Mnemonic Reader: Machine Comprehension with Iterative Aligning and Multi-hop Answer Pointing

Minghao Hu, Yuxing Peng, Xipeng Qiu

Abstract

In this paper, we introduce the Reinforced Mnemonic Reader for machine reading comprehension tasks, which enhances previous attentive readers in two aspects. First, a reattention mechanism is proposed to refine current attentions by directly accessing past attentions that are temporally memorized in a multi-round alignment architecture, so as to avoid the problems of attention redundancy and attention deficiency. Second, a new optimization approach, called dynamic-critical reinforcement learning, is introduced to extend the standard supervised method. It always encourages the prediction of a more acceptable answer so as to address the convergence suppression problem occurred in traditional reinforcement learning algorithms. Extensive experiments on the Stanford Question Answering Dataset (SQuAD) show that our model achieves state-of-the-art results. Meanwhile, our model outperforms previous systems by over 6% in terms of both Exact Match and F1 metrics on two adversarial SQuAD datasets.

Introduction

Machine Reading Comprehension (MRC) requires a model to read a passage (context) and answer a specific question (query). While previous models were effective at identifying local patterns, they faced three major hurdles: shallow interaction between query and context, the long-term dependency problem where RNNs struggle with long paragraphs, and the limitations of "one-shot" prediction for complex questions with multiple distractors. The Mnemonic Reader is designed to overcome these by mimicking human problem-solving: through rereading and iterative refinement.

The Mnemonic Reader Architecture

The Mnemonic Reader introduces a sophisticated neural architecture composed of three novel components designed to capture deep semantic links and handle long-distance dependencies.

The overall architecture of the Mnemonic Reader, featuring the Query-sensitive Encoder, Iterative Aligner, and Multi-hop Answer Pointer.

Iterative Aligner

This module performs two types of "alignment" over multiple rounds (hops). Interactive Alignment matches the query to the context, while Self-Alignment matches the context to itself. This self-alignment is crucial as it allows the model to link related ideas across distant parts of the text, effectively bypassing the "forgetting" problem of traditional RNNs.

Semantic Fusion Unit (SFU)

Instead of simple concatenation, the SFU uses a specialized gating mechanism to decide how much new "attended" information to merge into existing word representations. This ensures that new context is integrated without washing out the original semantic meaning.

Multi-hop Answer Pointer

Rather than predicting the answer span once, the model uses a memory vector that updates over several steps. It makes an initial guess, "remembers" that guess, and then uses that memory to refine the boundaries of the answer span in subsequent hops.

Query-Sensitive Encoding

The model explicitly categorizes the query (e.g., "who," "where," "why") and fuses this category into the initial representation. This helps the pointer understand the specific kind of answer required by the question type.

Experimental Evaluation

The authors evaluated the model on two massive benchmarks:

  • SQuAD (Stanford Question Answering Dataset): Over 100,000 Q&A pairs based on Wikipedia articles.
  • TriviaQA: A challenging dataset with contexts averaging nearly 3,000 words, requiring robust reasoning.

At the time of submission, the Mnemonic Reader ensemble reached the top of the SQuAD leaderboard with an F1 score of 81.69% and an Exact Match (EM) of 73.67%. On TriviaQA, the model significantly outperformed the standard BiDAF baseline, showing a 7% improvement in EM on the Wikipedia domain.

Analysis and Insights

Ablation studies revealed that Interactive Aligning was the most critical component; its removal caused the F1 score to drop by approximately 6%. Furthermore, the iterative nature of the model proved that "rereading" is as effective for AI as it is for humans.

Comparison of F1 scores between BiDAF and Mnemonic Reader across different question categories.

As shown in the performance breakdown by question type, while the model excels at "Who" and "When" questions, "Why" questions remain the most difficult. These questions require high-level causal reasoning rather than just pattern matching, representing a persistent challenge in the field.

Conclusion and Future Work

The Mnemonic Reader demonstrates that iterative refinement and category awareness are powerful tools for MRC. However, increasing the number of hops beyond two was found to decrease performance due to overfitting. Future research may involve incorporating external common knowledge bases to help the model understand concepts not explicitly stated in the text, particularly to bridge the reasoning gap in complex "Why" questions.

Summary updated

Abstract

In this paper, we introduce the Reinforced Mnemonic Reader for machine reading comprehension tasks, which enhances previous attentive readers in two aspects. First, a reattention mechanism is proposed to refine current attentions by directly accessing to past attentions that are temporally memorized in a multi-round alignment architecture, so as to avoid the problems of attention redundancy and attention deficiency. Second, a new optimization approach, called dynamic-critical reinforcement learning, is introduced to extend the standard supervised method. It always encourages to predict a more acceptable answer so as to address the convergence suppression problem occurred in traditional reinforcement learning algorithms. Extensive experiments on the Stanford Question Answering Dataset (SQuAD) show that our model achieves state-of-the-art results. Meanwhile, our model outperforms previous systems by over 6% in terms of both Exact Match and F1 metrics on two adversarial SQuAD datasets.
BibTeX
@inproceedings{hu2018mnemonic,
  author = {Minghao Hu and Yuxing Peng and Xipeng Qiu},
  title = {{Mnemonic Reader: Machine Comprehension with Iterative Aligning and Multi-hop Answer Pointing}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  doi = {10.24963/ijcai.2018/570},
}
Figure