Toward Improving the Automated Classification of Metonymy in Text Corpora

In this paper, we explore methods for improving the automatic classification of schematic metonymies in corpus-based text. Using a pre-existing dataset of thousands of samples, we formulate the hypothesis that a better modeling of the underlying syntactic, semantic and conceptual meanings within a document (sample) can aid automated metonymy classification. To test this hypothesis, we build upon previous researchers’ work on metonymy resolution systems but also introduce novel features, including, but not limited to, the extraction and analysis of conceptual paths between syntactically and semantically connected words. We initially explore three models for this classification task, but settle for final evaluation on two of them. Purely quantitatively, the results indicate that more work needs to be done, but on a more detailed analysis, we explain the strengths of our ideas behind our methods, and the weaknesses of the databases used for feature extraction. Finally, we present ways that this work can be continued and expanded. ∗This work was completed in partial fulfillment of the requirements for an Honors Bachelor of Science Degree in Computer Science from the Department of Computer Science at the University of Rochester, in Rochester, NY, USA.