MFBE: Leveraging Multi-field Information of FAQs for Efficient Dense Retrieval

In the domain of question-answering in NLP, retrieval Frequently Asked Questions (FAQ) is an important sub-area which well researched and has been worked upon for many languages. Here, response to a user query, system typically returns relevant FAQs from knowledge-base. The efficacy such depends on its ability establish semantic match between query real-time. task becomes challenging due inherent lexical gap queries FAQs, lack sufficient context FAQ titles, scarcity labeled data high latency. this work, we propose bi-encoder-based query-FAQ matching model that leverages multiple combinations fields (like, question, answer, category) both during training inference. Our proposed Multi-Field Bi-Encoder (MFBE) benefits additional resulting performs even with minimal data. We empirically support claim through experiments proprietary as open-source public datasets unsupervised supervised settings. achieves around 27% 23% better top-1 accuracy internal open datasets, respectively over best performing baseline.