Method and device for identifying fake news based on traceability reasoning, equipment and medium

Abstract

The application discloses a false news identification method and device based on traceability reasoning, equipment and medium. Based on text classification, text similarity and traceability natural language generation technology, a new false news identification process is designed, which solves the problem that new news cannot be identified due to the lag of fact database update in the fact-based method to a certain extent.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information processing, and particularly relates to a false news identification method and device based on trace reasoning, equipment and medium. BACKGROUND

[0002] With the globalization of computer technology and smart devices such as mobile phones and computers, people can easily obtain information from the Internet and various social platforms. However, various false news is rampant in social media, seriously damaging people's right to know the truth of things, and also disrupting the network order and social stability. Intelligent and automated false news identification is of great significance to improving the network information environment.

[0003] At present, false news identification methods mainly fall into two categories: one is a false news identification method based on content, which can be further divided into two categories based on facts and style; the second is a false news identification method based on transmission. The false news identification method based on facts is realized by automatic fact checking, and the inconsistencies in the to-be-identified news with the facts or common sense are judged according to the pre-established fact library, knowledge graph and other source data, mainly using text similarity, natural language reasoning and other technologies. The false news identification method based on style uses the emotion and writing style of news to identify the true or false of news, mainly using natural language processing and other technologies. The false news identification method based on transmission is realized by modeling the social transmission chain of news, mainly using graph neural networks and other technologies.

[0004] However, the method based on transmission lacks explainability in the identification results, and the identification results without evidence support are usually not convincing. The method based on style is often affected by some words with strong emotion (such as shock, serious), which leads to the method being more inclined to identify so-called "sensationalist" information. The method based on facts is currently the most accurate and has a certain explainability, but this method relies on pre-established fact library and knowledge graph and other source data, and these data are usually updated slowly, which leads to the inability to identify new false news in time. SUMMARY

[0005] The purpose of the present application is to overcome the defects of the prior art and provide a false news identification method, device, equipment and medium based on trace reasoning, which designs a new false news identification process based on text classification, text similarity and trace natural language generation.

[0006] The purpose of the present application is achieved by the following technical solutions:

[0007] A false news identification method based on trace reasoning, the method comprising:

[0008] Acquire a news text to be identified, and perform a preprocessing operation, the preprocessing including sentence segmentation and removing useless sentences, to obtain one or more event sentences to be identified;

[0009] Vectorize the event sentences to be identified and calculate the cosine similarity with all event vectors in the event graph to find an event node with the maximum similarity;

[0010] Query event nodes with multi-hop similarity and sequential relationship from the starting node and the multi-hop relationship nodes to form a relevant event set, and the relevant event set contains the preceding events related to the event sentences to be identified;

[0011] Combine each event in the relevant event set with the event sentences to be identified to form an event sequence, and perform abductive reasoning on each event sequence to obtain an explanation event set;

[0012] Take the event node corresponding to each explanation event in the explanation event set as an explanation event starting node, query event nodes with multi-hop similarity and causal relationship from the starting node and the multi-hop relationship nodes to form a causal event set;

[0013] Determine the truth of the events according to the maximum cosine similarity between the events in the explanation event set and the events in the causal event set, and obtain the authenticity judgment result of the news to be identified.

[0014] Further, the construction method of the event graph includes:

[0015] Acquire real news samples, segment the real news samples to obtain independent event representation sentences, and remove the useless sentence in the real news samples;

[0016] Tri-classify any two event representation sentences in the real news samples to determine whether the two events are unrelated, sequential, or causal;

[0017] Vectorize the events with causal and sequential relationships after classification, and then construct an event graph in the form of a triple by taking the event sentences and event vectors as nodes and the relationships as edges;

[0018] Further, the method further includes classifying the event representation sentences between real news samples with a publication time interval less than a preset threshold.

[0019] Further, the method establishes a similarity edge between similar event nodes by calculating the cosine similarity between event vectors.

[0020] Further, the judging true or false of the event according to the maximum cosine similarity of the event in the explanation event set and the event in the causal event set specifically comprises:

[0021] If the maximum cosine similarity of an event in the causal event set and an event in the explanation event set is greater than the similarity threshold, it is recorded as a real event and the judgment of the remaining explanation events in the explanation event set is skipped.

[0022] If the maximum cosine similarity of any event in the causal event set and any event in the explanation event set is not greater than the similarity threshold, it is recorded as a false event.

[0023] Further, the vectorization representation adopts a pre-trained language model to generate a sentence vector with semantic meaning using a twin structure.

[0024] Further, the abductive reasoning adopts a multi-hop reasoning flow generation model.

[0025] In another aspect, the present application also provides a false news recognition device based on abductive reasoning, which is used to implement the false news recognition method based on abductive reasoning described above, and the device comprises:

[0026] A text preprocessing module is configured to obtain a news text to be recognized, and perform preprocessing operations, the preprocessing including sentence segmentation and removing useless words, to obtain one or more event sentences to be recognized;

[0027] A similarity calculation module is configured to vectorize the event sentences to be recognized and calculate the cosine similarity with all event vectors in the event graph, to find an event node with the maximum similarity;

[0028] A related event set construction module is configured to query event nodes with multi-hop similarity and sequential relationship from a starting node with the maximum similarity, to form a related event set from the starting node and the multi-hop relationship nodes, and the related event set contains pre-event related to the event sentence to be recognized;

[0029] An explanation event set construction module is configured to form an event sequence from each event in the related event set and the event sentence to be recognized, and perform abductive reasoning on each event sequence to obtain an explanation event set;

[0030] A causal event set construction module is configured to take an event node corresponding to each explanation event in the explanation event set as an explanation event starting node, query event nodes with multi-hop similarity and causal relationship from the explanation event starting node, and form a causal event set from the starting node and the multi-hop relationship nodes;

[0031] The authenticity judgment module is configured to judge the authenticity of the events according to the maximum cosine similarity of the events in the explanation event set and the events in the causal event set, and obtain an authenticity judgment result of the news to be identified.

[0032] In another aspect, the present application also provides a computer device, which comprises a processor and a memory, and the memory stores a computer program, which is loaded and executed by the processor to implement any of the above-mentioned false news identification methods based on traceability reasoning.

[0033] In another aspect, the present application also provides a computer readable storage medium, which stores a computer program, and the computer program is loaded and executed by a processor to implement any of the above-mentioned false news identification methods based on traceability reasoning.

[0034] The present application has the following beneficial effects:

[0035] The present application can identify newly generated false news by combining event graph and news to be identified for traceability reasoning while ensuring that the identification result has certain explainability, and to some extent, solves the problem that the fact-based method cannot identify newly generated news due to the lag in updating the fact library. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a flowchart of the false news identification method based on traceability reasoning provided by the embodiment of the present application;

[0037] Figure 2 is a flowchart of constructing an event graph as a fact library based on existing real news according to the embodiment of the present application;

[0038] Figure 3 is a structural block diagram of the false news identification device based on traceability reasoning provided by the embodiment of the present application. DETAILED DESCRIPTION

[0039] The embodiments of the present application will be described below through specific and concrete examples, and those skilled in the art can easily understand other advantages and effects of the present application from the disclosure of the present specification. The present application can also be implemented or applied through other different specific embodiments, and each detail in the present specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict.

[0040] Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0041] The propagation-based method lacks explainability in the identification result, and the identification result without evidence support is usually not convincing. The style-based method is often affected by some words with strong emotions (such as shock and serious), so that the method tends to identify so-called "sensational" information. The fact-based method is currently an accurate and somewhat explainable identification method, but this method relies on pre-established fact library and knowledge graph and other source data, and these data are usually updated with a lag, so that the newly generated false news cannot be identified in time.

[0042] To solve the above technical problems, the following embodiments of the false news identification method, device, equipment and medium based on traceability reasoning of the present application are proposed.

[0043] Embodiment 1

[0044] Reference Figure 1 As Figure 1 The flowchart of the false news identification method based on traceability reasoning provided by the present embodiment is shown in the figure, and the method specifically includes the following steps:

[0045] Step 1: Obtain the news text to be identified, and perform preprocessing operation, which includes sentence segmentation and removing useless sentences, to obtain one or more event sentences to be identified.

[0046] Step 2: Vectorize the event sentence to be identified and calculate the cosine similarity with all event vectors in the event graph to find the event node with the maximum similarity.

[0047] Use SBert to vectorize the event sentence to be identified Q, and calculate the cosine similarity with all event vectors in the event graph to find the event node A with the maximum similarity.

[0048] It should be noted that Figure 2 As Figure 2 The flowchart of constructing the event graph as a fact library based on the existing real news in the present embodiment is shown in the figure, and the construction method of the event graph in the present embodiment is as follows:

[0049] In the present embodiment, the crawler technology is used to collect real news articles from various trusted news websites, and independent event representation sentences are obtained by sentence segmentation, and then regular expressions and other rules are used to remove useless sentences in the news articles.

[0050] Among them, the trusted news websites include various official news websites.

[0051] For any two event sentences in each news, use Ernie (Wenxin) text classification model for three classification, to judge whether the two events are irrelevant, consecutive or causal relationship, and in order to increase the connection between events, the event sentences between two adjacent news in time are paired and classified.

[0052] The Ernie text classification model is spliced with a classification layer composed of a Dropout layer and a Linear layer after the Ernie model. The Ernie model represents the sentence information as a 768-dimensional sentence vector. After the vector passes through the Dropout layer, it is sent to the Linear layer with an input dimension of 768 and an output dimension of the number of label categories, that is, 3, for classification. The output of the Linear layer is subjected to a softmax operation to obtain the prediction probability of each label. The label with the maximum probability is taken as the classification result of the model. The formula of the whole classification process is as follows:

[0053]

[0054] wherein, is the prediction result, is the information of input_ids, token_type_ids, position_ids and attention_mask obtained after tokenizer processing of the sentence, is the weight of the Linear layer, is the bias of the Linear layer, is the Ernie model function, is the Dropout layer function, is the function of taking the index of the maximum value, is the function of normalizing the input to probability.

[0055] The event vectors in the event pair with causal and coherent relationship are represented, and then the event sentence and the event vector are taken as nodes and the relationship is taken as edges to import the neo4j database in the form of triplets to construct an event graph. In order to avoid the phenomenon of too many "dumbbells", that is, the event nodes are relatively isolated and cannot be connected into a network structure, the nodes with a cosine similarity of more than 0.8 are added with a similar edge as a relationship.

[0056] As an embodiment, the event vectorization in the present embodiment uses a Sentence Bert (SBert) model. The SBert is modified based on the Bert network. The network generates sentence embedding vectors with semantic meaning using a twin structure. The embedding vectors of sentences with similar semantics are relatively close in distance, so the SBert is more suitable for calculating semantic similarity by cosine similarity than the original Bert pre-training model. The model adopts a sentence pair classification network structure for training, adopts a MEAN strategy, that is, calculates the average value of each token output vector to represent the sentence vector, and adopts a cross-entropy loss function for optimization.

[0057] It should be noted that in addition to using the Sentence Bert (SBert) model, the sentence vectorization in this embodiment can also use BERT, ERNIE, SpanBERT, RoBERTa and other pre-training language models.

[0058] Step three: taking the event node with the maximum similarity as the starting node, querying the event nodes with multi-hop similarity and continuity relationship with the starting node, and forming a related event set composed of the starting node and the multi-hop relationship node. The related event set contains the preceding events related to the to-be-identified event sentence.

[0059] Taking the most similar node A found in the previous step as the starting node, querying the event nodes with 4-hop similarity and continuity relationship in the neo4j database, and forming a related event set S composed of the starting node and the 4-hop relationship node. S contains as many preceding events related to the event sentence Q as possible.

[0060] Step four: forming an event sequence for each event in the related event set and the to-be-identified event sentence, and performing abductive reasoning on each event sequence to obtain an explanation event set.

[0061] Forming an event sequence (S i , Q) for each event S i in the related event set obtained in the previous step and the to-be-identified event sentence Q, and performing abductive reasoning on each event sequence using a GRF (multi-hop reasoning flow generation) model. The model performs abductive reasoning on the input two events S i and Q to produce an explanation event H i that may occur between S i and Q, so that the event sequence (S i , H i , Q) is reasonable, and thus an explanation event set H can be obtained.

[0062] As an implementation, the GRF model generates knowledge-rich language by performing multi-hop reasoning on an external knowledge graph (conceptnet Chinese part). The model mainly includes a pre-trained Chinese GPT2 model for context modeling module, a static multi-relationship graph encoding module, a dynamic multi-hop reasoning flow module, and a generation distribution module with gating.

[0063] Step five: taking each explanation event in the explanation event set as an explanation event starting node, querying the event nodes with multi-hop similarity and causal relationship with the explanation event starting node, and forming a causal event set composed of the starting node and the multi-hop relationship node.

[0064] Taking each explanation event H i in the explanation event set H obtained in the previous step as its corresponding S iThe event node is the starting node. The Neo4j database is queried for event nodes that have a 4-hop similarity and causal relationship with it. A causal event set T is formed by the starting node and the nodes with the 4-hop relationship. If event T is present in T... i With H i If T has the maximum cosine similarity and the similarity is greater than 0.9, then it is considered to be... i With H i Describing the same event, indicating a sequence of events (S) i T i Q) Reasonable, that is, based on the already occurred event S i and T i If event Q is indeed possible, record Q as true and skip the evaluation of the remaining interpretations in H. If the maximum similarity is not greater than the threshold, continue to evaluate the next interpretation H. j If none of the explanations in H find a T with a similarity greater than the threshold, j If the result is false, it means that no event has occurred that could have caused event sentence Q to occur, and Q is recorded as false. Therefore, new news events can be identified and explained through abductive reasoning. While ensuring that the identification results have a certain degree of interpretability, it also solves the shortcomings of fact-based methods caused by the lag in updating the fact base to a certain extent.

[0065] Step Six: Determine the authenticity of an event based on the maximum cosine similarity between events in the explanatory event set and events in the causal event set, thus obtaining the result of determining the authenticity of the news to be identified.

[0066] Based on the identification results of the event sentences in the aforementioned steps, the authenticity of the news to be identified and its interpretation are given.

[0067] The abductive reasoning-based fake news identification method provided in this embodiment ensures that the identification results have a certain degree of interpretability. It can identify newly generated fake news by combining event graphs with the news to be identified through abductive reasoning. To a certain extent, it solves the shortcomings of fact-based methods that cannot identify newly occurring news due to the sluggish updating of the fact base.

[0068] Example 2

[0069] Reference Figure 3 ,like Figure 3 The diagram shown is a structural block diagram of the fake news identification device based on abductive reasoning provided in this embodiment. This device is used to implement the fake news identification method based on abductive reasoning provided in the aforementioned embodiment, and specifically includes the following structure:

[0070] The text preprocessing module is used to acquire the news text to be identified and perform preprocessing operations, including sentence segmentation and removal of useless words and sentences, to obtain one or more event sentences to be identified;

[0071] The similarity calculation module is configured to calculate the cosine similarity between the vectorized representation of the event sentence to be identified and all event vectors in the event graph, and find an event node with the maximum similarity;

[0072] The related event set construction module is configured to query event nodes with multi-hop similarity and successive relationship from the starting node, and to form a related event set from the starting node and the multi-hop relationship nodes, the related event set containing pre-events related to the event sentence to be identified.

[0073] The explanation event set construction module is configured to form an event sequence from each event in the related event set and the event sentence to be identified, and to perform abductive reasoning on each event sequence to obtain an explanation event set.

[0074] The causal event set construction module is configured to take each event node corresponding to an explanation event in the explanation event set as an explanation event starting node, query event nodes with multi-hop similarity and causal relationship from the explanation event starting node, and form a causal event set from the starting node and the multi-hop relationship nodes.

[0075] The authenticity judgment module is configured to judge the authenticity of the event according to the maximum cosine similarity between the events in the explanation event set and the events in the causal event set, and obtain the authenticity judgment result of the news to be identified.

[0076] The false news identification device based on abductive reasoning provided in the embodiment can identify newly generated false news by combining the event graph with the news to be identified for abductive reasoning, while ensuring that the identification result has a certain explainability, and to some extent, solves the problem that the fact-based method cannot identify newly occurring news due to the lag in updating the fact library.

[0077] Embodiment 3

[0078] The preferred embodiment provides a computer device that can implement the steps in any embodiment of the false news identification method based on abductive reasoning provided in the embodiments of the application, and thus can achieve the beneficial effects of the false news identification method based on abductive reasoning provided in the embodiments of the application. For details, see the previous embodiments, which will not be described here.

[0079] Embodiment 4

[0080] Those skilled in the art can understand that all or part of the steps of various methods of the above embodiments can be completed by instructions or by controlling relevant hardware by the instructions, which can be stored in a computer readable storage medium and loaded and executed by a processor. To this end, the embodiments of the present application provide a storage medium, wherein a plurality of instructions are stored, which can be loaded by a processor to execute the steps of any embodiment of the method for identifying fake news based on provenance reasoning provided by the embodiments of the present application.

[0081] The storage medium can include a read only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0082] Since the instructions stored in the storage medium can execute the steps in any embodiment of the method for identifying fake news based on provenance reasoning provided by the embodiments of the present application, the beneficial effects that can be achieved by any method for identifying fake news based on provenance reasoning provided by the embodiments of the present application can be achieved, which are described in detail in the foregoing embodiments and will not be described here.

[0083] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for false news identification based on traceability reasoning, characterized in that, The method comprises: acquiring a news text to be identified, performing a preprocessing operation, the preprocessing comprising sentence segmentation and removing useless words, to obtain one or more event sentences to be identified; vectorizing the event sentences to be identified and calculating the cosine similarity with all event vectors in an event graph to find an event node with the maximum similarity; querying event nodes with multi-hop similarity and sequential relationship from the event node with the maximum similarity as a starting node, to form a relevant event set comprising the starting node and multi-hop relationship nodes, the relevant event set comprising a preceding event related to the event sentence to be identified; forming an event sequence by each event in the relevant event set and the event sentence to be identified, and performing a causal reasoning on each event sequence to obtain an explanation event set; taking each event node corresponding to an explanation event in the explanation event set as an explanation event starting node, querying event nodes with multi-hop similarity and causal relationship from the explanation event starting node to form a causal event set comprising the starting node and multi-hop relationship nodes; judging the truth of an event according to the maximum cosine similarity between the event in the explanation event set and the event in the causal event set, to obtain a true or false judgment result of the news to be identified; the event graph is constructed in the following manner: acquiring real news samples, segmenting the real news samples to obtain independent event representation sentences, and removing useless word sentences in the real news samples; performing three classification on any two event representation sentences in the real news samples to judge whether the two events are irrelevant, sequential, or causal; vectorizing the events with causal and sequential relationships after classification, and then constructing an event graph in the form of a triple comprising the event sentences and event vectors as nodes and the relationships as edges; the false news identification method establishes a similarity edge between similar event nodes by calculating the cosine similarity between event vectors; the judgment of the truth of an event according to the maximum cosine similarity between the event in the explanation event set and the event in the causal event set specifically comprises: if the maximum cosine similarity between an event in the causal event set and an event in the explanation event set is greater than a similarity threshold, it is recorded as an event that may have occurred and the judgment of the remaining explanation events in the explanation event set is skipped; if the maximum cosine similarity between any event in the causal event set and any event in the explanation event set is not greater than the similarity threshold, it is recorded as a false event. 2.The method of claim 1, wherein, The method further comprises classifying event representation sentences between real news samples with a publication time interval less than a preset threshold. 3.The method of claim 1, wherein, The vectorization uses a pre-trained language model to generate sentence vectors with semantic meaning using a twin structure. 4.The method of claim 1, wherein, The causal reasoning uses a multi-hop reasoning flow generation model. 5.A false news recognition device based on traceability reasoning, characterized by, The device is used to implement the false news identification method based on causal reasoning according to claim 1, and the device comprises: a text preprocessing module configured to acquire a news text to be identified, perform a preprocessing operation, the preprocessing comprising sentence segmentation and removing useless words, to obtain one or more event sentences to be identified; The similarity calculation module is configured to calculate the cosine similarity between the vectorized representation of the event sentence to be identified and all event vectors in the event graph, and find an event node with the maximum similarity; The related event set construction module is configured to query event nodes with multi-hop similarity and continuity from a starting node with the maximum similarity, and to form a related event set from the starting node and the multi-hop relationship nodes, wherein the related event set contains pre-events related to the event sentence to be identified; The explanation event set construction module is configured to form an event sequence from each event in the related event set and the event sentence to be identified, and to perform a causal reasoning on each event sequence to obtain an explanation event set; The causal event set construction module is configured to take each event node corresponding to an explanation event in the explanation event set as an explanation event starting node, to query event nodes with multi-hop similarity and causal relationship from the explanation event starting node, and to form a causal event set from the starting node and the multi-hop relationship nodes; The authenticity judgment module is configured to judge the authenticity of events according to the maximum cosine similarity between events in the explanation event set and events in the causal event set, and to obtain an authenticity judgment result of the news to be identified.

6. A computer device, comprising: The computer device includes a processor and a memory, and the memory stores a computer program, which is loaded and executed by the processor to implement the method for identifying fake news based on causal reasoning according to any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which is loaded and executed by the processor to implement the method for identifying fake news based on causal reasoning according to any one of claims 1 to 4.

Images