Transformer-Based Extractive Social Media Question Answering on TweetQA
Sabur Butt, Noman Ashraf, Hammad Fahim, Grigori Sidorov, Alexander Gelbukh
Abstract
The paper tackles the problem of question answering on social media data through an extractive approach. The task of question answering consists in obtaining an answer from the context given the context and a question. Our approach uses transformer models, which were fine-tuned on SQuAD. Usually, SQuAD is used for extractive question answering for comparing the results with human judgments in social mediaTweetQA dataset. Our experiments on multiple transformer models indicate the importance of application of pre-processing in the question answering on social media data and elucidates that extractive question answering fine-tuning even on other type of data can significantly improve the results reducing the gap with human evaluation. We use ROUGE, METEOR, and BLEU metrics.