An Analysis of Logic Rule Dissemination in Sentiment Classifiers

Shashank Gupta, Mohamed Reda Bouadjenek, Antonio Robles-Kelly

Abstract: Disseminating and incorporating logic rules in deep neural networks has been extensively explored for sentiment classification. Methods that are proposed for that goal rely on a component that aims to capture and model logic rules, followed by a sequence model to process the input sequence. While these methods claim to effectively capture syntactic structures that affect sentiment, they only show improvement in terms of accuracy to support their claims with no further analysis. Focusing on the A-but-B rule, we use the PERCY metric (a recently developed Post-hoc Explanation-based score for logic Rule dissemination ConsistencY assessment) to analyze and study the ability of these methods to identify the A-but-B structure, and to make their classification decision based on the B conjunct. PERCY proceeds by estimating feature attribution scores using LIME, a model-agnostic framework that aims to explain the predictions of any classifier in an interpretable and faithful manner. Our experiments show that (a) accuracy is misleading in assessing these methods, (b) not all these methods are effectively capturing the A-but-B structure, (c) often, the underlying sequence model is what captures the syntactic structure, and (d) the best method classifies less than 25% of test examples based on the B conjunct.