Extracting Clusters of Specialist Terms from Unstructured Text

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1426–1434, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Extracting Clusters of Specialist Terms from Unstructured Text Aaron Gerow Computation Institute University of Chicago Chicago, IL, USA gerow@uchicago.edu Abstract Automatically identifying related special- ist terms is a difficult and important task required to understand the lexical struc- ture of language. This paper develops a corpus-based method of extracting co- herent clusters of satellite terminology — terms on the edge of the lexicon — us- ing co-occurrence networks of unstruc- tured text. Term clusters are identi- fied by extracting communities in the co- occurrence graph, after which the largest is discarded and the remaining words are ranked by centrality within a community. The method is tractable on large corpora, requires no document structure and min- imal normalization. The results suggest that the model is able to extract coher- ent groups of satellite terms in corpora with varying size, content and structure. The findings also confirm that language consists of a densely connected core (ob- served in dictionaries) and systematic, se- mantically coherent groups of terms at the edges of the lexicon. 1 Introduction Natural language consists of a number of rela- tional structures, many of which can be obscured by lexical idiosyncrasies, regional variation and domain-specific conventions. Despite this, pat- terns of word use exhibit loose semantic struc- ture, namely that proximate words tend to be re- lated. This distributional hypothesis has been op- erationalized in a variety of ways, providing in- sights and solutions into practical and theoretical questions about meaning, intention and the use of language. Distributional analyses rely primarily on observing natural language to build statistical representations of words, phrases and documents (Turney and Pantel, 2010). By studying dictio- naries and thesauri, lexicographic and terminolog- ical research has proposed that a core lexicon is used to define the remaining portions of vocabu- lary (Itˆo and Mester, 1995; Mass´e et al., 2008). Though many words that comprise general lan- guage use reside in this core lexicon, even the most general language contains specialist or so- called “satellite” words. This paper introduces a method of extracting this peripheral structure, with co-occurrence networks of unstructured text. The core-periphery structure has been observed in dictionaries where definitions tend to use a re- stricted vocabulary, repetitively employing a core set of words to define others (Sinclair, 1996; Pi- card et al., 2013). In the farther regions of the lex- icon, it is more difficult to find systematic seman- tic definition with corpus-based techniques due to the overwhelming number of infrequent words. Unfortunately, the fringe of the lexicon can be more important than the core because this is where domain-specific terminology resides — features that may be more important than frequent. Examining dictionaries, (Picard et al., 2013) propose that the lexicon consists of four main parts: a core set of ubiquitous words used to de- fine other words, a kernel that makes up most of the lexicon, a …

Content truncated to 500 words