Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

DCN+: Mixed Objective and Deep Residual Coattention for Question Answering

Caiming Xiong, Victor Zhong, Richard Socher

TL;DR

DCN+ tackles the misalignment between training losses and QA evaluation by integrating cross-entropy with a word-overlap-based reinforcement learning objective. It also introduces a deep residual coattention encoder to capture long-range dependencies in the question-document pair. The combined objective and architecture yield state-of-the-art SQuAD results, especially for long questions, and ablation confirms the contributions of both components. The approach advances question answering by aligning optimization with evaluation and improving representational depth.

Abstract

Traditional models for question answering optimize using cross entropy loss, which encourages exact answers at the cost of penalizing nearby or overlapping answers that are sometimes equally accurate. We propose a mixed objective that combines cross entropy loss with self-critical policy learning. The objective uses rewards derived from word overlap to solve the misalignment between evaluation metric and optimization objective. In addition to the mixed objective, we improve dynamic coattention networks (DCN) with a deep residual coattention encoder that is inspired by recent work in deep self-attention and residual networks. Our proposals improve model performance across question types and input lengths, especially for long questions that requires the ability to capture long-term dependencies. On the Stanford Question Answering Dataset, our model achieves state-of-the-art results with 75.1% exact match accuracy and 83.1% F1, while the ensemble obtains 78.9% exact match accuracy and 86.0% F1.

DCN+: Mixed Objective and Deep Residual Coattention for Question Answering

TL;DR

DCN+ tackles the misalignment between training losses and QA evaluation by integrating cross-entropy with a word-overlap-based reinforcement learning objective. It also introduces a deep residual coattention encoder to capture long-range dependencies in the question-document pair. The combined objective and architecture yield state-of-the-art SQuAD results, especially for long questions, and ablation confirms the contributions of both components. The approach advances question answering by aligning optimization with evaluation and improving representational depth.

Abstract

Traditional models for question answering optimize using cross entropy loss, which encourages exact answers at the cost of penalizing nearby or overlapping answers that are sometimes equally accurate. We propose a mixed objective that combines cross entropy loss with self-critical policy learning. The objective uses rewards derived from word overlap to solve the misalignment between evaluation metric and optimization objective. In addition to the mixed objective, we improve dynamic coattention networks (DCN) with a deep residual coattention encoder that is inspired by recent work in deep self-attention and residual networks. Our proposals improve model performance across question types and input lengths, especially for long questions that requires the ability to capture long-term dependencies. On the Stanford Question Answering Dataset, our model achieves state-of-the-art results with 75.1% exact match accuracy and 83.1% F1, while the ensemble obtains 78.9% exact match accuracy and 86.0% F1.

Paper Structure

This paper contains 15 sections, 10 equations, 5 figures, 2 tables.

Table of Contents

  1. Introduction
  2. DCN+
  3. Deep residual coattention encoder
  4. Mixed objective using self-critical policy learning
  5. Experiments
  6. Results
  7. Comparison to baseline DCN with CoVe.
  8. Ablation study.
  9. Mixed objective convergence.
  10. Sample predictions.
  11. Related work
  12. Neural models for question answering.
  13. Neural attention models.
  14. Reinforcement learning in NLP.
  15. Conclusion

Figures (5)

  • Figure 1: Deep residual coattention encoder.
  • Figure 2: Computation of the mixed objective.
  • Figure 3: Performance comparison between DCN+ and the baseline DCN with CoVe on the SQuAD development set.
  • Figure 4: Training curve of DCN+ with and without reinforcement learning. In the latter case, only the cross entropy objective is used. The mixed objective initially performs worse as it begins policy learning from scratch, but quickly outperforms the cross entropy model.
  • Figure 5: Predictions by DCN+ ( red) and DCN with CoVe ( blue) on the SQuAD development set.