Enhancing Dual-Encoders with Question and Answer Cross-Embeddings for Answer Retrieval

TL;DR

This paper introduces a novel framework that enhances Dual-Encoders for answer retrieval by incorporating question-answer cross-embeddings and a Geometry Alignment Mechanism, significantly improving performance without sacrificing inference efficiency.

Contribution

The work proposes a new framework combining cross-embeddings and a geometry alignment mechanism to improve Dual-Encoders for answer retrieval, addressing training and inference challenges.

Findings

Significant performance improvements on multiple datasets

Outperforms state-of-the-art methods in answer retrieval

Effective alignment of embedding geometries enhances retrieval accuracy

Abstract

Dual-Encoders is a promising mechanism for answer retrieval in question answering (QA) systems. Currently most conventional Dual-Encoders learn the semantic representations of questions and answers merely through matching score. Researchers proposed to introduce the QA interaction features in scoring function but at the cost of low efficiency in inference stage. To keep independent encoding of questions and answers during inference stage, variational auto-encoder is further introduced to reconstruct answers (questions) from question (answer) embeddings as an auxiliary task to enhance QA interaction in representation learning in training stage. However, the needs of text generation and answer retrieval are different, which leads to hardness in training. In this work, we propose a framework to enhance the Dual-Encoders model with question answer cross-embeddings and a novel Geometry Alignment Mechanism (GAM) to align the geometry of embeddings from Dual-Encoders with that from Cross-Encoders. Extensive experimental results show that our framework significantly improves Dual-Encoders model and outperforms the state-of-the-art method on multiple answer retrieval datasets.