CN-121996793-A - User feedback detection method and device, electronic equipment and storage medium

Abstract

The application provides a user feedback detection method which comprises the steps of modeling preset language information according to a causal framework to obtain a structural causal graph. The structural causal graph is represented by the following formula: Among them, Representing the semantic content factors in the causal architecture, Representing the semantic style factors in the causal architecture, The text information included in the language information is represented, The tag information included in the language information is represented. And carrying out content focusing on the language information according to a preset large language model to obtain focusing representation. Modeling is carried out according to the focusing representation and the structural causal graph, and a causal decoupling model is obtained. And detecting user feedback according to the causal decoupling model to obtain a detection result. Thus, the language model can accurately identify the feedback of the user in the multi-style scenes such as dialect, slang, nonstandard sentence pattern and the like.

Inventors

• SHUANG JIE
• LI RUILING

Assignees

• 北京邮电大学

Dates

Publication Date

20260508

Application Date

20251209

Claims (10)

1. 1. A user feedback detection method, comprising: modeling preset language information according to a causal framework to obtain a structural causal graph, wherein the structural causal graph is represented by the following formula: ; Wherein, the Representing semantic content factors in the causal framework, Representing semantic style factors in the causal architecture, Representing text information included in the language information, Representing tag information included in the language information; performing content focusing on the language information according to a preset large language model to obtain focusing representation; modeling according to the focusing representation and the structural causal graph to obtain a causal decoupling model; and detecting user feedback according to the causal decoupling model to obtain a detection result.
2. 2. The user feedback detection method according to claim 1, wherein content focusing is performed on the language information according to a preset large language model to obtain a focused representation, and specifically includes: performing context learning on the text information according to the large language model to obtain middle layer output; Calculating the sensitivity of each token in the intermediate layer output to loss to obtain token characterization; determining the focus representation from the token representation and the intermediate layer output.
3. 3. The user feedback detection method according to claim 1, wherein the modeling according to the focus representation and the structural causal graph results in a causal decoupling model, specifically comprising: Mapping the structural causal graph according to the focusing representation to obtain latent variable information; And determining the causal decoupling model according to the latent variable information and the large language model.
4. 4. The user feedback detection method of claim 3, wherein the mapping the structural causal graph according to the focus representation to obtain latent variable information specifically comprises: Mapping the content factors according to the focusing representation to obtain a content latent variable; mapping the style factors according to the focusing representation to obtain style latent variables; And optimizing the content latent variable and the style latent variable according to the preset task prediction loss to obtain the latent variable information.
5. 5. The user feedback detection method of claim 4, wherein the latent variable information includes an optimized content latent variable and an optimized style latent variable; determining the causal decoupling model according to the latent variable information and the large language model specifically comprises the following steps: Mapping the optimized content latent variable according to a preset latent variable injection module to obtain a mapped content latent variable; Mapping the optimization style latent variable according to the latent variable injection module to obtain a mapping style latent variable; And injecting the mapping content latent variable and the mapping style latent variable into the large language model according to a residual mode to obtain the causal decoupling model.
6. 6. A user feedback detection apparatus, comprising: the causal graph construction module is used for modeling preset language information according to a causal framework to obtain a structural causal graph, wherein the structural causal graph is represented by the following formula: ; Wherein, the Representing semantic content factors in the causal framework, Representing semantic style factors in the causal architecture, Representing text information included in the language information, Representing tag information included in the language information; The focusing module is used for focusing the content of the language information according to a preset large language model to obtain focusing representation; the modeling module is used for modeling according to the focusing representation and the structural causal graph to obtain a causal decoupling model; And the detection module is used for detecting the user feedback according to the causal decoupling model to obtain a detection result.
7. 7. The user feedback detection apparatus of claim 6, wherein the focusing module specifically comprises: The characterization module is used for performing context learning on the text information according to the large language model to obtain middle-layer output; Calculating the sensitivity of each token in the intermediate layer output to loss to obtain token characterization; determining the focus representation from the token representation and the intermediate layer output.
8. 8. The user feedback detection apparatus of claim 6, wherein the modeling module specifically comprises: The mapping module is used for mapping the structural causal graph according to the focusing representation to obtain latent variable information; And determining the causal decoupling model according to the latent variable information and the large language model.
9. 9. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of any one of claims 1 to 5 when the program is executed.
10. 10. A non-transitory computer readable storage medium storing computer instructions for causing a computer to perform the method of any one of claims 1 to 5.

Description

User feedback detection method and device, electronic equipment and storage medium Technical Field The application relates to the technical field of semantic recognition, in particular to a user feedback detection method. Background With the rapid development of social media, online customer service and electronic commerce comment platforms, text expression forms of users in public or interactive scenes show a trend of increasing diversification, non-standardization and individuation. Users of different ages, regions or communities often have specific language style features, such as spoken expressions, dialect words, network slang, expression symbols, etc., when expressing opinions, moods or complaints. The widespread existence of these stylized languages makes traditional statistical-relevance-dependent text classification models face serious fairness and robustness challenges in user negative feedback detection. The main user feedback analysis method at present is mostly based on large-scale pre-training language models (such as BERT, roBERTa and the like) to carry out downstream fine adjustment, and adopts a supervision learning paradigm to complete the judgment of positive/negative feedback or complaint category. However, such models typically employ end-to-end feature compression and representation learning strategies, lack explicit structured attribution mechanisms, and cannot effectively distinguish semantic content factors that truly express emotion or disfavor in text from style features that reflect only personal expression habits. In the case of structural deviations in the training data distribution, such as more frequent occurrence of certain dialects, regional expressions or spoken sentence patterns in negative samples, the model is prone to learn pseudo-correlations "between language styles and negative labels" to systematically bias specific populations or expression styles during the inference phase, even misjudging common neutral complaints or advice as negative complaints. Disclosure of Invention In view of the foregoing, the present application aims to provide a user feedback detection method, a device, an electronic apparatus and a storage medium. Based on the above purpose, the application provides a user feedback detection method, which comprises the steps of modeling preset language information according to a causal framework to obtain a structural causal graph. The structural causal graph is represented by the following formula: . Wherein, the Representing the semantic content factors in the causal architecture,Representing the semantic style factors in the causal architecture,The text information included in the language information is represented,The tag information included in the language information is represented. And carrying out content focusing on the language information according to a preset large language model to obtain focusing representation. Modeling is carried out according to the focusing representation and the structural causal graph, and a causal decoupling model is obtained. And detecting user feedback according to the causal decoupling model to obtain a detection result. In some embodiments, content focusing is performed on language information according to a preset large language model to obtain focusing representation, and specifically includes performing context learning on text information according to the large language model to obtain middle layer output. And calculating the sensitivity of each token in the intermediate layer output to loss, and obtaining token characterization. A focus representation is determined from the token representation and the intermediate layer output. In some embodiments, modeling is performed according to the focus representation and the structural causal graph to obtain a causal decoupling model, and specifically comprises mapping the structural causal graph according to the focus representation to obtain latent variable information. And determining a causal decoupling model according to the latent variable information and the large language model. In some embodiments, the latent variable information is obtained by mapping the structural causal graph according to the focus representation, and specifically comprises the step of obtaining the content latent variable by mapping the content factor according to the focus representation. And mapping the style factors according to the focusing representation to obtain style latent variables. And optimizing the content latent variable and the style latent variable according to the preset task prediction loss to obtain the latent variable information. In some embodiments, the latent variable information includes an optimization content latent variable and an optimization style latent variable. Determining a causal decoupling model according to the latent variable information and the large language model, wherein the causal decoupling model specifically comprises the steps of mapping