How much should you ask? On the question structure in QA systems.

Proceedings of the 2018 EMNLP Workshop BlackboxNLP: Analyzing and Interpreting Neural Networks for NLP, pages 319–321 Brussels, Belgium, November 1, 2018. c⃝2018 Association for Computational Linguistics 319 How much should you ask? On the question structure in QA systems. Dominika Basaj∗1, Barbara Rychalska∗2, Przemysław Biecek3, and Anna Wróblewska4 d.basaj, b.rychalska, p.biecek, a.wroblewska@mini.pw.edu.pl 1,2,3,4Warsaw University of Technology, Warsaw, Poland 1Tooploox, Wroclaw, Poland Abstract Datasets that boosted state-of-the-art solutions for Question Answering (QA) systems prove that it is possible to ask questions in natural language manner. However, users are still used to query-like systems where they type in key- words to search for answer. In this study we validate which parts of questions are essential for obtaining valid answer. In order to conc- lude that, we take advantage of LIME - a fra- mework that explains prediction by local ap- proximation. We find that grammar and na- tural language is disregarded by QA. State- of-the-art model can answer properly even if ’asked’ only with a few words with high co- efficients calculated with LIME. According to our knowledge, it is the first time that QA mo- del is being explained by LIME. 1 Introduction Release of SQuAD (Rajpurkar et al., 2016) data- set boosted development of state-of-the-art solu- tions in Question Answering (QA) systems. Qu- estions asked in natural way give opportunity for human-computer interaction. However, in real life scenario, users are used to ’querying’ rather than ’asking’. This assumption inspired us to investi- gate whether QA systems trained on SQuAD da- taset could be used by people who prefer to write faster and more intuitive queries. Our experiments indicate that indeed, QA system returns true an- swer once we type in just selected keywords wi- thout keeping the sentence structure. We conclude that QA systems have a very limited understanding of natural language. They rather learn to distingu- ish specific words. This indifference to semantics- altering edits is called overstability (Jia and Liang, 2017). Research on this issue was also recently *Both authors contributed equally. conducted by Mudrakarta et al. (2018) who com- pute importance of words by application of Inte- grated Gradients. Inspired by LIME (Ribeiro et al., 2016) we per- turb questions and score newly created examples with context held constant. We prove that we can remove up to over 90% of words in question and still get the right answer. The contribution of our study is the following: 1. We use LIME in QA model for determining which parts of question are substantial for obta- ining right answer. We obtain valid results al- though QA systems are not natural candidates for explanation with LIME, since they do not solve a regular classification problem. 2. We show that QA models disregard gram- mar and syntax and thus can give the right an- swer once queried with most important keywords, which are the words with high coefficients retur- ned by LIME. Our findings can serve as a starting point for development of QA models that are im- mune to adversarial …

Content truncated to 500 words