Query Reformulator And Contrastive Answer Ranker For Conversational Question Answering Over Knowledge Graph

Conversational KGQA (Knowledge Graph Question Answering) is a sequential process of question-answering over a knowledge graph (KG). Conversational KGQA involves follow-up questions that are presented in anaphoric or abbreviated form, known as ellipsis. Another challenge is the difficulty in selecting corrects answer for follow-up questions from a large number of candidate answers from KGs. This paper aims to tackle the challenge of alleviating the ellipsis phenomena and selecting correct answers from copious candidates. We present a module called CSR, which effectively merges information-retrieval-based method with pretrained language models. This module aims to identify the most relevant entity for elliptical questions based on instructional templates and subsequently rephrase the questions. Additionally, we propose a contrastive answer ranker to address the challenge of selecting the correct answer from copious candidates by leveraging the relationship between the questions and answers. Leveraging fewer trainable parameters, our method outperforms baselines on ConvQuestions and surpass the F1-Score metric of state-of-the-art on ConvRef dataset.