CN-121998630-A - Social function identification method and system based on field perception and semantic conflict detection

CN121998630ACN 121998630 ACN121998630 ACN 121998630ACN-121998630-A

Abstract

The invention provides a social function identification method and a system based on field perception and semantic conflict detection, relates to the technical fields of multi-mode content understanding, computation social science and content function identification, and aims at solving the problems that the prior art has multi-mode semantic conflict identification failure, social field context deletion and uncontrollable generated model input. The method comprises the steps of obtaining social media content to be identified, carrying out multi-view coding and self-adaptive gating fusion on the social media content, calling different expert agents through a neural routing network to carry out reasoning based on fusion feature vectors and gating weights, generating a semantic decision vector, matching the semantic decision vector with a pre-constructed social function anchor point library, determining a target anchor point, carrying out bias correction on vocabulary probability related to the target anchor point by adopting Logits bias intervention algorithm, and outputting a function label and an interpretation text. The invention solves the problems existing in the prior art and realizes the automation and high-certainty identification required by the wind control of the industrial content.

Inventors

YU CHUAN
YAN SHU
LIU ZHAOXI
LIU YUANYUAN
Zhong Conghan

Assignees

青岛城市学院

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (10)

1. The social function identification method based on the field perception and the semantic conflict detection is characterized by comprising the following steps: acquiring social media content to be identified, wherein the social media content comprises an original image, a text and social interaction field information; performing multi-view coding and self-adaptive gating fusion on social media content to be identified to obtain fusion feature vectors and gating weights; based on the fusion feature vector and the gating weight, different expert agents in the mixed expert module are called through a neural routing network to conduct reasoning, and a semantic judgment vector is generated based on a reasoning result; performing similarity matching on the semantic decision vector and a pre-constructed social function anchor point library, and determining a target anchor point and a corresponding function category index thereof; inputting the target anchor point as constraint condition to a decoding stage of the visual language model, and adopting Logits bias intervention algorithm to carry out bias correction on vocabulary probability related to the target anchor point; under the constraint of the target anchor point, outputting the function label and the interpretation text.
2. The social function identification method based on field perception and semantic conflict detection as claimed in claim 1, wherein the steps of carrying out multi-view coding and adaptive gating fusion on the social media content to be identified are as follows: inputting the original image to a visual encoder, and extracting visual characteristics; inputting the text to a text encoder, and extracting text characteristics; Inputting social interaction field information into a context encoder, and extracting social interaction field characteristics; and inputting the visual features, the text features and the social interaction field features into the self-adaptive gating for fusion, and obtaining fusion feature vectors and gating weights.
3. The social function identification method based on field awareness and semantic conflict detection according to claim 2, wherein the social interaction field features are extracted, and the social interaction field features comprise explicit circle layer vectors, interaction portrait vectors and state atmosphere vectors, and the specific process is as follows: Embedding and mapping the dialogue question label or the super-dialogue ID to generate a dominant circle layer vector, and selecting a general default vector if the dominant circle layer vector is missing; obtaining an interactive image vector based on the historical content of the publisher; And obtaining the popular comment abstract of the similar content by utilizing a retrieval enhancement generation technology, and encoding to obtain a state atmosphere vector.
4. The social function recognition method based on field perception and semantic conflict detection according to claim 1, wherein in the self-adaptive gating batch training, a supervision contrast loss function is constructed, and a hard negative sample which satisfies that the visual or text similarity is higher than a preset threshold and social function labels are opposite is separated and optimized.
5. The social function identification method based on field awareness and semantic conflict detection according to claim 4, wherein the supervised contrast loss function formula is: ; Wherein, the The temperature coefficient is used for controlling the attention degree of the model to the difficult sample; representing sample indexes in the current lot; for the sample Is a fusion feature vector of (1); The fusion feature vector of the sample j; a fusion feature vector for sample p; As a positive sample set for sample i, Is the positive number of samples; a candidate sample set for participating in comparison; Representing candidate sample indices.
6. The social function identification method based on field awareness and semantic conflict detection according to claim 1, wherein the different expert agents in the hybrid expert module comprise: a literal expert agent whose roles are defined to describe visual elements of the image and to transcribe text content; A collision expert agent whose role is defined to detect emotional or semantic collision between the image and the text; the role of a domain expert agent is defined to explain the underlying meaning of content based on social circle layers and publisher settings.
7. The social function identification method based on the field awareness and the semantic conflict detection according to claim 1, wherein different expert agents in the hybrid expert module are called through a neural routing network to conduct reasoning, and the specific process is as follows: Based on the fusion feature vector and the gating weight, calculating a credibility weight vector of each expert agent; calculating the activation probability of each expert agent according to the credibility weight vector; And only when the activation probability of the expert agent is larger than a preset activation threshold, calling the corresponding expert agent to conduct reasoning.
8. The social function recognition method based on field awareness and semantic conflict detection according to claim 1, wherein the method is characterized in that a Logits bias intervention algorithm is adopted to carry out bias correction on vocabulary probability related to a target anchor point, and comprises the following specific processes: calculating semantic drift degree between the current decoding hidden state and the target anchor point; And carrying out bias correction on the vocabulary probability related to the target anchor point based on the semantic drift degree and the preset intervention strength, and generating biased corrected probability distribution.
9. The social function identification method based on field awareness and semantic conflict detection according to claim 8, wherein the probability distribution formula after bias correction is: ; Wherein, the In order to intervene in the intensity super-parameter, =5.0; To only lock anchor points A related specific security vocabulary, i.e. a vocabulary related to the target anchor point; Original Logits; the semantic drift degree is corresponding to the semantic drift degree; outputting vocabulary for candidates of the decoding stage time t; For bias corrected candidate output vocabulary Is a probability of (2).
10. The social function recognition system based on the field perception and the semantic conflict detection is characterized by comprising the following components: The perception and feature fusion module is used for acquiring social media content to be identified, wherein the social media content comprises an original image, a text and social interaction field information; The cognitive reasoning module is used for calling different expert agents in the hybrid expert module to conduct reasoning through the neural routing network based on the fusion feature vector and the gating weight, and generating a semantic judgment vector based on a reasoning result; The expression output module is used for carrying out similarity matching on the semantic decision vector and a pre-constructed social function anchor point library to determine a target anchor point and a corresponding function category index thereof, inputting the target anchor point as a constraint condition into a decoding stage of the visual language model, carrying out bias correction on vocabulary probability related to the target anchor point by adopting Logits bias intervention algorithm, and outputting a function label and an interpretation text under the constraint of the target anchor point.

Description

Social function identification method and system based on field perception and semantic conflict detection Technical Field The invention belongs to the technical fields of multi-mode content understanding, computing social science and content function identification, and particularly relates to a social function identification method and system based on field perception and semantic conflict detection. Background The statements in this section merely provide background information related to the present disclosure and may not necessarily constitute prior art. With the deep penetration of social media platforms, user-generated content exhibits highly multi-modal, strongly context-dependent, and rapidly evolving features. Typical social media content often contains information such as images, text, and even emoticons, topic labels, forwarding/commentary relationships, and the like. The true meaning of the system not only depends on the content, but also is commonly influenced by complex social interaction fields such as publishing community plates, user identity attributes, platform cultural atmosphere, popular speaking systems and the like. Therefore, the automatic and high-certainty recognition of social functions (such as information transmission, emotion releasing, group identification, irony/black humor and the like) carried by social contents is realized, and the method has become a key technology in the fields of social calculation and content auditing. However, in the prior art, when social contents such as graphics and texts irony are processed, certain defects still exist, firstly, the existing model mainly depends on a shallow feature splicing mechanism, when the scenes such as Meme, irony or metaphors and the like with inconsistent graphics and texts and ideas are faced, due to the fact that the important semantics are aligned and the structured perception of the social field is lacking, the deep semantic conflict high-frequency misjudgment is difficult to capture, secondly, the existing method often ignores key context features such as community culture, user settings and public opinion atmosphere and the like, so that semantic understanding is separated from an actual propagation scene, and the model performs one-sided reasoning under the condition of lacking a context anchor point, thirdly, the problem of uncertainty in output, semantic illusion and unexplainability exists in a large Visual Language Model (VLM) based on probability sampling, the result of the same input cannot be reproduced, and strict requirements of industrial content wind control and speech analysis on the result certainty and compliance basis are difficult to meet. In summary, the prior art has the problems of failure in multi-mode semantic conflict recognition, missing social field context and uncontrollable input of a generated model. Disclosure of Invention In order to overcome the defects of the prior art, the invention provides the social function identification method and the system based on the field perception and the semantic conflict detection, which solve the problems of semantic conflict misjudgment, context deletion and uncontrollable output faced by the prior art when the social content is processed by combining scene depth perception, expert collaborative reasoning and anchor point stability decision, and realize the automation and high-certainty identification required by industrial content wind control. To achieve the above object, one or more embodiments of the present invention provide the following technical solutions: The first aspect of the invention provides a social function identification method based on field awareness and semantic conflict detection, which comprises the following steps: acquiring social media content to be identified, wherein the social media content comprises an original image, a text and social interaction field information; performing multi-view coding and self-adaptive gating fusion on social media content to be identified to obtain fusion feature vectors and gating weights; based on the fusion feature vector and the gating weight, different expert agents in the mixed expert module are called through a neural routing network to conduct reasoning, and a semantic judgment vector is generated based on a reasoning result; performing similarity matching on the semantic decision vector and a pre-constructed social function anchor point library, and determining a target anchor point and a corresponding function category index thereof; inputting the target anchor point as constraint condition to a decoding stage of the visual language model, and adopting Logits bias intervention algorithm to carry out bias correction on vocabulary probability related to the target anchor point; under the constraint of the target anchor point, outputting the function label and the interpretation text. As an implementation mode, the multi-view coding and the self-adaptive gating fusion are carried out o