14 February – Pia Sommerauer

Tweet

Recognizing conventional nominal metaphors in natural language: Implementing the Metaphor Identification Procedure – VU with WordNet

In my term paper (supervised by Antske Fokkens), I tried to implement an approach to automatic metaphor identification in natural language that adheres to the most important aspects of the Metaphor Identification Procedure VU (MIPVU).
The rule-based approach aims to identify conventional nominal metaphors by distinguishing basic from metaphorical meanings in WordNet. The main goal of the system is to identify basic meanings of polysemous nouns and compare them to the output of a word sense disambiguation (WSD) module. If the output of the WSD module matches the basic meanings, a noun is annotated as used literally. Otherwise, it is annotated as used metaphorically.
I experimented with four ranking systems, each of which employs a different mechanism to rank the senses of a noun from basic to metaphorical. The ranking systems can be employed individually or in combination with each other. The results of the evaluation indicate that hyponymy of and similarity to a manually compiled list of concrete, body-related concepts yield the best results.
Even though the overall performance remains rather low, the experiments yield some insights into the distribution of basic and metaphorical senses in WordNet.