Automatic Extraction of Lexical Functions

Abstract

Lexical function is a concept which formalizes semantic and syntactic relations between lexical units. Relations between words are a vital part of any natural language system. Meaning of an individual word largely depends on various relations connecting it to other words in context. Collocational relation is a type of institutionalized lexical relations which holds between the base and its partner in a collocation (examples of collocations: gives a lecture, make a decision, lend support where the bases are lecture, decision, support and the partners, termed collocates, are give, make, lend). Collocations are opposed to free word combination where both words are used in their typical meaning (for example, give a book, make a dress, lend money). Knowledge of collocation is important for natural language processing because collocation comprises the restrictions on how words can be used together. There are many methods to extract collocations automatically but their result is a plain list of collocations. Such lists are more valuable if collocations are tagged with semantic and grammatical information. The formalism of lexical functions is a means of representing such information. If collocations are annotated with lexical functions in a computer readable dictionary, it will allow effective use of collocations in natural language applications including parsers, high quality machine translation, periphrasis system and computer-aided learning of lexica. In order to create such applications, we need to extract lexical functions from corpora automatically. It is our intent to extract Spanish verb-noun collocations belonging to a given lexical function from corpora. To achieve this task, it has been proposed to represent the lexical meaning of a given word with a set of all its hyperonyms and to use machine learning techniques for predicting lexical functions as values of the class variable for unseen collocations. Hyperonyms are extracted from the Spanish WordNet. We evaluate many machine learning algorithms on the training set and on an independent test set. The obtained results show that machine learning is feasible to achieve the task of automatic detection of lexical functions.