Affective Text: Generation Strategies and Emotion Measurement Issues

In affective natural language generation (NLG) a major aim is to be able to influence the emotional effects evoked in the addressee through the intelligent use of language. While previous work has shown that varying the form of the language, while keeping the content the same, can have a measurable effect on the emotions of the addressee, we report here on work which investigated which linguistic techniques to give the text a more or less positive slant contribute to these emotional effects. We report on three studies in which texts that gave positive feedback on an IQ test performance were tested for emotional effects on the recipient. The first study followed a comparison method on the sentence level, and the second study compared the texts as a whole. In both of these, participants were asked to rate the emotional effects that they thought the texts would have. On the other hand, in the third study different types of feedback were evaluated in a context of use, where participants were asked to perform an IQ test, read their feedback and report on their actual emotional state. In the first two studies, participants confirmed that the texts contained essentially the same content. The positive slanting techniques generally resulted in texts that were judged to be either positive or equal to neutral texts, although the effects were less strong than in previous work, which employed a variety of techniques, and there were a number of exceptions which impact on the usefulness of these techniques. However the IQ-test experiment did not show any emotional effects arising from variation in the form of the feedback. We reflect on possible reasons for this outcome and what it might mean for further work on Affective NLG.