French Multi Word Expressions: Using Data on Different Patterns for Extraction and Validation

Mutlti Word Epressions (MWE) are an important problem in NLP. Many researchers use association measures for collecting and evaluating MWE candidates. In this paper we propose to check if it is legitimate to use those measures when data are only collected on one pattern of MWE (e.g. NounAdjective) for evaluating candidates belonging to an other pattern (e.g. NounNoun). For this purpose, we run tests on the French Europarl corpus. Using association measures extracted from NounAdjective patterns as features, we train a model that we evaluate on instances of Noun-Noun candidates. We notice with this method that the model will still evaluate correctly a quarter of the candidates. However the result tend to be lower.