be identified, and the debiasing factor, the information authenticity identification result of the news data to be identified is obtained.
[0093] The specific limitations of the information authenticity identification device based on causal intervention can be found in the limitations of the information authenticity identification method based on causal intervention above and will not be repeated here. The various modules in the above-mentioned information authenticity identification device based on causal intervention can be implemented in whole or in part through software, hardware, or a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of each of the above modules.
[0094] In one embodiment, a computer device is provided. The computer device may be a terminal, and its internal structure diagram may be as follows: Figure 4As shown. The computer device includes a processor, a memory, a network interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal via a network connection. When the computer program is executed by the processor, a method for identifying the authenticity of information based on causal intervention is implemented. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covering the display screen, or a button, trackball or touchpad provided on the computer device housing, or an external keyboard, touchpad or mouse.
[0095] Those skilled in the art will understand that Figure 4 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than shown in the figure, or combine certain components, or have a different component arrangement.
[0096] In one embodiment, a computer device is provided, including a memory and a processor. The memory stores a computer program, and the processor implements the steps in the above method embodiment when executing the computer program.
[0097] In one embodiment, a computer readable memory is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above method embodiment are implemented.
[0098] Those skilled in the art will appreciate that all or part of the processes in the above-described method embodiments can be implemented by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable memory. When executed, the computer program can include the processes of the above-described method embodiments. Any reference to memory, storage, database, or other media used in the various embodiments provided herein may include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct RAMbus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM).
[0099] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0100] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.
Claims
1. A method for identifying the authenticity of information based on causal intervention. It is characterized in that The method comprises: According to the specific task orientation of news or information, collect text data in specific fields to build data sets; Collect evidence that can prove the authenticity of news and record the source to build a labeled training data set; A pre-built information authenticity identification model based on causal intervention is trained using a labeled training data set to obtain a trained information authenticity identification model; wherein the training process includes biased training, counterfactual prediction and unbiased training; the biased training is used to take news and evidence pairs as input, and obtain biased prediction results through the interaction and feature fusion of news and evidence; the counterfactual prediction is used to block evidence through causal intervention and output the counterfactual prediction results of the model; the unbiased training is used to obtain unbiased prediction results based on the biased prediction results and the counterfactual prediction results; Input the news data to be identified into the trained information authenticity identification model to obtain biased prediction results containing news and evidence; According to the average evidence representation of the data set, the counterfactual prediction results of the news to be identified are given through causal intervention; According to the biased prediction results including news and evidence and the counterfactual prediction results of the news to be identified, the authenticity identification results of the news data to be identified are obtained; The biased training process includes: taking news and evidence pairs as input, and passing through the evidence interaction module and the fusion module to obtain the biased prediction result containing news and evidence: in, represents a biased prediction result containing news and evidence, represents the fusion function, Indicates news, to indicate evidence; The counterfactual prediction process includes: inferring possible results by assuming that certain variables do not work, and the counterfactual prediction results of the output model are: in, represents the counterfactual prediction result, Indicates news, represents the average evidence representation on the dataset, represents the fusion function, f represents the normalized linear classification layer function; The unbiased training process comprises: optimizing the model parameters according to the biased prediction result and the counterfactual prediction result, taking the minimization of cross entropy loss as the training goal, and obtaining the unbiased prediction result; The information authenticity identification model based on causal intervention includes: a text embedding module, an evidence interaction module and a fusion module; the text embedding module is used to adopt a Bert encoder to utilize the world knowledge and understanding ability of a pre-trained language model; the evidence interaction module is used to adopt a multi-layer transformer structure with a multi-head self-attention mechanism, dynamically adjust the attention allocation according to the weights of different evidences, and enhance the correlation between evidences through cross-layer information transmission, so as to achieve more accurate and robust reasoning in a complex environment; the fusion module is used to add residual connections and adaptive activation functions on the basis of a multi-layer fully connected feedforward neural network, so as to enhance the expression ability of the model and the stability of gradient propagation.
2. The method according to claim 1, It is characterized in that The information truth and falsehood identification model is a causal inference model based on causal intervention.
3. The method according to claim 1, It is characterized in that According to the biased prediction results containing news and evidence and the counterfactual prediction results of the news to be identified, the true or false identification results of the news data to be identified are obtained, including: Determine the debiasing factor according to actual work requirements; the debiasing factor ranges from 0.1 to 0.9; According to the biased prediction results including news and evidence, the counterfactual prediction results of the news to be identified and the debiasing factor, the authenticity identification result of the news data to be identified is obtained.
4. A device for distinguishing true from false information based on causal intervention, It is characterized in that The device comprises: The module for building an annotated training dataset is used to collect text data in a specific field to build a dataset according to the specific task orientation of news or information; collect evidence that can prove the authenticity of the news and record the source to build an annotated training dataset; The information truth and falsehood identification model training module is used to train the pre-built information truth and falsehood identification model based on causal intervention with a labeled training data set to obtain a trained information truth and falsehood identification model; wherein the training process includes biased training, counterfactual prediction and unbiased training; the biased training is used to take news and evidence pairs as input, and obtain biased prediction results through the interaction and feature fusion of news and evidence; the counterfactual prediction is used to block evidence through causal intervention and output the counterfactual prediction results of the model; the unbiased training is used to obtain unbiased prediction results based on the biased prediction results and the counterfactual prediction results; The information authenticity identification module of news data is used to input the news data to be identified into the trained information authenticity identification model to obtain biased prediction results containing news and evidence; according to the average evidence representation of the data set, the counterfactual prediction results of the news to be identified are given through causal intervention; according to the biased prediction results containing news and evidence and the counterfactual prediction results of the news to be identified, the information authenticity identification results of the news data to be identified are obtained; Among them, the biased training process in the information authenticity identification model training module includes: taking the news and evidence pairs as input, and after passing through the evidence interaction module and the fusion module, the biased prediction result containing news and evidence is obtained as follows: in, represents a biased prediction result containing news and evidence, represents the fusion function, Indicates news, to indicate evidence; The counterfactual prediction process described in the information authenticity identification model training module includes: inferring possible results by assuming that certain variables do not work, and the counterfactual prediction results of the output model are: in, represents the counterfactual prediction result, Indicates news, represents the average evidence representation on the dataset, represents the fusion function, f represents the normalized linear classification layer function; The unbiased training process in the information authenticity identification model training module includes: optimizing the model parameters according to the biased prediction result and the counterfactual prediction result, taking the minimization of cross entropy loss as the training goal, and obtaining the unbiased prediction result; The information authenticity identification model based on causal intervention in the information authenticity identification model training module includes: a text embedding module, an evidence interaction module and a fusion module; the text embedding module is used to adopt a Bert encoder to utilize the world knowledge and understanding ability of the pre-trained language model; the evidence interaction module is used to adopt a multi-layer transformer structure with a multi-head self-attention mechanism, dynamically adjust the attention allocation according to the weights of different evidences, and enhance the correlation between evidences through cross-layer information transmission, so as to achieve more accurate and robust reasoning in a complex environment; the fusion module is used to add residual connections and adaptive activation functions on the basis of a multi-layer fully connected feedforward neural network to enhance the expression ability of the model and the stability of gradient propagation.
5. A computer device comprising a memory and a processor, wherein the memory stores a computer program, It is characterized in that When the processor executes the computer program, the method according to any one of claims 1 to 3 is implemented.
6. A computer readable memory having a computer program stored thereon, It is characterized in that When the computer program is executed by a processor, the method according to any one of claims 1 to 3 is implemented.
Citation Information
Patent Citations
Cross-modal retrieval model and method based on anti-fact reasoning and computer equipment
CN115146100A
Bias Identification in Cognitive Computing Systems
US20210097405A1
