Sentiment Analysis of Text Guided by Semantics and Structure

As moods and opinions play a pivotal role in various business and economic processes, keeping track of one's stakeholders' sentiment can be of crucial importance to decision makers. Today's abundance of user-generated content allows for the automated monitoring of the opinions of many stakeholders, like consumers. One challenge for such automated sentiment analysis systems is to identify whether pieces of natural language text are positive or negative. Typical methods of identifying this polarity involve low-level linguistic analysis. Existing systems predominantly use morphological, lexical, and syntactic cues for polarity, like a text's words, their parts-of-speech, and negation or amplification of the conveyed sentiment. This dissertation argues that the polarity of text can be analysed more accurately when additionally accounting for semantics and structure. Polarity classification performance can benefit from exploiting the interactions that emoticons have on a semantic level with words – emoticons can express, stress, or disambiguate sentiment. Furthermore, semantic relations between and within languages can help identify meaningful cues for sentiment in multi-lingual polarity classification. An even better understanding of a text's conveyed sentiment can be obtained by guiding automated sentiment analysis by the rhetorical structure of the text, or at least of its most sentiment-carrying segments. Thus, the sentiment in, e.g., conclusions can be treated differently from the sentiment in background information. The findings of this dissertation suggest that the polarity of natural language text should not be determined solely based on what is said. Instead, one should account for how this message is conveyed as well.

• View at Worldcat

Free Full Text ( Final Version , 1mb )