DeModify: A Dataset for Analyzing Contextual Constraints on Modifier Deletion

DeModify: A Dataset for Analyzing Contextual Constraints on Modifier Deletion Vivi Nastase⋄, Devon Fritz, Anette Frank⋄ Leibniz ScienceCampus ”Empirical Linguistics and Computational Language Modeling”⋄ Institute of Computational Linguistics, Heidelberg University Heidelberg, Germany {nastase, frank}@cl.uni-heidelberg.de, devon.s.fritz@gmail.com Abstract Tasks such as knowledge extraction, text simplification and summarization have in common the fact that from a text fragment a smaller (not necessarily contiguous) portion is obtained by discarding part of the context. This may cause the text fragment to acquire a new meaning, or even to become false. The smallest units that can be considered disposable in a larger context are modifiers. In this paper we describe a dataset collected and annotated to facilitate the study of the influence of modifiers on the meaning of the context they are part of, and to support the development of models that can determine whether a modifier can be removed without undesirable semantic consequences. Keywords: annotation, modifier deletion, text simplification 1. Introduction Knowledge extraction, keyword identification, text simpli- fication, or summarization are useful tasks that rely on the assumption that certain information from texts can be dis- carded without negative consequences. Certain details are peripheral and can be disregarded. Removing subordinate clauses is a common practice in extractive summarization, for example (Vanderwende et al., 2007; Zajic et al., 2007). While modifiers can be quite complex – ranging from a sin- gle word to a full clause – we focus here on single-word modifiers, the smallest unit of context that can be removed. While even their name suggests a peripheral role in the overall meaning of a text, modifiers can impact both their local and global context. The impact of modifiers on their local context (themselves plus their syntactic head), also called modification distortion (Murphy, 2002), has been studied mostly out of context, and there are several datasets that allow for the study of this phenomenon (Kruszewski and Baroni, 2014; Schulte im Walde et al., ). The dataset we built and present here is focused on the effect of modi- fiers on the larger context. For example, the following title of a TED talk would mean something completely different should the modifier nearly be removed: AJ Jacobs: How healthy living nearly killed me While nearly could be argued to be a special modifier, the same situation may arise for ”normal” modifiers, such as the adjective old – it can be removed from the following context without a dramatic impact on the meaning of the sentence: Then she saw the old parish priest pull up in his car. but it is an essential element of the story, when the entire context is included – a short story from the ROC corpus (Mostafazadeh et al., 2016): Joan entered the confessional and kneeled. She thought she was confessing to the old parish priest. Joan confessed she had fantasized about the young visiting priest. Joan felt relief as she left the confessional. Then she saw the old parish priest pull up in his car. Understanding the influence of modifiers is important, …

Content truncated to 500 words