Using unmarked contexts in nominal lexical semantic classification

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 508–519, Dublin, Ireland, August 23-29 2014. Using unmarked contexts in nominal lexical semantic classification Lauren Romeo*, Sara Mendes*,†, Núria Bel* *Universitat Pompeu Fabra, Roc Boronat, 138, Barcelona, Spain †Centro de Linguística da Universidade de Lisboa, Av. Prof. Gama Pinto, 2, Lisboa, Portugal {lauren.romeo,sara.mendes,nuria.bel}@upf.edu Abstract The work presented here addresses the use of unmarked contexts in pattern-based nominal lexical semantic classification. We define unmarked contexts to be the counterposition of the class-indicatory, or marked, contexts. Its aim is to evaluate how unmarked contexts can be used to improve the accuracy and reliability of lexical semantic classifiers. Results demonstrate that the combined use of both types of distributional in- formation (marked and unmarked) is crucial to improve classification. This result was replicated using two different corpora, demonstrating the robustness of the method proposed. 1 Introduction Lexical resources annotated with lexical semantic classes have been successfully incorporated into a wide range of NLP applications, such as grammar induction (Agirre et al., 2011) and the building and extending of semantic ontologies (Abbès et al., 2011). However, lexical semantic tagging in large lexica is mostly done by hand, implying high costs with regard to maintenance and domain tuning. As the use of an inadequate lexicon is one of the causes of poor performance of NLP applications, current research to improve the automatic production of rich language resources, and of class-annotated lexica, in particular, is critical. One way to approach this task is through supervised cue-based lexical semantic classification. Based on the distributional hypothesis (Harris, 1954), according to which words occurring in the same contexts can be said to belong to the same class, cue-based lexical semantic classification uses particu- lar linguistic contexts where nouns occur as cues that represent distinctive distributional traits of a lexical class. Yet, training a classifier with information about word occurrences in a corpus within a selected number of contexts can present a challenge, mainly because specific words might be observed in a number of class-indicative contexts but not always are. This type of marked, or class-indicative, context (e.g. co-occurrence with specific prepositions, predicate selectional restrictions, and grammatical information, such as indirect objects) are sparse in any corpus as, being so specific, they do not occur often with each target noun. Using only exclusive class-indicative contexts as features in nominal lexical semantic classification has been shown to not always provide sufficient information to make a decision regarding class membership of a noun (Bel et al., 2012), especially when the data does not contain relevant co-occurrences or when those co- occurrences are too disperse to be correlated. Recent work on the use of distributional models for nominal classification tasks (Romeo et al., 2014) discusses potential bottlenecks of models using data extracted with lexico-syntactic patterns as features, identifying data sparsity as one of the major issues affecting the performance of these sys- tems. In fact, the selection of class-indicative information, in an attempt to provide relevant informa- tion to …

Content truncated to 500 words