Behavior Analysis of NLI Models: Uncovering the Influence of Three Factors on Robustness

Proceedings of NAACL-HLT 2018, pages 1975–1985 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Behavior Analysis of NLI Models: Uncovering the Influence of Three Factors on Robustness V. Ivan Sanchez Carmona and Jeff Mitchell and Sebastian Riedel University College London Department of Computer Science {i.sanchezcarmona, j.mitchell, s.riedel}@cs.ucl.ac.uk Abstract Natural Language Inference is a challenging task that has received substantial attention, and state-of-the-art models now achieve impres- sive test set performance in the form of ac- curacy scores. Here, we go beyond this sin- gle evaluation metric to examine robustness to semantically-valid alterations to the input data. We identify three factors - insensitiv- ity, polarity and unseen pairs - and compare their impact on three SNLI models under a va- riety of conditions. Our results demonstrate a number of strengths and weaknesses in the models’ ability to generalise to new in-domain instances. In particular, while strong perfor- mance is possible on unseen hypernyms, un- seen antonyms are more challenging for all the models. More generally, the models suffer from an insensitivity to certain small but se- mantically significant alterations, and are also often influenced by simple statistical correla- tions between words and training labels. Over- all, we show that evaluations of NLI models can benefit from studying the influence of fac- tors intrinsic to the models or found in the dataset used. 1 Introduction The task of Natural Language Inference (NLI)1 has received a lot of attention and has elicited models which have achieved impressive results on the Stanford NLI (SNLI) dataset (Bowman et al., 2015). Such results are impressive due to the linguistic knowledge required to solve the task (LoBue and Yates, 2011; Maccartney, 2009). However, the ever-growing complexity of these models inhibits a full understanding of the phe- nomena that they capture. 1Also known as Recognizing Textual Entailment. As a consequence, evaluating these models purely on test set performance may not yield enough insight into the complete repertoire of abil- ities learned and any possible abnormal behav- iors (Kummerfeld et al., 2012; Sammons et al., 2010). A similar case can be observed in models from other domains; take as an example an image classifier that predicts based on the image’s back- ground rather than on the target object (Zhao et al., 2017; Ribeiro et al., 2016), or a classifier used in social contexts that predicts a label based on racial attributes (Crawford and Calo, 2016). In both ex- amples, the models exploit a bias (an undesired pattern hidden in the dataset) to enhance accuracy. In such cases, the models may appear to be robust to new and even challenging test instances; how- ever, this behavior may be due to spurious factors, such as biases. Assessing to what extent the mod- els are robust to these contingencies just by look- ing at test accuracy is, therefore, difficult. In this work we aim to study how certain fac- tors affect the robustness of three pre-trained NLI models (a conditional encoder, the DAM model (Parikh et al., …

Content truncated to 500 words