Robust Domain Adaptation for Machine Reading Comprehension

Liang Jiang, Zhenyu Huang, Jia Liu, Zujie Wen, Xi Peng

Abstract

Most domain adaptation methods for machine reading comprehension (MRC) use a pre-trained question-answer (QA) construction model to generate pseudo QA pairs for MRC transfer. Such a process will inevitably introduce mismatched pairs (i.e., noisy correspondence) due to i) the unavailable QA pairs in target documents, and ii) the domain shift during applying the QA construction model to the target domain. Undoubtedly, the noisy correspondence will degenerate the performance of MRC, which however is neglected by existing works. To solve such an untouched problem, we propose to construct QA pairs by additionally using the dialogue related to the documents, as well as a new domain adaptation method for MRC. Specifically, we propose Robust Domain Adaptation for Machine Reading Comprehension (RMRC) method which consists of an answer extractor (AE), a question selector (QS), and an MRC model. Specifically, RMRC filters out the irrelevant answers by estimating the correlation to the document via the AE, and extracts the questions by fusing the candidate questions in multiple rounds of dialogue chats via the QS. With the extracted QA pairs, MRC is fine-tuned and provides the feedback to optimize the QS through a novel reinforced self-training method. Thanks to the optimization of the QS, our method will greatly alleviate the noisy correspondence problem caused by the domain shift. To the best of our knowledge, this could be the first study to reveal the influence of noisy correspondence in domain adaptation MRC models and show a feasible way to achieve robustness to mismatched pairs. Extensive experiments on three datasets demonstrate the effectiveness of our method.

Introduction: The Challenge of Noisy Correspondence

Machine Reading Comprehension (MRC) aims to enable models to answer questions based on a given context. While state-of-the-art models perform exceptionally well on general-domain datasets like Wikipedia, their performance often degrades when applied to specialized target domains (e.g., medical or customer service) where labeled data is scarce.

Standard Domain Adaptation (DA) techniques attempt to bridge this gap by generating "pseudo" question-answer (QA) pairs from target documents using a pre-trained generator. However, this process introduces a critical issue: Noisy Correspondence (NC). Unlike traditional "noisy labels"—where a data point is simply assigned the wrong category—noisy correspondence refers to a fundamental misalignment between a question and its supposed answer. This occurs because target documents may not naturally contain the information required to answer generated questions, or because the generator itself suffers from domain shift.

Comparison between Noisy Labels and Noisy Correspondence.

The RMRC Framework

To address the NC problem, we propose Robust Domain Adaptation for Machine Reading Comprehension (RMRC). The core innovation lies in leveraging associated dialogues (such as customer service logs) rather than relying solely on static documents. Dialogues provide a natural source of human intent and question-answer relationships that are often missing in raw documents.

The RMRC framework consists of three primary components:

  1. Answer Extractor (AE): This module identifies potential answer spans within a document by correlating them with responses found in the associated dialogue. It uses a BERT-based similarity scoring mechanism to filter out irrelevant or "junk" answers.
  2. Question Selector (QS): Instead of generating questions from scratch, the QS identifies relevant questions from the dialogue history. By fusing multiple rounds of chat, it constructs comprehensive pseudo-questions that are grounded in actual human interaction.
  3. MRC Model: A BERT-based reader is then fine-tuned on these pseudo-QA pairs to perform the final comprehension task.
The architecture of the RMRC framework, illustrating the Question-Answer Construction and MRC/QS Optimization phases.

Reinforced Self-Training

A key technical contribution of RMRC is the use of Reinforced Self-Training to optimize the Question Selector. Since the selection of questions involves discrete operations that are non-differentiable, we employ Reinforcement Learning (REINFORCE). The MRC model serves as a critic, providing a reward signal (based on the F1-score of its predictions) back to the QS. This feedback loop allows the QS to refine its selection strategy, effectively teaching the generator how to produce better training data for the reader.

Experimental Evaluation

We evaluated RMRC across three diverse datasets:

  • d-QuAC & d-CoQA: Modified versions of standard MRC datasets where questions were shuffled to simulate the discontinuous nature of real-world conversations.
  • Alipay Dataset: A real-world customer service dataset containing 1,526 dialogues and 3,813 human-annotated QA pairs, characterized by significant natural noise and domain-specific jargon.

Key Results

RMRC demonstrated superior performance over several strong baselines, including UQACT and CASe:

  • Performance Gains: On the d-QuAC dataset, RMRC achieved an F1-score of 24.46, significantly outperforming the best baseline (19.33). On the real-world Alipay dataset, RMRC improved the F1-score from 44.28 to 55.67.
  • Robustness to Noise: The results confirm that by explicitly modeling and filtering noisy correspondence, RMRC maintains high accuracy even in environments where traditional DA methods fail.
  • Data Efficiency: RMRC remains effective even with limited target-domain data, proving its utility in low-resource scenarios.

Conclusion

RMRC represents a significant step forward in domain adaptation for MRC. By identifying the "noisy correspondence" problem and introducing a dialogue-augmented, reinforced self-training framework, we provide a robust solution for transferring MRC models to specialized domains. This work highlights the value of using auxiliary conversational data to ground pseudo-QA generation in real-world human intent.

Summary updated

Abstract

Most domain adaptation methods for machine reading comprehension (MRC) use a pre-trained question-answer (QA) construction model to generate pseudo QA pairs for MRC transfer. Such a process will inevitably introduce mismatched pairs (i.e., noisy correspondence) due to i) the unavailable QA pairs in target documents, and ii) the domain shift during applying the QA construction model to the target domain. Undoubtedly, the noisy correspondence will degenerate the performance of MRC, which however is neglected by existing works. To solve such an untouched problem, we propose to construct QA pairs by additionally using the dialogue related to the documents, as well as a new domain adaptation method for MRC. Specifically, we propose Robust Domain Adaptation for Machine Reading Comprehension (RMRC) method which consists of an answer extractor (AE), a question selector (QS), and an MRC model. Specifically, RMRC filters out the irrelevant answers by estimating the correlation to the document via the AE, and extracts the questions by fusing the candidate questions in multiple rounds of dialogue chats via the QS. With the extracted QA pairs, MRC is fine-tuned and provides the feedback to optimize the QS through a novel reinforced self-training method. Thanks to the optimization of the QS, our method will greatly alleviate the noisy correspondence problem caused by the domain shift. To the best of our knowledge, this could be the first study to reveal the influence of noisy correspondence in domain adaptation MRC models and show a feasible way to achieve robustness to mismatched pairs. Extensive experiments on three datasets demonstrate the effectiveness of our method.
BibTeX
@inproceedings{jiang2023robust,
  author = {Liang Jiang and Zhenyu Huang and Jia Liu and Zujie Wen and Xi Peng},
  title = {{Robust Domain Adaptation for Machine Reading Comprehension}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2023},
  url = {https://openreview.net/forum?id=SWiXbJ8wb5},
}
Figure