The 31st UW/MS Symposium in Computational Linguistics

Опубликовано 8 августа 2016, 19:57

Woodley Packard and Emily M. Bender: Predicting the Scope of Negation using Minimal Recursion Semantics: Negation is a pervasive phenomenon in natural language, occurring in every language and every genre. Despite the obviously profound impact of negation on the meaning of a sentence, the most common approach to handling negation in NLP systems is to ignore it, leading to all manner of (frequently) comical errors. To encourage the exploration of better solutions, the 2012 *SEM Shared Task focused (among other things) on automatically identifying negation and determining its scope. Several of the resulting systems were quite successful, but despite the semantic nature of the task, the vast majority of them were based on surface or syntactic methods. In this talk, we will describe a semantics-based method of attacking the same problem. Our system is based on the Minimal Recursion Semantics structures produced by the English Resource Grammar, a broad-coverage, precision, computational HPSG account of English. We show that it is relatively straightforward to design high precision rules to determine what portion of a sentence is within the scope of negation, by "crawling" through these graphs. In a system combination with the winner of the 2012 competition, our method yields improved precision and F1. Moreover, our "crawling" rules can be seen as a first-pass formalization of the shared task annotation guidelines. Margaret Mitchell: Generating human reference to visible objects: In this talk, I will detail some previous work on how people refer to everyday objects in real world settings, and discuss how to model this in a generation system that produces humanlike descriptions. This talk will tie in aspects of linguistics, cognitive science, and statistical natural language processing.