Inherent Disagreements in Human Textual Inferences

Inherent Disagreements in Human Textual Inferences Ellie Pavlick Brown University ellie pavlick@brown.edu Tom Kwiatkowski Google Research tomkwiat@google.com Abstract We analyze human’s disagreements about the validity of natural language inferences. We show that, very often, disagreements are not dismissible as annotation ‘‘noise’’, but rather persist as we collect more ratings and as we vary the amount of context provided to raters. We further show that the type of uncertainty captured by current state-of-the-art models for natural language inference is not reflective of the type of uncertainty present in human disagreements. We discuss implications of our results in relation to the recognizing textual entailment (RTE)/natural language inference (NLI) task. We argue for a refined evaluation objective that requires models to explicitly capture the full distribution of plausible human judgments. 1 Introduction Entailment is arguably one of the most fun- damental of language understanding tasks, with Montague himself calling entailment ‘‘the basic aim of semantics’’ (Montague, 1970). Compu- tational work on recognizing textual entailment (RTE) (also called natural language inference, or NLI) has a long history, ranging from early efforts to model logical phenomena (Cooper et al., 1996), to later statistical methods for modeling practical inferences needed for applications like informa- tion retrieval and extraction (Dagan et al., 2006), to current work on learning common sense hu- man inferences from hundreds of thousands of examples (Bowman et al., 2015; Williams et al., 2018). Broadly speaking, the goal of the NLI task is to train models to make the inferences that a human would make. Currently, ‘‘the inferences that a human would make’’ are determined by asking multiple human raters to label pairs of sentences, and then seeking some consensus among them. For example, having raters choose among discrete labels and taking a majority vote (Dagan et al., 2006; Bowman et al., 2015; Williams et al., 2018), or having raters use a continuous Likert scale and taking an average (Pavlick and Callison-Burch, 2016a; Zhang et al., 2017). That is, the prevailing assumption acrossannotation methods is that there is a single ‘‘true’’ inference about h given p that we should train models to predict, and that this label can be approximated by aggregating multi- ple (possibly noisy) human ratings as is typical in many other labelling tasks (Snow et al., 2008; Callison-Burch and Dredze, 2010). Often, however, we observe large disagree- ments among humans about whether or not h can be inferred from p (see Figure 1). The goal of this study is to establish whether such disagree- ments can safely be attributed to ‘‘noise’’ in the annotation process (resolvable via aggregation), or rather are a reproducible signal and thus should be treated as part of the NLI label assigned to the p/h pair. Specifically, our primary contributions are: • We perform a large-scale study of humans’ sentence-level inferences and measure the degree to which observed disagreements persist across samples of annotators. • We show that current state-of-the-art NLI systems do not capture this disagreement by default (by virtue of treating NLI as probabilistic) and …

Content truncated to 500 words