Learning compound noun semantics

This thesis investigates computational approaches for analysing the semantic relations in compound nouns and other noun-noun constructions. Compound nouns in particular have received a great deal of attention in recent years due to the challenges they pose for natural language processing systems. One reason for this is that the semantic relation between the constituents of a compound is not explicitly expressed and must be retrieved from other sources of linguistic and world knowledge. I present a new scheme for the semantic annotation of compounds, describing in detail the motivation for the scheme and the development process. This scheme is applied to create an annotated dataset for use in compound interpretation experiments. The results of a dual-annotator experiment indicate that good agreement can be obtained with this scheme relative to previously reported results and also provide insights into the challenging nature of the annotation task. I describe two corpus-driven paradigms for comparing pairs of nouns: lexical similarity and relational similarity. Lexical similarity is based on comparing each constituent of a noun pair to the corresponding constituent of another pair. Relational similarity is based on comparing the contexts in which both constituents of a noun pair occur together with the corresponding contexts of another pair. Using the flexible framework of kernel methods, I develop techniques for implementing both similarity paradigms. A standard approach to lexical similarity represents words by their co-occurrence distributions. I describe a family of kernel functions that are designed for the classification of probability distributions. The appropriateness of these distributional kernels for semantic tasks is suggested by their close connection to proven measures of distributional lexical similarity. I demonstrate the effectiveness of the lexical similarity model by applying it to two classification tasks: compound noun interpretation and the 2007 SemEval task on classifying semantic relations between nominals. To implement relational similarity I use kernels on strings and sets of strings. I show that distributional set kernels based on a multinomial probability model can be computed many times more efficiently than previously proposed kernels, while still achieving equal or better performance. Relational similarity does not perform as well as lexical similarity in my experiments. However, combining the two models brings an improvement over either model alone and achieves state-of-the-art results on both the compound noun and SemEval Task 4 datasets. 3 4 Acknowledgments The past four years have been a never-boring mixture of learning and relearning, experiments that worked in …