GHOST at SemEval-2021 Task 5: Is explanation all you need?

Proceedings of the 15th International Workshop on Semantic Evaluation (SemEval-2021), pages 852–859 Bangkok, Thailand (online), August 5–6, 2021. ©2021 Association for Computational Linguistics 852 GHOST at SemEval-2021 Task 5: Is explanation all you need? Kamil Pluci´nski and Hanna Klimczak Institute of Computer Science Poznan University of Technology, 60-965 Pozna´n, Poland {kamil.plucinski97,hanna.klimczak}@gmail.com Abstract This paper discusses different approaches to the Toxic Spans Detection task. The prob- lem posed by the task was to determine which words contribute mostly to recognising a doc- ument as toxic. As opposed to binary classi- fication of entire texts, word-level assessment could be of great use during comment modera- tion, also allowing for a more in-depth com- prehension of the model’s predictions. As the main goal was to ensure transparency and understanding, this paper focuses on the cur- rent state-of-the-art approaches based on the explainable AI concepts and compares them to a supervised learning solution with word- level labels. The work consists of two xAI approaches that automatically provide the ex- planation for models trained for binary classi- fication of toxic documents: an LSTM model with attention as a model-specific approach and the Shapley values for interpreting BERT predictions as a model-agnostic method. The competing approach considers this problem as supervised token classification, where models like BERT and its modifications were tested. The paper aims to explore, compare and assess the quality of predictions for different meth- ods on the task. The advantages of each ap- proach and further research direction are also discussed. 1 Introduction The popularity of social media platforms has been continuously increasing over time. As reported by Social (2020), 3.8 billion people have been active users of social media in January 2020. While user inter-connectivity carries a lot of positive effects, there is still a significant threat of cyberbullying and harassment caused by the illusion of anonymity online. Statistics released by Facebook (Richter, 2020) identified that in the first quarter of 2020, there were 2.3 million bullying/harassment posts and 9.6 million hate speech posts detected as a violation of Community Standards. The importance of keeping the online commu- nity safe for users has caused many researchers to focus their work on detecting toxic contents in order to assist moderators in their difficult and men- tally exhausting work. A lot of progress has been done in the task of toxic comment classification, but unfortunately, complex models still lack clear ex- planation and cannot gain much moderators’ trust. A step towards increasing the transparency and therefore trust in automatic comment classifiers would be extending current approaches to operate on word-level rather than document-level. The abil- ity to extract and highlight key text fragments that cause the toxic character of a comment could be of great use in post moderation. The lack of extensive datasets for word-level toxicity detection poses an obstacle to traditional, supervised learning classification, as current state- of-the-art models are very complex and normally require large-scale data. Therefore, with the subject of this task being defined in terms of understanding and transparency …

Content truncated to 500 words