Interpretable Semantic Role Relation Table for Supporting Facts Recognition of Reading Comprehension

YanQing Bai, Zhichang Zhang, HaoYuan Chen, Xiaohui Qin, Yanglong Qiu

Keywords: interpretability, semantic role relation table, supporting facts

Abstract:

The current Machine Reading Comprehension (MRC) model has poor interpretability.
Interpretable semantic features can enhancethe interpretability of the model.
Semantic role labeling (SRL) captures predicate-argument relations, such as "who did what to whom," which
are critical to comprehension and interpretation.
To enhance the interpretability of the model, we propose the semantic role relation table,
which represents the semantic relation of the sentence itself and the semantic relations among sentences.
We use the name of entities to integrate into the semantic role relation table to establish the semantic relation between sentences.
This paper makes the first attempt to utilize contextual semantic's explicit relation to the recognition supporting fact of reading
comprehension.
We have established nine semantic relationtables between target sentence, question, and article.
Then we take each semantic relationship table's overall semantic role relevance and each
semantic role relevance as important judgment information.
Detailed experiments on HotpotQA, a challenging multi-hop MRC data set, our method achieves better performance.
With few training data sets, the model performance is still stable.

One-sentence Summary: We believe that the model uses interpretable features for reasoning, which contributes to enhanced interpretability. So we propose a handy and interpretable feature semantic role relationship table.