Text-based NP Enrichment

Text-based NP Enrichment Yanai Elazar∗ Victoria Basmov∗ Yoav Goldberg Reut Tsarfaty Computer Science Department, Bar Ilan University Allen Institute for Artificial Intelligence, Israel {yanaiela,vikasaeta,yoav.goldberg,reut.tsarfaty}@gmail.com Abstract Understanding the relations between entities denoted by NPs in a text is a critical part of human-like natural language understanding. However, only a fraction of such relations is covered by standard NLP tasks and bench- marks nowadays. In this work, we propose a novel task termed text-based NP enrichment (TNE), in which we aim to enrich each NP in a text with all the preposition-mediated relations—either explicit or implicit—that hold between it and other NPs in the text. The relations are represented as triplets, each denoted by two NPs related via a preposition. Humans recover such relations seamlessly, while current state-of-the-art models struggle with them due to the implicit nature of the problem. We build the first large-scale dataset for the problem, provide the formal framing and scope of annotation, analyze the data, and report the results of fine-tuned language mod- els on the task, demonstrating the challenge it poses to current technology. A webpage with a data-exploration UI, a demo, and links to the code, models, and leaderboard, to foster further research into this challenging problem can be found at: yanaiela.github.io/TNE/. 1 Introduction A critical part of understanding a text is detecting the entities in the text, denoted by NPs, and deter- mining the different semantic relations that hold between them. Some semantic relations between NPs are explicitly mediated via verbs, as in (1): Much work in NLP addresses the recovery of such verb-mediated relations (SRL) (Gildea and Jurafsky, 2002; Palmer et al., 2010), either using pre-specified role ontologies such as PropBank or FrameNet (Palmer et al., 2005; Ruppenhofer et al., 2016), or, more recently, using natural-language- based representations (QA-SRL) (He et al., 2015; ∗Equal contribution. FitzGerald et al., 2018). Another well-studied kind of semantic relations between NPs is that of core- ference (Vilain et al., 1995; Pradhan et al., 2012), where two (or more) NPs refer to the same entity. Such NP-NP relations, which are either medi- ated by verbs (as in SRL or Relation Extraction) or form coreference relations, represent only a subset of the NP-NP relations that are naturally expressed in texts. Consider, for instance, the fol- lowing sentences: All of the above cases contain examples of NP- NP relations, where the type of relation can be expressed via an English preposition. The prepo- sition may be explicit in the text, as in (2), where the relation A person with blue eyes is explicitly expressed, or they may be implicit and left to the reader to infer, as in (3)-(4); in (3) readers easily infer that the roots are of the plant. Likewise, in (4) readers infer that the window is in the room.1 Properly understanding the text means knowing that these relations hold, even when they are not explicitly stated in the utterance. Figure 1 shows additional examples. These relations, both explicit and implicit, are indispensable for understanding the text. …

Content truncated to 500 words