Exploring Dual Encoder Architectures for Question Answering

less than 1 minute read

Introduction

  • What is the name of the Dual Encoder Question Answering paper?

    Exploring Dual Encoder Architectures for Question Answering

    google

  • What are the main contributions of the Dual Encoder Question Answering paper?
    • show superior performance of SDEs over ADEs
    • improvement of ADEs by sharing projection layer
  • In an Assymetric Dual Encoder parameters are not shared between the query and passage encoder
    • example: DPR
    • generally worse than Siamese Dual Encoder

Method

  • For Dual Encoders there are several options for parameter sharing for example can share input embedding matrices

    Untitled

Results

  • Assymetric Dual Encoders fail to separate query and passage in the embedding space

    Untitled

  • Best results for Assymetric Dual Encoders are achieved with shared output projection weights

    Untitled

Conclusions

This is a nice paper giving some empirical insight into different retrieval techniques. It’s a little disappointing that assymetric dual encoders can’t perform better given that there is often a much different data distribution captured by the query and passages in an information retrieval system.

Reference

@article{dong2022exploring,
  title={Exploring Dual Encoder Architectures for Question Answering},
  author={Dong, Zhe and Ni, Jianmo and Bikel, Dan and Alfonseca, Enrique and Wang, Yuan and Qu, Chen and Zitouni, Imed},
  journal={arXiv preprint arXiv:2204.07120},
  year={2022}
}

Updated: