Performance Impact Caused by Hidden Bias of Training Data for Recognizing Textual Entailment
Masatoshi Tsuchiya
TL;DR
The paper investigates training-data quality for Recognizing Textual Entailment (RTE) by proposing a hypothesis-testing framework to detect hidden biases. It introduces a TE-label prediction test without premises using Naive Bayes and contrasts SNLI with SICK, revealing a hypothesis-only bias in SNLI. The results show that a large portion of neural RTE performance on SNLI can be attributed to this bias, effectively turning NN systems into TE-label predictors for biased data. These findings highlight the need to account for dataset biases when evaluating NN-based RTE models and dataset construction.
Abstract
The quality of training data is one of the crucial problems when a learning-centered approach is employed. This paper proposes a new method to investigate the quality of a large corpus designed for the recognizing textual entailment (RTE) task. The proposed method, which is inspired by a statistical hypothesis test, consists of two phases: the first phase is to introduce the predictability of textual entailment labels as a null hypothesis which is extremely unacceptable if a target corpus has no hidden bias, and the second phase is to test the null hypothesis using a Naive Bayes model. The experimental result of the Stanford Natural Language Inference (SNLI) corpus does not reject the null hypothesis. Therefore, it indicates that the SNLI corpus has a hidden bias which allows prediction of textual entailment labels from hypothesis sentences even if no context information is given by a premise sentence. This paper also presents the performance impact of NN models for RTE caused by this hidden bias.