This thesis is about the annotation of metaphorical language in historical corpora and what computational resources can support that process. Metaphor is a kind of extended language use in which the literal meaning nevertheless plays a central role in what speakers intend to communicate. Literal meanings themselves are established through use, and they are in constant flux, both on the micro level, from use to use, and on the macro level, over the course of decades and centuries. Therefore it can be difficult to know, from a historical distance, what literal meanings were established at a given time and place, which makes metaphor analysis in historical corpora especially challenging. Thus it is the goal of this project to investigate computational methods of determining the presence and prevalence of word senses in a way that is tunable to time- or genre-specific portions of a particular corpus and still usable when only small amounts of corpus data are available. It is hypothesized that this kind of information about the literal meanings reflected in a particular corpus would provide the kind of resource that would be helpful in the analysis of metaphor in a historical setting.
After an initial theoretical consideration of the properties of metaphor and a survey of the means of metaphor identification suggested in metaphor annotation guidelines, I select two existing algorithms, one for learning sense embeddings and one for specializing embeddings for particular subcorpora, to use for further investigations. I perform a series of evaluations of various aspects of these algorithms and explain what modifications were necessary to combine them. The evaluations showed that each of these algorithms provides a usable degree of performance in isolation, and the combination of the two results in no decline in performance.
Then, in a further set of experiments, I examine whether or not the outputs that are available from this diachronic sense embedding approach correlate with the metaphor annotations in existing datasets. Such a correlation is something I would consider to be indicative of the potential utility of this kind of information for supporting the metaphor annotation process. The initial results do reflect such a correlation. A further exploration of the data indicates that the system outputs could provide useful information for characterizing conventional and novel metaphor in distributional terms, suggesting a promising avenue for future work.
erwv2cdo: