What does BERT Learn from Multiple-Choice Reading Comprehension Datasets?
Chenglei Si1, Shuohang Wang2, Min-Yen Kan3, Jing Jiang4 1River Valley High School, 2Microsoft Dynamics 365 AI Research, 3School of Computing, National University of Singapore 4Singapore Management University
Abstract
Multiple-Choice Reading Comprehension (MCRC) requires the model to read the passage and question, and select the correct answer among the given options. Recent state-of-the-art models have achieved impressive performance on multiple MCRC datasets. However, such performance may not reflect the model’s true ability of language understanding and reasoning. In this work, we adopt two approaches to investigate what BERT learns from MCRC datasets: 1) an un-readable data attack, in which we add keywords to confuse BERT, leading to a significant performance drop; and 2) an un-answerable data training, in which we train BERT on partial or shuffled input. Under un-answerable data training, BERT achieves unexpectedly high performance. Based on our experiments on the 5 key MCRC datasets — RACE, MCTest, MCScript, MCScript2.0, DREAM — we observe that 1) fine-tuned BERT mainly learns how keywords lead to correct prediction, instead of learning semantic understanding and reasoning; and 2) BERT does not need correct syntactic information to solve the task; 3) there exists artifacts in these datasets such that they can be solved even without the full context.
中文速览
大规模预训练模型BERT在多项选择阅读理解（MCRC）任务上表现亮眼，但这究竟是真正"读懂"了文章，还是在走捷径？研究者设计了两类探测实验：一是"不可读数据攻击"，在测试时往文章或选项里塞入打乱词序、语义破碎的干扰内容，结果BERT的准确率大幅下滑，说明它严重依赖关键词匹配而非真正理解文意；二是"不可作答数据训练"，在训练时把段落或问题整段删掉、或把词序全部打乱，人类从这样的数据里根本学不到任何解题规律，但BERT在五个主流MCRC数据集（RACE、MCTest、MCScript、MCScript2.0、DREAM）上仍远超随机猜测的基线，表明它无需正确的句法结构也能"答题"。这两方面证据共同揭示：BERT在阅读理解上主要学到的是统计关键词规律，而非语义推理能力，同时现有基准数据集本身存在大量可被利用的统计捷径，并不能真实衡量模型的语言理解水平，这对如何设计更可靠的评测数据集和模型训练范式具有重要警示意义。
原文 arXiv:1910.12391;中英对照 + 大白话阅读 https://aha.fim.ai/paper/1910.12391v1