Diversify Your Datasets: Analyzing Generalization via Controlled Variance in Adversarial Datasets

Proceedings of the 23rd Conference on Computational Natural Language Learning, pages 196–205 Hong Kong, China, November 3-4, 2019. c⃝2019 Association for Computational Linguistics 196 Diversify Your Datasets: Analyzing Generalization via Controlled Variance in Adversarial Datasets Ohad Rozen1, Vered Shwartz2,3, Roee Aharoni1, and Ido Dagan1 1Computer Science Department, Bar-Ilan University, Ramat-Gan, Israel 2Allen Institute for Artificial Intelligence 3Paul G. Allen School of Computer Science & Engineering, University of Washington {ohadrozen,roee.aharoni}@gmail.com, vereds@allenai.org, dagan@cs.biu.ac.il Abstract Phenomenon-specific “adversarial” datasets have been recently designed to perform targeted stress-tests for particular inference types. Recent work (Liu et al., 2019a) pro- posed that such datasets can be utilized for training NLI and other types of models, of- ten allowing to learn the phenomenon in fo- cus and improve on the challenge dataset, in- dicating a “blind spot” in the original training data. Yet, although a model can improve in such a training process, it might still be vul- nerable to other challenge datasets targeting the same phenomenon but drawn from a differ- ent distribution, such as having a different syn- tactic complexity level. In this work, we ex- tend this method to drive conclusions about a model’s ability to learn and generalize a target phenomenon rather than to “learn” a dataset, by controlling additional aspects in the adver- sarial datasets. We demonstrate our approach on two inference phenomena – dative alterna- tion and numerical reasoning, elaborating, and in some cases contradicting, the results of Liu et al.. Our methodology enables building bet- ter challenge datasets for creating more robust models, and may yield better model under- standing and subsequent overarching improve- ments. 1 Introduction To successfully recognize textual entailment (RTE; Dagan et al., 2013), also known as natural language inference (NLI) (MacCartney and Man- ning, 2008; Bowman et al., 2015), a system needs to model a broad range of inference phenomena. Pre-neural systems often included explicit compo- nents, such as engineered features or syntax-based transformations (e.g. Stern and Dagan, 2012; Stern et al., 2012; Bar-Haim et al., 2015), to address particular inference types such as syntactic, lexi- cal, and logical inferences. Today’s neural mod- els do not explicitly model such inferences, but instead attempt to learn them implicitly from the training data. Despite their success on common NLI dataset, recent challenge datasets designed for probing different linguistic phenomena showed that neural models often fail on particular infer- ence types, like recognizing semantic relations and negation (Poliak et al., 2018; Naik et al., 2018; Glockner et al., 2018). Recently, Liu et al. (2019a) showed that when probing reveals a model’s failure on a specific lin- guistic phenomenon, it is often possible to amend this failure. They suggested to fine-tune the model on (a training section of) the challenge dataset it- self, in order to teach it to address the specific tar- get phenomenon, or in other words, to “inoculate” it against the adversarial data. Inoculation has two possible outcomes. The first - a success to ad- dress the phenomenon after fine-tuning - suggests the original training set did …

Content truncated to 500 words