A rumor identification method based on multi-source information fusion

Abstract

The present application relates to the field of artificial intelligence and network information security technology, and particularly relates to a rumor identification method based on multi-source information fusion. The method comprises: obtaining a text to be identified; obtaining content audit original labels returned by a plurality of third-party content audit platforms; mapping heterogeneous labels of different platforms to a unified clue category system to obtain a clue set based on platform content audit; using a rumor clue supplement vocabulary to identify language expression features of the text to be identified to obtain a clue set based on language features; calculating fusion scores of various clues according to the clue set based on platform content audit and the clue set based on language features, and generating a rumor clue representation vector according to a preset threshold; fusing a text representation of the text to be identified with the rumor clue representation vector, and inputting the fusion result into a rumor identification model to obtain a rumor identification result. The present application can unify mapping, supplement and fusion of multi-platform content audit labels and language expression features, reduce the influence of single information source label loss, label heterogeneity and semantic inconsistency on the identification result, and help to improve the accuracy and stability of rumor identification in the public opinion dissemination scenario.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and network information security technology, specifically to a rumor identification method based on multi-source information fusion. Background Technology

[0002] With the rapid development of social media, online information dissemination is characterized by speed, wide reach, and strong interactivity. Rumors spread online through text, images, and short videos, not only misleading public perception but also potentially triggering public opinion fluctuations and affecting social order and public safety. Therefore, rumor identification in the context of public opinion dissemination has significant application value.

[0003] Existing technologies mainly improve rumor recognition performance through text semantic modeling, multimodal fusion, and the introduction of auxiliary features, but they still have the following shortcomings in the context of public opinion dissemination: (1) Existing rumor identification tasks typically use true / false labels as the main supervisory information. However, in the actual process of public opinion dissemination, rumors do not always appear through explicit true / false statements, but often enhance their dissemination effect by using exaggerated wording, emotional inducement, fictional narratives, or low-quality expressions. Such expressions are not directly equivalent to rumor labels, but can provide auxiliary evidence for judging authenticity. Although existing methods can learn the overall semantic representation of the text and complete the true / false classification, the explicit utilization of the above-mentioned fine-grained expression signals is still insufficient, thus limiting the model's ability to identify complex expression patterns in public opinion dissemination scenarios.

[0004] (2) Some existing methods attempt to introduce auxiliary information such as sentiment tendency, keyword features, image-text consistency, and dissemination behavior to improve the effectiveness of rumor identification by reflecting the abnormality of information from different perspectives. However, existing methods mostly use the above information as discrete features or additional inputs, lacking a unified definition, semantic merging, and fusion mechanism for multi-source heterogeneous information, making it difficult to form a consistent auxiliary discrimination representation. Especially in the context of public opinion dissemination, exaggerated wording, inducing forwarding, and fictional narratives related to rumor dissemination are scattered across different sources. Existing technologies are still unable to systematically organize, uniformly represent, and effectively fuse such auxiliary discrimination information, thus affecting the accuracy and stability of rumor identification.

[0005] Therefore, there is an urgent need to provide a technical solution that can uniformly merge and integrate multi-source heterogeneous information and use it for rumor identification, so as to improve the effectiveness of rumor identification in public opinion dissemination scenarios. Summary of the Invention

[0006] To address the aforementioned problems, this invention proposes a rumor identification method based on multi-source information fusion. This method unifies the semantic representation of content moderation tags across multiple platforms and generates rumor clue representation vectors by combining linguistic expression features. These rumor clue representation vectors are then introduced as auxiliary information into the rumor identification model for use in public opinion dissemination scenarios. In this invention, rumor clues refer to feature information extracted from the text to be identified and its associated information, which can help determine whether the text is a rumor. This feature information is not directly equivalent to a rumor tag, but rather represents abnormal expressions, dissemination inducements, fictitious narratives, risk orientations, and platform moderation results related to rumor dissemination.

[0007] This invention aims to address the shortcomings of existing rumor detection methods in utilizing fine-grained signal representation and the lack of a unified definition, merging, and fusion mechanism for multi-source heterogeneous auxiliary information. The technical solution of this invention is as follows: Figure 1 As shown, it includes the following steps: S1. Obtain the text to be recognized. ; S2. Obtain the original tags for content review from multiple platforms, and set the text to be recognized. The content moderation tags on internet platforms are: in, Indicates the platform number. The original tags for the content output by the corresponding platform are reviewed. Indicates the first The first platform output The original tags; the content review platform is one or more third-party content review platforms; the original content review tags may include, but are not limited to, tags such as "low-quality spam", "abusive content", and "advertising traffic". S3. Map heterogeneous content moderation tags from multiple platforms to a unified clue category, specifically including: S31. Construct a unified clue category system and define a mapping function: in This indicates that the unified clue category system defines the first... Clues related to rumors; S32. Add content moderation tags Mapping to a unified clue category system yields a set of clues based on platform content review: S4. Utilize language expression features to supplement clues that are difficult to cover by platform tags, specifically including: S41. Read the supplementary word bank for rumor clues, targeting clue categories that require identification through language expression features: in Indicates the first The set of keywords corresponding to the category clues; S42. Identify language expression features. When the text t to be identified contains elements belonging to the keyword set... Keywords At that time, the text to be identified is determined. Includes corresponding rumor clues This yields a set of clues based on language features: S5. Merge multi-source rumor clues to generate a unified clue vector, specifically including: S51. For each type of rumor clue Based on the set of clues from the platform's content review and sets of clues based on language features The clue fusion score is calculated according to the preset weights. The preset weights are pre-set based on the credibility of tags and the importance of language expression features on different content review platforms; S52. Generate a cue representation vector, when the cue fusion score is... Exceeding the preset threshold At that time, the corresponding clues to the rumor were identified. This is valid, and we obtain the rumor clue representation vector: S6. The text representation of the text to be identified is concatenated, weighted and fused or jointly encoded with the rumor clue representation vector, and then input into the rumor recognition model to obtain the rumor recognition result.

[0008] The beneficial effects of this invention are: (1) By uniformly mapping, supplementing and integrating the content review tags and language expression features of multiple platforms, this invention can form a consistent auxiliary feature representation, reduce the impact of missing tags, heterogeneous tags and semantic inconsistencies of a single information source on the rumor identification results, thereby improving the accuracy and robustness of rumor identification in the context of public opinion dissemination. (2) The rumor clue representation vector generated by this invention can be used as a reusable auxiliary feature to access different types of rumor recognition models, and has good model compatibility and scalability; (3) This invention can be applied to scenarios such as content review on social platforms, network information governance, public opinion monitoring, risk warning and network information security protection, providing more stable and accurate technical support for the identification of false information in the process of public opinion dissemination. Attached Figure Description

[0009] Figure 1 This is a flowchart of the rumor identification method based on multi-source information fusion of the present invention.

[0010] Figure 2 This is an embodiment of the unified clue category system of the present invention.

[0011] Figure 3 This is an embodiment of the rumor clue supplementary vocabulary database of the present invention.

[0012] Figure 4 This is a comparison of the accuracy of the rumor identification model using this invention with traditional methods on publicly available rumor datasets. Detailed Implementation

[0013] The present invention will be further described below with reference to specific embodiments to enable those skilled in the art to better understand and implement the present invention. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Various modifications, substitutions, or improvements can be made to the present invention by those skilled in the art without departing from the spirit and substance of the invention, and all such modifications, substitutions, or improvements should fall within the scope of protection of the present invention.

[0014] Example 1

[0015] This embodiment provides a rumor identification method based on multi-source information fusion. The method uniformly maps content moderation tags returned by multiple third-party content moderation platforms, constructs a rumor clue representation vector by combining linguistic expression features, and then fuses the rumor clue representation vector with the text representation of the text to be identified to obtain the rumor identification result. This method can be applied to scenarios such as rumor identification, content moderation, network information governance, public opinion monitoring, and risk warning in public opinion dissemination scenarios.

[0016] The process in this embodiment includes the following steps: S1. Obtain the text to be identified t, which may be post text, news headline, news body, comment text, forwarded postscript, or other online text content to be identified from a social media platform; S2. Obtain the text to be recognized from multiple third-party content review platforms. The returned original content moderation tags, in this embodiment, refer to the third-party content moderation platforms as Baidu, Alibaba Cloud, and Tencent Cloud. For the text to be identified... , No. The original set of content review tags returned by each platform is represented as follows: in, Indicates the platform number. This refers to the original tag set for content review output by the corresponding platform. Indicates the first The first platform output The original tag; in the specific implementation, the text to be recognized can be obtained by calling the review interface provided by various third-party content review platforms. Send to each platform and receive the review tag results returned by each platform to form a set of original tags for content review across multiple platforms; S3. Map heterogeneous content moderation tags from multiple platforms to a unified clue category, specifically including: S31. Construct a unified clue category system and define a mapping function: in This indicates that the unified clue category system defines the first... Rumor-related clues; In this embodiment, the unified clue category system includes low-quality information clues, sensitive content clues, commercial guidance clues, aggressive expression clues, privacy leakage clues, political and ideological clues, prohibited language clues, violence and terrorism clues, negative tendency clues, and normal content categories, etc., for example... Figure 2 As shown; the above unified clue category system is only an example. The number of unified clue categories, names, and the merging relationship of tags from various platforms can be added, deleted, and adjusted according to specific application scenarios. S32. According to the mapping function Add content moderation tags Mapping to a unified clue category system yields a set of clues based on platform content review: in, Indicates the first A set of clues obtained by the platform based on content moderation tag mapping; that is, if the first... Any original tag returned by the platform is mapped to the first... If the clues are consistent across categories, then it is believed that the platform treats the text to be identified as... Hit the clue ; S4. Utilize language expression features to supplement clues that are difficult to cover by platform tags, specifically including: S41. Read the rumor clue supplementary vocabulary and construct a rumor clue supplementary vocabulary for the categories of clues that need to be identified through language expression features: in Indicates the first The set of keywords corresponding to category clues; in this embodiment, a supplementary thesaurus is constructed for exaggerated expression clues and fictional narrative clues, for example... Figure 3 As shown; the above-mentioned vocabulary is only an example. In other implementations, the keyword set can be expanded or adjusted according to the actual application scenario, or phrase dictionaries, rule templates, regular expressions, syntactic patterns and other methods can be used to supplement the identification of clues. S42. Identify language expression features. When the text t to be identified contains elements belonging to the keyword set... Keywords At that time, the text to be identified is determined. Includes corresponding rumor clues This yields a set of clues based on language features: In the specific implementation, the text to be identified can first be segmented, stop words removed, and normalized, and then keyword matching can be performed based on a preset thesaurus to determine the category of clues that the text hits. S5. Merge multi-source rumor clues to generate a unified clue vector, specifically including: S51. For each type of rumor clue Based on the set of clues from the platform's content review and sets of clues based on language features Calculate the clue fusion score: in, Indicates the first The weight corresponding to each platform; Weights corresponding to language expression features; This is an indicator function; it takes a value of 1 when the condition within the parentheses is true, and a value of 0 otherwise; the weight... and It can be pre-set based on the credibility of tags on different content review platforms, the platform's review capabilities, the importance of language features, historical verification results, or empirical rules. For example, higher weights can be set for platforms with higher review accuracy or wider coverage. S52. Generate a cue representation vector, when the cue fusion score is... Exceeding the preset threshold At that time, the corresponding clues to the rumor were identified. Based on whether each unified clue category holds true, construct a rumor clue representation vector: in, This represents the total number of clue categories in the unified clue category system. Indicates the text to be recognized The rumor clue representation vector, when the rumor clue At the time of its establishment, The value is 1 if it is set to 1, otherwise the value is 0. S6. The text representation of the text to be identified is concatenated, weighted and fused or jointly encoded with the rumor clue representation vector, and then input into the rumor recognition model to obtain the rumor recognition result.

[0017] Example 2 To verify the effectiveness of the method of this invention, experiments were conducted on the publicly available rumor datasets CHECKED and Weibo. CHECKED is a Chinese COVID-19 rumor dataset, and Weibo is a Chinese fake news detection dataset. The rumor cue representation vectors generated by this invention were fused with the text representations of TextCNN, BiLSTM, and BERT models, respectively, and compared with the corresponding baseline models that did not incorporate rumor cue representation vectors. The experimental results are as follows: Figure 4 As shown, after introducing the rumor clue representation vector (+Clue) constructed in this invention, the recognition performance of each model is improved, indicating that the method of this invention can effectively supplement the overall semantic representation of the text and improve the accuracy of rumor recognition.

[0018] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A rumor identification method based on multi-source information fusion, characterized in that, Includes the following steps: S1. Obtain the text to be recognized; S2. Obtain the original content review tags returned by multiple third-party content review platforms for the text to be identified; S3. Map the original content review tags to a unified clue category system to obtain a clue set based on platform content review; S4. Use language expression features to identify the text to be identified, and obtain a set of clues based on language features; S5. Based on the clue set based on platform content review and the clue set based on language features, perform fusion processing on various clues to generate a rumor clue representation vector; S6. The text representation of the text to be identified is fused with the rumor clue representation vector, and then input into the rumor identification model to obtain the rumor identification result of the text to be identified. The rumor clues are feature information extracted from the text to be identified and its associated information, used to help determine whether the text is a rumor. The feature information is not directly equivalent to the rumor label, but is used to characterize at least one of the following related to the spread of rumors: abnormal expression, spread inducement, fictional narrative, risk tendency and platform review results.

2. The rumor identification method based on multi-source information fusion according to claim 1, characterized in that, The process of mapping the original content review tags to a unified clue category system to obtain a clue set based on platform content review includes: Construct a unified clue category system and establish a mapping relationship between the original content review tags and the unified clue categories; Based on the mapping relationship, the original content review tags returned by each third-party content review platform are mapped to the unified clue category system to obtain the platform-level clue set corresponding to each platform.

3. The rumor identification method based on multi-source information fusion according to claim 1, characterized in that, The process of identifying the text to be identified using linguistic expression features to obtain a set of clues based on linguistic features includes: Construct a supplementary vocabulary for rumor clues, which includes a set of keywords corresponding to one or more clue categories; Perform keyword matching on the text to be identified; When the text to be identified contains keywords belonging to a certain clue category, it is determined that the text to be identified contains a rumor clue of the corresponding category, and a clue set based on language features is obtained.

4. The rumor identification method based on multi-source information fusion according to claim 1, characterized in that, The step involves fusing various clues based on the platform content review-based clue set and the language feature-based clue set to generate a rumor clue representation vector, including: For each type of clue, a corresponding clue fusion score is calculated based on the clue set based on platform content review and the clue set based on language features. The clue fusion score is compared with a preset threshold. When the clue fusion score corresponding to a certain type of clue reaches or exceeds the preset threshold, the clue is determined to be valid. Based on whether the various clues are valid, a rumor clue representation vector is generated.

5. A rumor identification device based on multi-source information fusion, characterized in that, include: The text tag acquisition module is used to acquire the text to be identified and to acquire the original content review tags returned by multiple third-party content review platforms for the text to be identified. The tag mapping module is used to map the original tags of the content review to a unified clue category system to obtain a set of clues based on the platform's content review. The language feature recognition module is used to recognize the text to be recognized by utilizing language expression features, and to obtain a set of clues based on language features; The multi-source clue fusion module is used to fuse various clues based on the clue set based on platform content review and the clue set based on language features to generate a rumor clue representation vector. The rumor identification module is used to fuse the text representation of the text to be identified with the rumor clue representation vector, and input it into the rumor identification model to obtain the rumor identification result of the text to be identified.

6. The apparatus according to claim 8, characterized in that, The multi-source clue fusion module is specifically used for: calculating the corresponding clue fusion score for each type of clue based on the clue set based on platform content review and the clue set based on language features; and comparing the clue fusion score with a preset threshold. When the clue fusion score corresponding to a certain type of clue reaches or exceeds the preset threshold, the clue of that type is determined to be valid, and the rumor clue representation vector is generated based on whether the various types of clues are valid.

7. An electronic device, characterized in that, include: At least one processor; At least one memory, wherein a computer program is stored; When the computer program is executed by the at least one processor, the electronic device performs the rumor identification method based on multi-source information fusion as described in any one of claims 1-7.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the rumor identification method based on multi-source information fusion as described in any one of claims 1-7.

Images