| Resumen |
In such verb-noun combinations as draw a conclusion, lend support, take a step, the verb acquires a meaning different from its typical meaning usually represented by the first sense in WordNet thus making a correct compositional analysis hard or even impossible. Such non-compositional word combinations are called collocations. The semantics and syntactical properties of collocations can be formalized using lexical functions, a concept of the Meaning-Text Theory. In this paper we realized two series of experiments, both with supervised learning methods on automatic detection of lexical functions in verb-noun collocations using WordNet hypernyms. In the first experimental series, we used hypernyms which correspond to the manually annotated WordNet senses of verbs and nouns in the dataset. In the second series, we used hypernyms corresponding to the typical (first) sense of the verbs. Comparing the results of both experiments we found that the performance of supervised learning on some lexical functions was better in the second case in spite of the fact that the first sense was not the sense of the verbs they have in collocations. This shows that for such lexical functions, the semantics of the verbs is closer to their typical senses and thus non-compositionality of such collocations is weaker. We propose to use the difference in lexical function detection based on the actual sense and the first sense as a simple measure of non-compositionality of verb-noun collocations. |
