Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Consensus Attention-based Neural Networks for Chinese Reading Comprehension

Yiming Cui, Ting Liu, Zhipeng Chen, Shijin Wang, Guoping Hu

TL;DR

This work tackles the lack of Chinese Cloze-style reading comprehension data by introducing two large Chinese datasets (People Daily and Children’s Fairy Tale) and a human-evaluated test set. It proposes the Consensus Attention Sum Reader (CAS Reader), a neural architecture that aggregates attention across all query time steps via a merging function to produce word-level predictions directly from the document. Empirical results show CAS Reader achieves strong performance on public benchmarks and significantly improves results on the Chinese datasets, with insights about merging strategies and domain transfer. The paper also establishes a baseline for Chinese RC and highlights directions for handling real-world questions and domain shifts in future work.

Abstract

Reading comprehension has embraced a booming in recent NLP research. Several institutes have released the Cloze-style reading comprehension data, and these have greatly accelerated the research of machine comprehension. In this work, we firstly present Chinese reading comprehension datasets, which consist of People Daily news dataset and Children's Fairy Tale (CFT) dataset. Also, we propose a consensus attention-based neural network architecture to tackle the Cloze-style reading comprehension problem, which aims to induce a consensus attention over every words in the query. Experimental results show that the proposed neural network significantly outperforms the state-of-the-art baselines in several public datasets. Furthermore, we setup a baseline for Chinese reading comprehension task, and hopefully this would speed up the process for future research.

Consensus Attention-based Neural Networks for Chinese Reading Comprehension

TL;DR

This work tackles the lack of Chinese Cloze-style reading comprehension data by introducing two large Chinese datasets (People Daily and Children’s Fairy Tale) and a human-evaluated test set. It proposes the Consensus Attention Sum Reader (CAS Reader), a neural architecture that aggregates attention across all query time steps via a merging function to produce word-level predictions directly from the document. Empirical results show CAS Reader achieves strong performance on public benchmarks and significantly improves results on the Chinese datasets, with insights about merging strategies and domain transfer. The paper also establishes a baseline for Chinese RC and highlights directions for handling real-world questions and domain shifts in future work.

Abstract

Reading comprehension has embraced a booming in recent NLP research. Several institutes have released the Cloze-style reading comprehension data, and these have greatly accelerated the research of machine comprehension. In this work, we firstly present Chinese reading comprehension datasets, which consist of People Daily news dataset and Children's Fairy Tale (CFT) dataset. Also, we propose a consensus attention-based neural network architecture to tackle the Cloze-style reading comprehension problem, which aims to induce a consensus attention over every words in the query. Experimental results show that the proposed neural network significantly outperforms the state-of-the-art baselines in several public datasets. Furthermore, we setup a baseline for Chinese reading comprehension task, and hopefully this would speed up the process for future research.

Paper Structure

This paper contains 11 sections, 9 equations, 2 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Chinese Reading Comprehension Datasets
  3. Existing Cloze-style Datasets
  4. People Daily and Children's Fairy Tale Datasets
  5. Consensus Attention Sum Reader
  6. Experiments
  7. Experimental Setups
  8. Results on Public Datasets
  9. Results on Chinese Reading Comprehension Datasets
  10. Related Work
  11. Conclusion

Figures (2)

  • Figure 1: Example training sample in People Daily datasets (the English translation is given in the right box). The " XXXXX" represents the missing word. In this example, the document consists of 10 sentences, and the 9th sentence is chosen as the query.
  • Figure 2: Architecture of the proposed Consensus Attention Sum Reader (CAS Reader).