Learning Negation Scope from Syntactic Structure

Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics (*SEM), pages 137–142 Barcelona, Spain (Online), December 12–13, 2020 137 Learning Negation Scope from Syntactic Structure Nick McKenna School of Informatics University of Edinburgh nick.mckenna@ed.ac.uk Mark Steedman School of Informatics University of Edinburgh steedman@inf.ed.ac.uk Abstract We present a semi-supervised model which learns the semantics of negation through analysis of syntactic structure.purely Lin- guistic theory posits that the semantics of nega- tion can be understood purely syntactically, though recent research relies on combining a variety of features including part-of-speech tags, word embeddings, and semantic repre- sentations to achieve high task performance. Our simplified model returns to syntactic the- ory and achieves state-of-the-art performance on the task of Negation Scope Detection while demonstrating the tight relationship between the syntax and semantics of negation. 1 Introduction Negation is a semantic phenomenon in natural lan- guage which varies significantly. For example, “Sherlock did not solve the case” contains a simple negation cued by “not.” However, there are many cue words such as “without” and “nothing,” and word affixes like “un-” which instantiate negation, and their effect on meaning can be different and dependent on context. We approach the meaning of negation using a logical semantics, in which natural language nega- tion is expressed with the negation operator on logical expressions (Horn, 1989; Horn and Wans- ing, 2017), such as in ¬SOLVE(Sherlock, case). Further, in truth-theoretic logic the meaning of negation is simply the inverted truth value of this expression. Capturing the meaning of a negation cue in language can thus be understood as simply identifying the negated expression. This is the task of Negation Scope Detection (NSD): given a nega- tion cue in a sentence, identify the sentence tokens which make up the negated expression. This work is licensed under a Creative Commons Attri- bution 4.0 International License. License details: http: //creativecommons.org/licenses/by/4.0/. 1.1 Negation Scope Detection Task The task of NSD has been approached in several formulations using different annotation schemes (Kim et al., 2008; Szarvas et al., 2008), but these older tasks do not define negation semantically, in contrast to the commonly accepted *SEM2012 Shared Task competition (*SEM, 2012). Under the *SEM definition cues scope over events and partic- ipants, either directly or via predicate arguments and complements. Linguistic theory explains nega- tion semantics in terms of events, which are built from syntactic phrase structures (Huddleston and Pullum, 2005), i.e. that negation doesn’t just scope over individual words, but rather whole phrases and clauses which make up events. The following examples drawn from the *SEM2012 dataset of Conan Doyle writing illustrate how scope is built from phrase structures. Cues are in bold* with underlined underlined underlined underlined underlined underlined underlined underlined underlined underlined underlined underlined underlined underlined underlined underlined underlined scope. (1) Well, sir, I thought no* good could come of it good could come of it good could come of it good could come of it good could come of it good could come of it good could come of it …

Content truncated to 500 words