Figure 5

Word shift graphs for the movie review corpus. By analyzing the words that contribute most significantly to the sentiment score produced by each sentiment dictionary we are able to identify words that are problematic for each sentiment dictionary at the word-level, and generate an understanding of the emotional texture of the movie review corpus. Again we find that coverage of the lexicon is essential to produce meaningful word shift graphs, with the labMT dictionary providing the most coverage of this corpus and producing the most detailed word shift graphs. All words on the left hand side of these word shift graphs are words that individually made the positive reviews score more negatively than the negative reviews, and the removal of these words would improve the accuracy of the ratings given by each sentiment dictionary. In particular, across each sentiment dictionary the word shift graphs show that domain-specific words such as ‘war’ and ‘movie’ are used more frequently in the positive reviews and are not useful in determining the polarity of a single review.