Natural Language Processing used in sentiment analysis of poetry: a study of six common techniques

Abstract

This paper explores and compares the accuracies of six different NLP model techniques when applied to analysis of English poetry. It was found that bigram-based and word embedding-based models were the most accurate in deriving emotions from the bodies of text in the corpus, with respective accuracies of 61.67% and 61.68%. The least accurate models were the unigram-based model and distributed dictionary-based model, constructed using the traditional approach rather than the new methodology, and had accuracies of 11.173% and 48.17% respectively. Most models passed the benchmark accuracies of 49.42% and 47.86%, the higher accuracy one being a word count model and lower one being a sentiment model. The need for newer methodologies that allow for higher dimensional levels of semantic analysis to be performed is also discussed in this paper, along with the potential impacts of this research to the field of Natural Language Processing.