Bootstrapping Semantic Lexicons for Technical Domains

International Joint Conference on Natural Language Processing, pages 1321–1329, Nagoya, Japan, 14-18 October 2013. Bootstrapping Semantic Lexicons for Technical Domains Patrick Ziering1 Lonneke van der Plas1 Hinrich Sch¨utze2 1Institute for NLP, University of Stuttgart, Germany 2CIS, University of Munich, Germany {Patrick.Ziering, Lonneke.vanderPlas}@ims.uni-stuttgart.de Abstract We address the task of bootstrapping a se- mantic lexicon from a list of seed terms and a large corpus. By restricting to a small subset of semantically strong pat- terns, i.e., coordinations, we improve re- sults significantly. We show that the re- striction to coordinations has several ad- ditional benefits, such as improved ex- traction of multiword expressions, and the possibility to scale up previous efforts. 1 Introduction High-quality semantic lexicons are needed for many natural language processing (NLP) tasks like information extraction and discourse process- ing (Riloff and Shepherd, 1997). Building such lexicons manually is costly and time-consuming. Automatic lexicon construction is therefore an im- portant task and much prior work has addressed it. This paper adopts Basilisk (Thelen and Riloff, 2002) as its basic approach, a system that uses lexico-syntactic patterns for bootstrapping. We have adapted Basilisk to our setting in several ways. Whereas the original Basilisk covers a wide variety of lexico-syntactic patterns, we restrict ourselves to a specific type of patterns, i.e., coordi- nations. Coordinations have been exploited in lex- ical acquisition before (e.g., Roark and Charniak (1998); Caraballo (1999); Goyal et al. (2010)). Most of this previous work uses a pairwise per- spective (i.e., a focus on whether two words co- occur in a coordination). However, we use coor- dinations in a Basilisk approach, for which pat- terns contain several terms in general. We there- fore do not split up coordinations in pairs but keep the complete coordination intact. Coordinations in technical domains frequently contain more than 2 elements. We argue that bootstrapping methods, known to be particularly sensitive to the ambiguity of terms and contexts and prone to semantic drift, bene- fit from the strong semantic coherence found in coordinations. The elements of a coordination often have a common hypernym; i.e., they are co-hyponyms or members of a common semantic class. Furthermore, the high arity of coordinations, the fact that coordinations have two or more ar- guments (e.g., “platinum, nickel, palladium, cop- per, silver, or gold”) further constrains the seman- tics. For example, if two out of three elements in a coordination have already been identified as sub- stances, this makes it likely that the third is also a substance. General or lexico-syntactic patterns in lexical bootstrapping have arity 1, i.e., they have a single argument. These patterns are too often not restrictive enough to prevent false terms in- fecting the semantic lexicon, which can lead to semantic drift. For example, given a corpus in which the semantic class DISEASE is not predomi- nant, after some iterations the weak pattern “treat- ment of <X>” is selected, which also provides many off-class candidates (e.g., treatment of pris- oners). In coordinations, the semantic coherence among terms leads to more selective patterns. For …

Content truncated to 500 words