基于深度学习的多文档机器阅读理解综述

Summary

This Chinese-language survey covers deep-learning approaches to multi-document machine reading comprehension (MRC). In this task a system must read several documents together, filter and combine their content, and answer a question. Compared with single-document MRC, the task stresses question-guided logical links across documents. It also adds three demands: handling heterogeneous or noisy sources, stronger reasoning and interpretability, and scalability in time and memory as document sets grow.

The authors formalise both task variants. They compare eight representative datasets released between 2016 and 2020, grouping them as multi-source, linear-reasoning and fine-grained semi-structured. They also review the standard evaluation metrics. Methods are organised into a four-tier reading architecture. At the comprehension tier the survey examines four families: document selection, Transformer language models, graph convolutional networks and external-knowledge fusion.

The paper reports no new experiments. Its contribution is this taxonomy plus a closing list of open problems. For newcomers it offers a structured map of datasets and model families, including Chinese resources such as DuReader.

Key points

• Datasets surveyed (2016–2020): MS MARCO, TriviaQA, DuReader, WikiHop, MedHop, HotpotQA, AmazonQA and R4C. They span open and closed domains and generative, extractive and classification answer formats.
• Reported dataset scales:
  • MS MARCO: about 1.01M Q&A pairs and 8.8M documents.
  • DuReader: 201K questions, 417K answers and 946K documents, the largest Chinese MRC set when released.
  • AmazonQA: 923K questions and 14M documents.
  • HotpotQA: about 113K Q&A pairs with 10 documents each in the distractor setting.
  • MedHop: 2,508 questions.
• Dataset construction:
  • WikiHop is built from Wikidata triples via a bipartite entity–document graph and breadth-first search. Annotators judged about 74% of its samples to need cross-document reasoning.
  • MedHop applies the same procedure to DrugBank and Medline abstracts to predict drug–drug interactions.
• TriviaQA evidence coverage: the gathered documents allow the question to be answered in 79.7% of samples for Wikipedia and 75.4% for web documents.
• Metrics in use: BLEU, ROUGE-L, EM, F1, accuracy and joint F1. Most benchmarks evaluate on a blind test set.