PRO-ConvQA: Phrase Retrieval for Open-Domain Conversational QA with Contrastive Conversational Dependency Modeling

Forum topic · 小凯 · 2026-07-05

Summary

This paper (arXiv:2306.04293, June 2023) by Soyeong Jeong, Jinheon Baek, Sung Ju Hwang, and Jong C. Park addresses limitations of the conventional retriever-reader pipeline in Open-Domain Conversational Question Answering (ODConvQA). Standard pipelines retrieve passages and then extract answers, which propagates retriever errors to the reader, requires developing two separate models, and prevents parallel execution. The authors propose a phrase retrieval approach that directly predicts answers as spans over word sequences, collapsing the two subtasks into one, and study its capability for ODConvQA for the first time. Because conversational turns are interdependent, naive phrase retrieval fails on multi-turn contexts, so the authors introduce a novel contrastive learning strategy that maximizes representational similarity between consecutive turns while minimizing similarity to irrelevant conversational contexts, ensuring previous turns inform current-turn retrieval. Experiments on two ODConvQA datasets show the method substantially outperforms retriever-reader baselines. Code is available at https://github.com/starsuzi/PRO-ConvQA.

PRO-ConvQA: Phrase Retrieval for Open-Domain Conversational QA via Contrastive Learning

Paper: Phrase Retrieval for Open-Domain Conversational Question Answering with Conversational Dependency Modeling via Contrastive Learning Authors: Soyeong Jeong, Jinheon Baek, Sung Ju Hwang, Jong C. Park Published: 2023-06-07 Code: https://github.com/starsuzi/PRO-ConvQA

Motivation

Open-Domain Conversational Question Answering (ODConvQA) traditionally relies on a retriever-reader pipeline: passages are retrieved first, then a reader predicts answers from them. This design has three drawbacks:

Errors from the retriever propagate to the reader.
Two separate components must be developed and maintained.
The stages cannot run in parallel, making the system slower.

Approach

The paper proposes collapsing retrieval and answer extraction into a single phrase retrieval step: the model directly retrieves answer phrases from a sequence of words. This is the first study of phrase retrieval applied to ODConvQA.

A naive adoption, however, struggles because conversation turns are dependent — the meaning of the current query depends on prior turns. To handle this, the authors introduce a contrastive learning strategy that:

maximizes representational similarity between consecutive turns in a conversation;
minimizes similarity to irrelevant conversational contexts;

so that previous turns are properly reflected when retrieving phrases for the current context.

Method Outline (typical pipeline)

1. Input & representation — encode query, passages, and conversational context into dense representations. 2. Core module — a phrase retriever that directly selects answer spans, replacing the separate retriever + reader cascade. 3. Learning strategy — contrastive learning over consecutive vs. irrelevant turn representations. 4. Inference — single-stage phrase retrieval per conversational turn.

Results

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

Why It Matters

Unifying retrieval and extraction into one phrase retrieval stage reduces error propagation and system latency.
The conversational dependency modeling via contrastive learning is a general technique applicable to multi-turn retrieval settings.
It anticipates later trends in conversational search where retrieval granularity and multi-turn context handling are learned end-to-end.

Limitations & Open Questions

As with much research in this area, open issues include evaluation reliability beyond static metrics, latency and cost at scale, hallucination/safety concerns, and cross-lingual or multimodal extension.

Original Abstract

> Open-Domain Conversational Question Answering (ODConvQA) aims at answering questions through a multi-turn conversation based on a retriever-reader pipeline, which retrieves passages and then predicts answers with them. However, such a pipeline approach not only makes the reader vulnerable to the errors propagated from the retriever, but also demands additional effort to develop both the retriever and the reader, which further makes it slower since they are not runnable in parallel. In this work, we propose a method to directly predict answers with a phrase retrieval scheme for a sequence of words, reducing the conventional two distinct subtasks into a single one. Also, for the first time, we study its capability for ODConvQA tasks. However, simply adopting it is largely problematic, due to the dependencies between previous and current turns in a conversation. To address this problem, we further introduce a novel contrastive learning strategy, making sure to reflect previous turns when retrieving the phrase for the current context, by maximizing representational similarities of consecutive turns in a conversation while minimizing irrelevant conversational contexts. We validate our model on two ODConvQA datasets, whose experimental results show that it substantially outperforms the relevant baselines with the retriever-reader. Code is available at: https://github.com/starsuzi/PRO-ConvQA.