Research On Technologies For Text Stance Detection On Social Media Platforms

In recent years,people have posted a massive amount of subjective text content on social media platforms,and mining this content has significant economic,social,and political value.In order to achieve automated understanding of this massive amount of subjective text,scholars have conducted extensive and in-depth research on sentiment analysis tasks.Among them,stance detection is currently one of the hot research topics,which aims to discriminate the stance tendency of text expressed towards specific targets,topics,or opinions.The core problem of stance detection is how to effectively learn the semantic relationship between the target and the comment text.In existing stance detection datasets,the number of targets is usually limited,the length is relatively short,and there are highly correlated features with stance labels in the comment text(biased features).Existing stance detection models generally model the target in simple ways and often ignore the interdependence between comment texts and targets,and are also prone to rely solely on biased features in the comment text for stance prediction,further reducing the modeling of the semantic relationship between the target and the comment text.These models do not perform the task according to the definition of stance detection and have poor generalization performance on data containing unknown targets.To address the above challenges,this paper draws on human reasoning processes and models the interdependence of <target,comment text>,introduces the intrinsic task knowledge of stance reasoning,and external background knowledge provided by large language models,to study and solve the incapability of modeling complex text based targets and ignorance of target information when modeling stance in existing works.This paper focuses on the text stance detection task in social media and conducts research in the following four aspects:1.Multi-turn Interaction Modeling-Based Stance Detection.To address the problem of existing stance detection models’ simple modeling of targets and the ignorance of the semantic information of comment text in modeling targets,this paper proposes a stance detection model based on recurrent interdependence modeling that explicitly learns stance features involving the interaction between targets and comment text.This method alternately depends on comment text information to model targets and target information to model comment text in the semantic interaction process.Experimental results show that the proposed method achieves significant performance improvement on traditional social media Twitter stance detection datasets and community QA stance detection datasets.2.Incorporating Stance Reasoning Process for Stance Detection.To address the problem that existing stance detection models often only fit the biased features of the comment text while ignoring the information of the target,this paper proposes a stance detection method that integrates the intermediate tasks of stance reasoning to model the features of the interaction between the target and the comment text.Inspired by the process of human expert analysis of stance,this paper defines these features as ”effective for both stance classification and intermediate reasoning subtasks” and introduces two feasible and simplified stance reasoning subtasks to assist in learning the aforementioned interaction features.In addition,this paper constructs multiple challenging test sets to evaluate the model’s understanding of the stance detection task from multiple perspectives.Compared with the baseline model,the proposed method further improves the performance on out-of-distribution and challenging data sets while maintaining the performance on in-distribution data.3.Counterfactual Reasoning for Stance Detection.To address the problem that existing stance detection models tend to ignore the information of the target and rely only on biased features from the comment text for stance prediction,this paper proposes a stance detection method based on adversarial bias learning and counterfactual reasoning,which models features that do not include the interaction between the two.The biased features that are ”effective for stance classification but not for intermediate reasoning subtasks”are treated as features that do not include the interaction between the two,and adversarial learning is used to model these features.Based on this,the counterfactual reasoning method is applied to estimate and eliminate the impact of such features on stance prediction.The experiment shows that the proposed method models the biased features more accurately,and the counterfactual reasoning method plays a positive role in eliminating the impact of biased features,and also improves the performance on both in-distribution and generalization data compared to baseline models.4.Incorporating Chain-of-Thoughts Knowledge for Stance Detection.In response to the problem that the comment text may lack the background knowledge necessary to establish semantic associations with the target,this article proposes a method to obtain missing background knowledge using large language models,which supplement the content of the interaction between the comment text and the target.It combines the reasoning chain inference knowledge generated by the large model with the intermediate steps of stance inference,providing fine-grained supervision signals to the stance inference model,thereby improving the reliability and interpretability of stance inference.On benchmark data of stance detection tasks,the proposed method only needs to use a small amount of data to surpass the baseline model trained on the entire data.In summary,this paper investigates methods for enhancing the semantic interaction between comments and targets in stance detection.The paper explicitly integrates the stance reasoning process and introduces bidirectional semantic dependencies between targets and comments,reducing the model’s dependence on biased features in the dataset and improving the model’s generalization performance and stance reasoning ability.In the era of large models,the paper aligns the knowledge of the large model’s thinking chain with the stance reasoning process,further improving the performance of the stance detection model.