Toward Deconfounding the Effect of Entity Demographics for Question Answering Accuracy

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 5457–5473 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 5457 Toward Deconfounding the Influence of Entity Demographics for Question Answering Accuracy Maharshi Gor∗ CS University of Maryland mgor@cs.umd.edu Kellie Webster Google Research, New York websterk@google.com Jordan Boyd-Graber∗ CS, UMIACS, iSchool, LCS University of Maryland jbg@umiacs.umd.edu Abstract The goal of question answering (QA) is to answer any question. However, major QA datasets have skewed distributions over gen- der, profession, and nationality. Despite that skew, model accuracy analysis reveals little ev- idence that accuracy is lower for people based on gender or nationality; instead, there is more variation on professions (question topic). But QA’s lack of representation could itself hide ev- idence of bias, necessitating QA datasets that better represent global diversity. 1 Introduction Question answering (QA) systems have impressive recent victories—beating trivia masters (Ferrucci et al., 2010) and superhuman reading (Najberg, 2018)—but these triumphs hold only if they gener- alize; QA systems should be able to answer ques- tions even if they do not look like training exam- ples. While other work (Section 4) focuses on demographic representation in NLP resources, our focus is how well QA models generalize across demographic subsets. After mapping mentions to a knowledge base (Section 2), we show existing QA datasets lack diversity in the gender and national origin of the people mentioned: English-language QA datasets mostly ask about US men from a few professions (Section 2.2). This is problematic because most En- glish speakers (and users of English QA systems) are not from the US or UK. Moreover, multilin- gual QA datasets are often translated from English datasets (Lewis et al., 2020; Artetxe et al., 2019). However, no work has verified that QA systems generalize to infrequent demographic groups. Section 3 investigates whether statistical tests reveal patterns on demographic subgroups. De- spite skewed distributions, accuracy is not corre- lated with gender or nationality, though it is with ∗ Work completed while at Google Research Entity Type NQ QB SQuAD TriviaQA Train Dev Train Dev Train Dev Train Dev Person 32.14 32.27 58.49 57.18 14.56 9.56 40.89 40.69 No entity 21.97 22.61 15.88 20.13 37.27 47.95 14.85 14.93 Location 28.42 27.82 34.50 35.11 33.52 29.34 44.32 44.10 Work of art 17.31 16.19 27.98 28.16 3.15 1.15 17.53 17.76 Other 2.66 2.70 9.83 9.88 5.08 3.10 14.38 14.61 Organization 17.73 18.05 20.37 17.78 21.34 19.14 22.15 21.14 Event 8.15 8.89 8.87 8.75 3.16 3.01 8.59 8.73 Product 0.85 1.06 0.88 0.63 1.39 0.33 3.99 3.99 Total Examples 106926 2631 112927 2216 130319 11873 87622 11313 Table 1: Coverage (% of examples) of entity-types in QA datasets. Since examples can mention more than one entity, columns can sum to >100%. Most datasets except SQuAD are rich in people. professional field. For instance, Natural Ques- tions (Kwiatkowski et al., 2019, NQ) systems do well with entertainers but poorly with scientists, which are handled well in TriviaQA. However, ab- sence of evidence is not evidence of absence …

Content truncated to 500 words