What’s Great and What’s Not: Learning to Classify the Scope of Negation for Improved Sentiment Analysis

by Councill et. al - Proceedings of the Workshop on Negation and Speculation in NLP (July, 2010)

This paper uses conditional random fields to detect negations and their scope for sentiment detection. Based on Tottie (1991) the authors distinguish between the following kinds of negations:

1. Denials, which are the most common forms of negations
2. Rejections, which often occur in a discurse ("Can I get you anything? No.")
3. Imperatives such as "Do not use the elevator!"
4. Questions, can often indicate negation -- Why wouldn't they sell this good at a resonable price?
5. Supports and Repetitions

The authors draw upon a manually annotated Product Review Corpus to evaluate their work and present a list of explicat negation clues which they use to detect negated sentences. Annotating text with (i) sentence boundaries and (ii) token annotations from a dependency parser helps determining the negation scope. The authors also discuss how the word POS (adverb, adjective, noun, etc.) influences the negation scope. The paper's evaluation focuses on (i) the performance of the negation scope detection, and (ii) the improvement negation detection yields to sentiment detection. It shows that considering negation for sentiment detection considerably improved the F-scores of the sentiment predictions (+ 29.5% for positive and +11.4% for negative sentences)

Interesting Remarks:

The paper uses a sentiment lexicon extracted from the web by using label propagation (see Velikovich et al. - "The viability of web-derived polarity lexicons" Proceedings of the 11th Annual Conference of the North American Chapter of the Association for Computational Linguistics).