# Gated Group Self-Attention for Answer Selection

**Authors:** Dong Xu, Jianhui Ji, Haikuan Huang, Hongbo Deng, Wu-Jun Li  
**arXiv:** [1905.10720](https://arxiv.org/abs/1905.10720)  
**Date:** 2019-05-26  
**Summary page:** https://www.summarizepaper.com/en/paper/1905.10720v1/

## Key points

- Answer selection is crucial in various question answering (QA) applications
- Deep models have achieved state-of-the-art performance in answer selection
- Recurrent neural network (RNN) based models struggle to capture long-range dependencies between words in sentences
- Gated group self-attention (GGSA) is a new deep model proposed for answer selection
- GGSA is inspired by global self-attention, originally developed for machine translation and not explored in answer selection before
- GGSA overcomes the problem of global self-attention where local and global information cannot be well distinguished
- An interaction mechanism between questions and answers is proposed to enhance GGSA using a residual structure
- Experimental results on two popular QA datasets demonstrate that GGSA outperforms existing answer selection models to achieve state-of-the-art performance with lower computation cost than global self-attention.
- The paper's authors are Dong Xu, Jianhui Ji, Haikuan Huang, Hongbo Deng, and Wu-Jun Li.

## Summary

Answer selection is a crucial step in various question answering (QA) applications and deep models have achieved state-of-the-art performance in this area. Recurrent neural network (RNN) based models are the most popular among these deep models but struggle to capture long-range dependencies between words in sentences. To address this issue, the authors propose a new deep model called gated group self-attention (GGSA) for answer selection. GGSA is inspired by global self-attention which was originally developed for machine translation and has not been explored in answer selection before. GGSA overcomes the problem of global self-attention where local and global information cannot be well distinguished. Additionally, an interaction mechanism between questions and answers is proposed to enhance GGSA using a residual structure. Experimental results on two popular QA datasets demonstrate that GGSA outperforms existing answer selection models to achieve state-of-the-art performance with lower computation cost than global self-attention. The paper's authors are Dong Xu, Jianhui Ji, Haikuan Huang, Hongbo Deng, and Wu-Jun Li.