Abstract:
The use of figurative language, with a major emphasis on metaphors and a
minor emphasis on oxymorons, has been widely accepted as part of everyday
language, not just in literary language. Over the last few years, there has also
been a large move towards automated sentiment analysis tlu·ough which diverse
corporations seek feedback on the sentiment (or affect: emotions, value
judgments, etc.) that customers bear towards their goods and services. The need
for this feedback has been particularly challenged by the use of social media,
which allow the use of non-literal language, including shorthand, abbreviations
and emoticons. Begilming with an overview of metaphors, sentiment analysis,
modifiers, and how they relate to each other in terms of conveying affect, this
thesis examines the accuracy of relying on lexical libraries like SentiWordNet
and WordNet in an attempt to extract sentiment-related information on language
in discourse. Following a set of empirical studies and experiments, I examined
how some existing systems are carrying out analysis of metaphors and
oxymorons, and how those systems evaluate metaphors that have made use of
modifiers. I demonstrate that modifiers do enhance the sentiment conveyed by
metaphors, though their placement within the metaphor, if not done well, can
distort the intended meaning, providing a good motivation for non-literal text
identification systems to be integrated into existing sentiment analysis systems.
I also prove by analysis that SentiWordNet has inherent inaccuracies that
introduce errors in sentiment extractions, and recommend that it is crucial to
identify non-literal text before sentiment is extracted in order to avoid incorrect
judgments
