Extracting sentiment as a function of discourse structure and topicality

We present an approach to extracting sentiment from texts that makes use of contextual information. Using two di¤erent approaches, we extract the most relevant sentences of a text, and calculate semantic orientation weighing those more heavily. The …rst approach makes use of discourse structure via Rhetorical Structure Theory, and extracts nuclei as the relevant parts; the second approach uses a topic classi…er built using support vector machines, which extracts topic sentences from texts. The use of weights on relevant sentences shows an improvement over word-based methods that consider the entire text equally. In the paper, we also describe an enhancement of our previous word-based methods in the treatment of intensi…ers and negation, and the addition of other parts of speech beyond adjectives. This work was supported by the Natural Sciences and Engineering Research Council of Canada, under a Discovery Grant and a University Faculty Award to Maite Taboada. We also thank members of the Sentiment Research Group at SFU for their feedback, and in particular Vita Markman, Milan To…loski and Ping Yang for their help in ranking our dictionaries.