Valence Shifting: Is It A Valid Task?

Valence Shifting: Is It A Valid Task? Mary Gardiner Centre for Language Technology Macquarie University mary@mary.gardiner.id.au Mark Dras Centre for Language Technology Macquarie University mark.dras@mq.edu.au Abstract This paper looks at the problem of va- lence shifting, rewriting a text to preserve much of its meaning but alter its senti- ment characteristics. There has only been a small amount of previous work on the task, which appears to be more difficult than researchers anticipated, not least in agreement between human judges regard- ing whether a text had indeed had its va- lence shifted in the intended direction. We therefore take a simpler version of the task, and show that sentiment-based lexical para- phrases do consistently change the senti- ment for readers. We then also show that the Kullback-Leibler divergence makes a useful preliminary measure of valence that corresponds to human judgements. 1 Introduction This paper looks at the problem of VALENCE SHIFTING, rewriting a text to preserve much of its meaning but alter its sentiment characteris- tics. For example, starting with the sentence If we have to have another parody of a post-apocalyptic America, does it have to be this bad?, we could make it more negative by changing bad to abom- inable. Guerini et al. (2008) say about valence shifting that it would be conceivable to exploit NLP techniques to slant original writings to- ward specific biased orientation, keep- ing as much as possible the same mean- ing . . . as an element of a persuasive system. For instance a strategic plan- ner may decide to intervene on a draft text with the goal of “coloring” it emo- tionally. There is only a relatively small amount of work on this topic, which we review in Section 2. From this work, the task appears more difficult than re- searchers originally anticipated, with many fac- tors making assessment difficult, not least the re- quirement to be successful at a number of differ- ent NLG tasks, such as producing grammatical output, in order to properly evaluate success. One of the fundamental difficulties is that it is difficult to know whether a particular approach has been successful: researchers have typically had some trouble with inter-judge agreement when evaluat- ing whether their approach has altered the senti- ment of a text. This casts doubt on valence shift- ing as a well-defined task, although intuitively it should be, given that writing affective text has a very long history, computational treatment of sentiment-infused text has recently been quite ef- fective. This encourages us to start with a simpler task, and show that this simpler version of valence shifting achieves agreement between judges on the direction of the change. Consequently, in this paper we limit ourselves to exploring valence shifting by lexical substitution rather than explor- ing richer paraphrasing techniques, and testing this on manually constructed sentences. We then explore two questions: 1. Is it in fact true that altering a single lexi- cal item in a sentence noticeably changes its sentiment for readers? 2. Is there …

Content truncated to 500 words