APER: AdaPtive Evidence-driven Reasoning Network for machine reading comprehension with unanswerable questions
This is the paper that published in 2021 Knowledge-Based System. Impact Factor, 8.038 (2020).
动机
在解决不可回答任务上,先前的方法有两个问题:
-
First, most of them utilize a simple classifier
or a verifiable module to determine whether a question is unanswerable, which lacks the explicit process of explanation. -
Second, these methods treat the answer extraction task and the unanswerable MRC task as two
independent tasks without considering the logical consistency of their results
贡献
-
An Evidence Refining Reasoner is designed to fuse questionrelevant information globally to refine the key evidence which is a basis for determining the non-answerability.
-
A novel logical consistency training objective is introduced to keep logical consistency of the results between the answer extraction task and the unanswerable MRC task.
-
Experimental results and ablation study on two datasets show that the APER has strong competitiveness and improves the Recall of unanswerable questions significantly.
做法
- Global Encoder
- 利用PLMs和task-specific global encoder 对上下文进行编码,获取 task-specific global semantic information
- Evidence Refining Reasoner (the interaction between the passage and question is sometimes performs multiple times during the rereading process of human comprehension)
- Evidence Refining Reasoner contains a stack of reasoning and fusion blocks, each of which contains a start and an end sub-block. See the figure below.
- Finally, Evidence Refining Reasoner refines the key evidence information for subsequent prediction.
-
Answer Consistency Detector
- outputs an unanswerable score based on the refined evidences (clues)
- compute a consistency training loss to keep logical consistency between the answer extraction task and the unanswerable MRC task.
-
Training process
-
Dataset processing
实验
-
Dureader and Squad 2.0
-
Precision §, Recall ® and F1 of unanswerable questions
-
ablation study
-
Case study
-
Different question types
- Steps of the reasoning