Data-driven identification of fixed expressions and their modifiability

This thesis originated from the need for a description of the linguistic behavior of Dutch fixed expressions in a computational model. This work addresses the following problems: (i) which criteria determine that an expression in language use has to be described as a fixed expression, (ii) how (automatic) identification models can capture such criteria and (iii) how to predict whether such expressions allow variation and modification. The properties and the behavior of fixed expressions are still poorly understood. The type and number of irregularities vary across expressions. The author argues that the idiosyncratic behavior affecting one or more levels of linguistic analysis suffices to establish that an expression qualifies as a fixed expression. To identify fixed expressions, Villada Moirón proposes the use of hybrid models that combine statistics and linguistic nowledge. In order to be successful, the identification models need to capture irregularities that affect the morphology, syntactic distribution and meaning, in addition to the lexical associations between the word components. To establish the potential for variation and modification, a corpus-based method extracts relevant evidence from anannotated corpus. The evidence allows us distinguish various types of support verb constructions. Villada Moiron observes that the presence of certain determiner changes correlate with a potential for modification, albeit less than perfectly. The methods need to be tested further with other types of fixed expressions but it seems that their design features are appropriate for the task. Such methods are crucial to improve the description of fixed expressions in lexical and computational resources.