Are Red Roses Red? Evaluating Consistency of Question-Answering Models

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6174–6184 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 6174 Are Red Roses Red? Evaluating Consistency of Question-Answering Models Marco Tulio Ribeiro Microsoft Research marcotcr@microsoft.com Carlos Guestrin University of Washington guestrin@cs.uw.edu Sameer Singh University of California, Irvine sameer@uci.edu Abstract Although current evaluation of question- answering systems treats predictions in isola- tion, we need to consider the relationship be- tween predictions to measure true understand- ing. A model should be penalized for answer- ing “no” to “Is the rose red?” if it answers “red” to “What color is the rose?”. We propose a method to automatically extract such impli- cations for instances from two QA datasets, VQA and SQuAD, which we then use to eval- uate the consistency of models. Human evalu- ation shows these generated implications are well formed and valid. Consistency evalua- tion provides crucial insights into gaps in ex- isting models, and retraining with implication- augmented data improves consistency on both synthetic and human-generated implications. 1 Introduction Question-answering (QA) systems have become popular benchmarks for AI systems, as they re- quire the ability to comprehend and employ com- plex reasoning about the question and the associ- ated context. In order to really excel in machine comprehension (Rajpurkar et al., 2016), for exam- ple, models need to understand the entities, coref- erences, and relations in the paragraph, and align them to the information need encoded in the ques- tion. Similarly, Visual Question Answering (An- tol et al., 2015) requires not only perception abili- ties (fine-grained recognition, object detection), but also “higher level reasoning” about how the ques- tion is related to the visual information, common- sense reasoning, knowledge based reasoning, and the understanding of location/color/size attributes. However, recent work has shown that popular benchmarks have crucial limitations in their ability to test reasoning and comprehension. For example, Weissenborn et al. (2017) show that models can do well in the SQuAD dataset by using heuristic (a) Input image from the VQA dataset. How many birds? A: 1 Is there 1 bird? A: no Are there 2 birds? A: yes Are there any birds? A: no (b) Model (Zhang et al., 2018) provides inconsistent answers. Kublai originally named his eldest son, Zhenjin, as the Crown Prince, but he died before Kublai in 1285. (c) Excerpt from an input paragraph, SQuAD dataset. Q: When did Zhenjin die? A: 1285 Q: Who died in 1285? A: Kublai (d) Model (Peters et al., 2018) provides inconsistent answers. Figure 1: Inconsistent QA Predictions: Models that are accurate for questions from these datasets (first row in (b) and (d)) are not able to correctly answer follow- up questions whose answers are implied by the original question/answer. We generate such questions automati- cally, and evaluate existing models on their consistency. lexical and type overlap between the context and the question. Biases have also been observed in the popular VQA dataset, e.g. answering questions starting with “Do you see a …

Content truncated to 500 words