Unsupervised User Stance Detection on Tweets Against Web Articles Using Sentence Transformers

Bhashithe Abeysinghe, Gyandeep Reddy Vulupala, Anu G. Bourgeois, Rajshekhar Sunderraman

Abstract: Social media is an integral part of our daily lives, allowing users share opinions and content easily. This has been beneficial in many ways, but has also become a source for sharing misinformation and disinformation. Users can take different stances on various news or story articles that are being shared online. Predicting or detecting the stance of a user is a challenging problem in the Natural Language Processing and Inference circles, yet there is little work on the disinformation front. Infering users stance on disinformation is an important during a global pandemic. In this work we use a self-curated dataset to detect the stance of a Twitter user, based on a tweet and an article that they shared along with the tweet. Our process of collecting and curating the large dataset involves parallel computing methods to improve the timing performance. We present a novel unsupervised method for user stance detection that involves three phases: 1) embedding text from articles and tweets using a sentence transformer, 2) MeanShift clustering the article embeddings to determine themes, and 3) performing cosine similarity to find the semantic similarity between the tweet and article. Results of our methodology are promising and provide multiple directions for future work.