Recognising affect in text using pointwise-mutual information

Abstract

This dissertation describes experiments conducted to evaluate an algorithm that attempts to automatically recognise emotions (affect) in written language. Examples from several areas of research that can inform affect recognition experiments are reviewed, including sentiment analysis, subjectivity analysis, and the psychology of emotion. An affect annotation exercise was carried out in order to build a suitable set of test data for the experiment. An algorithm to classify according to the emotional content of sentences was derived from an existing technique for sentiment analysis. When compared against the manual annotations, the algorithm achieved an accuracy of 32.78%. Several factors indicate that the method is making slightly informed choices, and could be useful as part of a holistic approach to recognising the affect represented in text. ii Acknowledgements Thanks to John Carroll, not only for supervision and helpful advice during this project but also for lectures in Natural Language Processing. Thanks also to Diana McCarthy and the other members of NLCL for their kind welcome and discussions during the STATNLP reading group.