[Paper Review] Performance Impact Caused by Hidden Bias of Training Data for Recognizing Textual Entailment

Masatoshi Tsuchiya|arXiv (Cornell University)|Apr 22, 2018

Topic ModelingComputer Science21 references123 citations

TL;DR

The paper proposes a two-phase method to detect hidden bias in RTE corpora using a Naive Bayes TE-label predictor and a baseline, finds hidden bias in SNLI but not in SICK, and shows this bias can distort neural NLP models’ performance for RTE.

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.

Motivation & Objective

• Assess the quality of large RTE corpora.
• Introduce a null hypothesis about TE-label predictability without context.
• Develop a Naive Bayes TE-label prediction model.
• Compare SNLI and SICK corpora to reveal hidden biases.
• Discuss how hidden bias impacts neural network models for RTE.

Proposed method

• Define TE-label predictability without premises as the null hypothesis.
• Use a multinomial Naive Bayes model with unigram features on hypothesis sentences to predict TE labels.
• Use a baseline model that assigns the most frequent TE label from the corpus when premise and hypothesis contexts are absent.
• Test null hypothesis by comparing TE-label prediction model to baseline using a sign test.
• Apply the method to SNLI and SICK corpora to assess hidden bias.
• Discuss implications for NN-based RTE models and how biases can masquerade as learning signals.

Experimental results

Research questions

• RQ1 Do RTE corpora contain hidden bias that allows predicting TE labels without premises?
• RQ2 Can a NB TE-label predictor outperform a corpus-based majority baseline on hypothesis-only data?
• RQ3 Is hidden bias present in SNLI, or is it absent in SICK?
• RQ4 How does detected bias affect the evaluation and learning behavior of NN models for RTE?

Key findings

• The TE label prediction model achieves 63.3% accuracy on SNLI hypothesis sentences without premises, versus 34.3% for the baseline.
• For SICK, the TE label predictor and baseline perform similarly (56.7%).
• The difference between models on SNLI is statistically significant (p = 5.7e−202).
• The hidden bias in SNLI enables TE-label prediction without context, while SICK shows no such bias (null hypothesis rejected for SICK, not SNLI).
• NN models for RTE exhibit large performance drops on empirical hard test sets, indicating reliance on the bias rather than genuine context understanding.
• Replacing premise words with unknown tokens reduces context, yet NN models still perform above chance on empirical easy test sets, suggesting TE-label prediction rather than true RTE behavior.