FAQ Search using Transformers

Abstract

Many websites have bots as a guiding agent, for answering FAQ questions or directing users to human support. Many of them already have a curated FAQ page that can be used to bootstrap these bots. In this paper, we want to tackle a real-world problem of question answering for Bots. Given a user query, the system needs to pick the most relevant answer from a data source such as FAQ or Manuals. So, the ranking system needs to consider not just the passage but also the provided support questions or titles. This technique also provides the flexibility to add and delete support questions to continuously improve bot's quality, suggestions can be provided by system and the bot developer has control over their data instead of a black box system. We explore novel techniques to improve the results on a few public sets and on our own judged real user data. For the paper, We limit our experiments to transformers since it has proven to be significantly better in all question answering tasks. We show that significant gains can be observed using an extra segment embedding as well as pre-training new separators in transformers.

Tasks

Reproductions