Do Neural Language Models Overcome Reporting Bias?

Proceedings of the 28th International Conference on Computational Linguistics, pages 6863–6870 Barcelona, Spain (Online), December 8-13, 2020 6863 Do Neural Language Models Overcome Reporting Bias? Vered Shwartz and Yejin Choi Allen Institute for AI Paul G. Allen School of Computer Science & Engineering, University of Washington {vereds,yejinc}@allenai.org Abstract Mining commonsense knowledge from corpora suffers from reporting bias, over-representing the rare at the expense of the trivial (Gordon and Van Durme, 2013). We study to what extent pre-trained language models overcome this issue. We find that while their generalization capacity allows them to better estimate the plausibility of frequent but unspoken of actions, outcomes, and properties, they also tend to overestimate that of the very rare, amplifying the bias that already exists in their training corpus. 1 Introduction Apart from several notable efforts to collect commonsense knowledge from experts (Lenat, 1995) or through crowdsourcing (Speer and Havasi, 2012; Sap et al., 2019), most work has been on extracting such knowledge from large text corpora (Mitchell et al., 2018). While the latter approach is scalable and low cost, it also suffers from reporting bias: due to Grice’s conversational maxim of quantity (Grice et al., 1975), people rarely state the obvious, thus many trivial facts (“people breathe”) are rarely men- tioned in text, while uncommon events (“people murder”) are reported disproportionately (Gordon and Van Durme, 2013; Sorower et al., 2011). Traditionally, knowledge acquisition from text was extractive. In recent years, the generalization capacity of neural language models (LMs) and their ability to aggregate knowledge across contexts have facilitated estimating the plausibility of facts, even when they don’t appear in the corpus explicitly. Recent pre-trained LMs such as GPT-2 (Radford et al., 2019) and BERT (Devlin et al., 2019), trained on massive texts, dominate the NLP leaderboards, and are considered a source of commonsense knowledge (Petroni et al., 2019). Does this mean that pre-trained LMs overcome reporting bias? In this paper we revisit the experiments conducted by Gordon and Van Durme (2013) (henceforth G&V), applying them to various pre-trained LMs (based on the nature of the experiment, we test either masked LMs or standard left-to-right LMs). We find that LMs, compared to extractive methods:1 1. Provide a worse estimate of action frequency, mostly due to overestimating very rare actions. 2. Predict both expected outcomes as well as sensational and unlikely outcomes. 3. Are capable of learning associations between concepts and their properties indirectly, but tend to over-generalization, which leads to confusing semantically-similar but mutually exclusive values. 2 Actions and Events G&V demonstrate the discrepancy between corpus occurrences and actual action frequency by showing that if you believe the corpus, people murder more than they breathe. Breathing is an activity we take for granted and thus rarely talk about (Grice et al., 1975). That murder is frequent in the corpus is a reflection of the same issue: we talk more about uncommon or newsworthy events (van Dalen, 2012). We follow G&V’s qualitative analysis of actions and events performed by or which happen to people by …

Content truncated to 500 words