Temporality and modality in entailment graph induction

The ability to draw inferences is core to semantics and the field of Natural Language Processing. Answering a seemingly simple question like ‘Did Arsenal play Manchester yesterday’ from textual evidence that says ‘Arsenal won against Manchester yesterday’ requires modeling the inference that ‘winning’ entails ‘playing’. One way of modeling this type of lexical semantics is with Entailment Graphs, collections of meaning postulates that can be learned in an unsupervised way from large text corpora. In this work, we explore the role that temporality and linguistic modality can play in inducing Entailment Graphs. We identify inferences that were previously not supported by Entailment Graphs (such as that ‘visiting’ entails an ‘arrival’ before the visit) and inferences that were likely to be learned incorrectly (such as that ‘winning’ entails ‘losing’). Temporality is shown to be useful in alleviating these challenges, in the Entailment Graph representation as well as the learning algorithm. An exploration of linguistic modality in the training data shows, counterintuitively, that there is valuable signal in modalized predications. We develop three datasets for evaluating a system’s capability of modeling these inferences, which were previously underrepresented in entailment rule evaluations. Finally, in support of the work on modality, we release a relation extraction system that is capable of annotating linguistic modality, together with a comprehensive modality lexicon