A Study on Various Approaches Towards Non-Factoid Question Answering Systems

Question Answering (QA) is a branch of Natural Language Processing (NLP) which develop systems that automatically provide answer to questions queried by humans in any natural language. Depending upon the type of question, QA systems can be both factoid and non-factoid type. There are numerous methods for retrieving answers for the input queries. Due to its complex behavior, non-factoid QA still remains a challenge to many researchers. This survey mainly focuses on works done so far in the area of non-factoid QA systems. Beginning with traditional deep learning approaches and moving towards the usage of Knowledge-Graphs (KG), this paper focuses on studying how to improve the efficiency of such QA systems. Transformer models like Bidirectional Encoder Representations from Transformers (BERT) and its variants reflect a significant improvement in QA implementation. One of the main challenges faced by those models was to answer the questions from knowledge-dependable domains. The study shows that knowledge-graphs help to address these issues to some extent. Also, it provides an analytical discussion on various approaches in non-factoid QA system implementation and how KG outperform previous methods.