Combining Wordnet and Morphosyntactic Information in Terminology Clustering

Proceedings of COLING 2012: Technical Papers, pages 1951–1962, COLING 2012, Mumbai, December 2012. Combining Wordnet and Morphosyntactic Information in Terminology Clustering Agnieszka Mykowiecka Małgorzata Marciniak Institute of Computer Science, PAS, Jana Kazimierza 5, 01-248 Warsaw, Poland agn@ipipan.waw.pl, mm@ipipan.waw.pl ABSTRACT The paper presents results of clustering terms extracted from economic articles in Polish Wikipedia. First, we describe the method of automatic term extraction supported by linguistic knowledge. Then, we define different types of term similarities used in the clustering experiment. Term similarities are based on Polish Wordnet and morphosyntactic analysis of data. The latter takes into account: term contexts, coordinated sequences of terms, syntactic patterns in which terms appear and words that are parts of terms (such as their heads and modifiers). Then we performed several experiments with hierarchical clustering of the 400 most frequent terms. We present the results of clustering when different groups of similarity coefficients are applied. Finally, we present an evaluation that compares the results with manually obtained groups. Our results prove that morphosyntactic information can help or even serve themselves for initial clustering of terms in semantically coherent groups. KEYWORDS: terminology extraction, terminology clustering, Polish. 1951 1 Introduction Many NLP applications, like text indexing, information extraction or question answering, rely on sets of predefined concepts which can be identified within texts, on a given subject. Such concepts form either hierarchical ontologies or flat terminology lists. Although there are already very many works aimed at making these kinds of resources available, still, the existing data is very limited with regard to the chosen application/knowledge domain as well as natural language addressed. An overview of existing approaches to automatic terminology extraction is included in (Pazienza et al., 2005) while problems with ontology creation are described, among others, in (Cimiano, 2006). For some NLP purposes flat terminology lists are sufficient, but for others, like IE, ontologies representing relations between particular terms are more adequate. Unfortunately, domain ontologies’ availability is very limited and what is even more crucial – they are rarely adequate for the purpose at hand. They are either too specific or too general, or do not cover the appropriate domain, or they are outdated. Ontology reuse and projection is still a very difficult and unresolved issue. As a result, when general ontologies like SUMO (Pease and Niles, 2001) are insufficient, or are not available in a particular language, a new dedicated ontology has to be created. In such a case, a list of concepts to be included in the ontology can be prepared manually by a domain expert, or can be (at least initially) extracted from texts similar to those which are to be processed. In this paper we address the problem of finding interesting concepts in Polish economic texts and organizing them in coherent groups which can be further analyzed more easily than a long unstructured term list. The identified sets of terms will be used for developing a domain model which will constitute the base of an information extraction system. The basic idea is to …

Content truncated to 500 words