Analysis on the TGA Model for Stance Detection
Abstract
Stance detection, a problem concerned with finding the stance that an author takes on a specific issue, is a large subset of NLP and A.I, and its uses can already be seen in a multitude of applications. The majority of stance detection machine learning models are tested against a popular dataset called SemEval2016, which is a collection of tweets, authors, topics and stances that were derived from Twitter data and the Twitter API. Many researchers across the globe have created machine learning models to accurately predict the stance of authors based on their tweets regarding a certain topic. However, recently, researchers at Columbia university have created a new dataset called VAST along with a model called Topic-Grouped Attention (TGA), or better known as the TGANet, that claims to perform well on zero-shot and few-shot stance detection, which is a subset of stance detection that focuses on determining the stance of authors on new, never seen topics. Their VAST dataset focuses on this zero-shot and few-shot sub-problem by including a large variety of topics. This VAST dataset has many more topics than traditional stance detection datasets, which often focus on a particular subject to focus their topics around. In this thesis paper, we analyze how the TGA model performs on the SemEval2016 dataset and determine whether the TGA model improves on the current existing zero shot and few-shot stance detection models.
Description
Keywords
NLP, AI, Stance Detection