This paper presents a new approach to phrase-level sentiment analysis that first determines whether an expression is neutral or polar and then disambiguates the polarity of the polar expressions. With this approach, the system is able to automatically identify the contextual polarity for a large subset of sentiment expressions, achieving results that are significantly better than baseline. The paper introduces a new method for phrase-level sentiment analysis that involves determining whether an expression is neutral or polar, and then disambiguating the polarity of polar expressions. The system uses a two-step process that employs machine learning and various features. The first step classifies each phrase containing a clue as neutral or polar. The second step takes all phrases marked in step one as polar and disambiguates their contextual polarity (positive, negative, both, or neutral). The system is able to automatically identify the contextual polarity for a large subset of sentiment expressions, achieving results that are significantly better than baseline. To create a corpus for the experiments, the authors added contextual polarity judgments to existing annotations in the Multi-perspective Question Answering (MPQA) Opinion Corpus. They developed an annotation scheme for marking the contextual polarity of subjective expressions. Annotators were instructed to tag the polarity of subjective expressions as positive, negative, both, or neutral. The authors conducted an agreement study with two annotators, using 10 documents from the MPQA Corpus. The results showed a high level of agreement between the annotators. The authors also developed a prior-polarity subjectivity lexicon containing over 8,000 subjectivity clues. These clues are words and phrases that may be used to express private states. The authors used this lexicon to train a classifier that classifies the contextual polarity of expressions containing instances of the subjectivity clues. The classifier uses a two-step process to disambiguate the contextual polarity of the expressions. The first step classifies each phrase containing a clue as neutral or polar. The second step takes all phrases marked in step one as polar and disambiguates their contextual polarity. The authors evaluated the performance of the classifier on a development set and found that it achieved an accuracy of 75.9%, with a polar F-measure of 63.4 and a neutral F-measure of 82.1. The results showed that the classifier performed significantly better than simpler classifiers. The authors also conducted experiments to determine the contribution of various polarity features to the performance of the classifier. The results showed that the combination of features was needed to achieve significant results over baseline for polarity classification. The authors conclude that their approach to phrase-level sentiment analysis is able to automatically identify the contextual polarity for a large subset of sentiment expressions, achieving results that are significantly better than baseline.This paper presents a new approach to phrase-level sentiment analysis that first determines whether an expression is neutral or polar and then disambiguates the polarity of the polar expressions. With this approach, the system is able to automatically identify the contextual polarity for a large subset of sentiment expressions, achieving results that are significantly better than baseline. The paper introduces a new method for phrase-level sentiment analysis that involves determining whether an expression is neutral or polar, and then disambiguating the polarity of polar expressions. The system uses a two-step process that employs machine learning and various features. The first step classifies each phrase containing a clue as neutral or polar. The second step takes all phrases marked in step one as polar and disambiguates their contextual polarity (positive, negative, both, or neutral). The system is able to automatically identify the contextual polarity for a large subset of sentiment expressions, achieving results that are significantly better than baseline. To create a corpus for the experiments, the authors added contextual polarity judgments to existing annotations in the Multi-perspective Question Answering (MPQA) Opinion Corpus. They developed an annotation scheme for marking the contextual polarity of subjective expressions. Annotators were instructed to tag the polarity of subjective expressions as positive, negative, both, or neutral. The authors conducted an agreement study with two annotators, using 10 documents from the MPQA Corpus. The results showed a high level of agreement between the annotators. The authors also developed a prior-polarity subjectivity lexicon containing over 8,000 subjectivity clues. These clues are words and phrases that may be used to express private states. The authors used this lexicon to train a classifier that classifies the contextual polarity of expressions containing instances of the subjectivity clues. The classifier uses a two-step process to disambiguate the contextual polarity of the expressions. The first step classifies each phrase containing a clue as neutral or polar. The second step takes all phrases marked in step one as polar and disambiguates their contextual polarity. The authors evaluated the performance of the classifier on a development set and found that it achieved an accuracy of 75.9%, with a polar F-measure of 63.4 and a neutral F-measure of 82.1. The results showed that the classifier performed significantly better than simpler classifiers. The authors also conducted experiments to determine the contribution of various polarity features to the performance of the classifier. The results showed that the combination of features was needed to achieve significant results over baseline for polarity classification. The authors conclude that their approach to phrase-level sentiment analysis is able to automatically identify the contextual polarity for a large subset of sentiment expressions, achieving results that are significantly better than baseline.