Phrase Retrieval for Open-Domain Conversational Question Answering with Conversational Dependency Modeling via Contrastive Learning

• Paper: arXiv:2306.04293
• Authors: Soyeong Jeong, Jinheon Baek, Sung Ju Hwang, Jong C. Park
• Published: 2023-06-07
• Code: https://github.com/starsuzi/PRO-ConvQA

Summary

Open-Domain Conversational Question Answering (ODConvQA) typically uses a retriever-reader pipeline: passages are retrieved first, then the reader predicts answers from them. The paper identifies three drawbacks of this pipeline:

1. Error propagation — reader performance is vulnerable to retriever mistakes. 2. Development overhead — two separate models must be built and maintained. 3. Latency — the two components cannot run in parallel.

Method

The authors propose directly predicting answers with a phrase retrieval scheme, reducing the two subtasks into a single one. This is the first study of phrase retrieval's capability for ODConvQA. Naive adoption is problematic because of dependencies between previous and current conversation turns. They therefore introduce a contrastive learning strategy that:

• maximizes representational similarities of consecutive turns in a conversation, and
• minimizes similarities to irrelevant conversational contexts,

so that phrase retrieval for the current turn properly reflects previous turns.

Experiments

The model is validated on two ODConvQA datasets. Results show it substantially outperforms relevant retriever-reader baselines. Exact figures should be checked against the original paper's tables.

Relevance

This work sits at the intersection of conversational search and LLM-era information retrieval, illustrating how the retrieval/ranking/generation boundary can be redrawn to reduce cascade errors and latency. Related entries include surveys on conversational search, ChatRetriever, CoSearchAgent, and agentic conversational search.