Enhanced English Universal Dependencies: An Improved Representation for Natural Language Understanding Tasks

Enhanced English Universal Dependencies: An Improved Representation for Natural Language Understanding Tasks Sebastian Schuster and Christopher D. Manning Linguistics Department and Computer Science Department Stanford University, Stanford, CA 94305 {sebschu,manning}@stanford.edu Abstract Many shallow natural language understanding tasks use dependency trees to extract relations between content words. However, strict surface-structure dependency trees tend to follow the linguistic structure of sentences too closely and frequently fail to provide direct relations between content words. To mitigate this problem, the original Stanford Dependencies representation also defines two dependency graph representations which contain additional and augmented relations that explicitly capture otherwise implicit relations between content words. In this paper, we revisit and extend these dependency graph representations in light of the recent Universal Dependencies (UD) initiative and provide a detailed account of an enhanced and an enhanced++ English UD representation. We further present a converter from constituency to basic, i.e., strict surface structure, UD trees, and a converter from basic UD trees to enhanced and enhanced++ English UD graphs. We release both converters as part of Stanford CoreNLP and the Stanford Parser. Keywords: Universal Dependencies, semantic representation, treebank conversion 1. Introduction Since its first version, the Stanford Dependencies (SD) rep- resentation (de Marneffe et al., 2006) has had the status of being both a syntactic and a shallow semantic representa- tion. This dual status is also reflected in the usage of SD in natural language processing tasks which broadly fall into two categories. The first category is composed of tasks that require a syntactic tree such as source-side reorder- ing for machine translation (e.g., Genzel (2010)) and sen- tence compression (Galanis and Androutsopoulos, 2010). For these tasks, a sound syntactic representation is more important than the relations between individual words. The second and much larger category is composed of a wide range of shallow natural language understanding (NLU) tasks such as biomedical text mining (e.g., Airola et al. (2008)), open domain relation extraction (e.g., Mausam et al. (2012)), and unsupervised semantic parsing (Poon and Domingos, 2009). For these tasks, the relations be- tween content words are more important than the overall tree structure. Not surprisingly, we observe a similar divide if we look at which one of the three SD representations is being used for the individual downstream tasks. Most systems that re- quire a syntactic representation use basic SD trees which are guaranteed to be a strict surface syntax tree. On the other hand, most systems that are concerned with the rela- tions between content words use the collapsed or CCpro- cessed SD representations. These representations may be graphs instead of trees, and may contain additional and aug- mented relations that explicitly capture otherwise implicit relations between content words. To illustrate one of the differences between these represen- tations, consider the sentence “Fred started to laugh”. The basic SD representation of this sentence lacks a direct re- lation between the controlled verb laugh and its controller, Fred, while in the CCprocessed SD representation, this re- lation is made explicit with an additional subject edge. …

Content truncated to 500 words