Integrate Candidate Answer Extraction with Re-Ranking for Chinese Machine Reading Comprehension.

Junjie Zeng,Qi Zhang,Xiaoya Sun,Xinmeng Li

doi:10.3390/e23030322

Abstract

Machine Reading Comprehension (MRC) research concerns how to endow machines with the ability to understand given passages and answer questions, which is a challenging problem in the field of natural language processing. To solve the Chinese MRC task efficiently, this paper proposes an Improved Extraction-based Reading Comprehension method with Answer Re-ranking (IERC-AR), consisting of a candidate answer extraction module and a re-ranking module. The candidate answer extraction module uses an improved pre-training language model, RoBERTa-WWM, to generate precise word representations, which can solve the problem of polysemy and is good for capturing Chinese word-level features. The re-ranking module re-evaluates candidate answers based on a self-attention mechanism, which can improve the accuracy of predicting answers. Traditional machine-reading methods generally integrate different modules into a pipeline system, which leads to re-encoding problems and inconsistent data distribution between the training and testing phases; therefore, this paper proposes an end-to-end model architecture for IERC-AR to reasonably integrate the candidate answer extraction and re-ranking modules. The experimental results on the Les MMRC dataset show that IERC-AR outperforms state-of-the-art MRC approaches.

Highlights

Machine Reading Comprehension (MRC), one of the most difficult problems in the field of natural language processing (NLP), concerns how to endow machines with the ability to comprehend a given context paragraph and answer corresponding questions
Regarding Recall-Oriented Understudy for Gisting Evaluation-L (ROUGE-L), the score of IERC-AR is 16.12% higher than Robustly optimized BERT Pretraining approach (RoBERTa)-whole word masking (WWM) based on Fine-Tuning and 1.33% higher than R-net. These experimental results show that IERC-AR generated the word embedding representation with more precise semantic information by using the improved pre-training language model RoBERTa-WWM, and that it effectively integrated the candidate answer extraction module with the candidate answer re-ranking module based on self-attention mechanism by proposing a novel end-to-end architecture, which can predict answers more precisely and increase the model’s robustness
This paper proposed an Improved Extraction-based Reading Comprehension methods with Answer Re-ranking (IERC-AR), which can improve performance on Chinese machine reading comprehension tasks

Summary

Introduction

Machine Reading Comprehension (MRC), one of the most difficult problems in the field of natural language processing (NLP), concerns how to endow machines with the ability to comprehend a given context paragraph and answer corresponding questions. Reading comprehension models can be divided into generation-based models and extraction-based models according to the way answers are generated. The former generates answers word by word, which can give neutral answers, but it lacks a global perspective, which makes it difficult to understand the answer in context; the latter extracts continuous fragments directly from the given passages, which can use the given context information to locate semantic units related to the questions [1]. Wang et al [2] compared the performance of the generation-based and extraction-based models in its proposed Match-LSTM method. This paper focuses on how to improve the performance of extraction-based reading comprehension methods

Methods

Results

Discussion

Conclusion