Stance detection in tweets: A topic modeling approach supporting explainability

Stance detection improves fake information recognition in social media. This task encourages interpreting and explaining the misinformation identification, thus aligning with the importance of improving human trust in classification results. Nonetheless, most of the stance detection studies do not engage in understanding the reasons behind a solution. We propose a two-phase classification system for stance detection in tweets. We mainly exploit topic modeling features. Our proposal is remarkably different from previous ones since we provide an explanation of stance labels through the most relevant terms within topics over the tweets. Therefore, our approach is flexible because it adjusts to the vocabulary leveraging topic information. We additionally construct sets of features seeking tweet-specific content, sentiment and subjectivity markups, target attributes, term dis-tribution, and word embeddings. Our classification system ranks second in the state-of-the-art regarding SemEval-2016 task 6 dataset with a 74.63% overall F-measure. To the best of our knowledge, our results are superior to deep learning-based proposals and competitive with studies that do not provide an explanation for stance labels. Hence, we affirm that topic modeling features enhanced classification results and provided textual information about a plausible explanation of stance labels. We report the performance of our system considering variations in the feature set, besides discussing the explanation of the results.

Automated summary

Stance detection plays a crucial role in recognizing fake information in social media. Most previous studies on stance detection focus on classification results without providing explanations. In this paper, a two-phase classification system is proposed for stance detection in tweets, utilizing topic modeling features and explaining stance labels through relevant terms within topics. The approach is flexible and adjusts to vocabulary leveraging topic information. The system's performance ranks second in the SemEval-2016 task 6 dataset, outperforming deep learning-based proposals. The results affirm that topic modeling features improve classification results and provide textual information about stance labels.