Abstract/Summary

Sentiment analysis is viewed generally as a text classification problem involving the prediction of the semantic orientation of a text. Much of the analysis has focused on the sentiment expressed in the sentence or by the writer but not the sentiment of the recipient. For example, the sentence “Housing costs have dropped significantly” might be assigned a negative classification by a sentiment analysis model, however humans from different works of life might express different sentiments. A landlord will likely express a negative sentiment while a renter might express a positive sentiment. Therefore, traditional sentiment analysis methods fail to capture the human centric aspects that motivate diverse sentiments. Additionally, attempts at predicting recipient sentiment have involved considerable human effort in the form of content analysis and empirical surveys, making the process expensive and time-consuming. Thus, the aim of this research is to develop a method of recipient sentiment analysis that is devoid of human input in the form of annotations or empirical surveys. The approach taken in this research involves applying a model of human values towards recipient sentiment prediction. The justification for this approach is based on the well-established principle that values influence human behaviour of which sentiment is a form. Therefore, if a persons’ values can be modelled quantitatively, when presented with some text, in theory the sentiment of the recipient can be predicted. This research proposes that the application of values in developing sentences is a generative process, that can be represented as a language model. A mechanism called Feature Switching (FS) that enables the determination of recipient’s sentiment from the value language model is also discussed. The resulting sentiment prediction model has an accuracy in the range of 72.2%-72.5% which is in and about the range of performance of existing systems which make use of content analysis and human annotated data.