Information Technology Reference
In-Depth Information
to find the most suitable feature set for both the subjectivity detection and polarity
classification task. We train and evaluate our approach on a human-annotated corpus
of German quotations. The corpus consists of 742 neutral, 71 positive, and 38 negative
quotations. It can be made available for research purposes after signing an agreement.
1.4.2 Related Work
Related research on sentiment analysis varies from simple lexicon-based approaches
looking up words in opinion lexicons to supervised approaches exploiting linguistic
features and enhanced machine learning algorithms. The main part concentrates on
the classification of text as either positive, negative or neutral toward a specific
entity or topic explicitly mentioned in the text.
Sentiment analysis treats texts at different levels. There is work examining entire
document texts like entire reviews or news articles, attempting to predict the overall
sentiment of a document text [
37
,
49
]. But there is also work that performs sentiment
analysis at statement level [
5
,
18
,
46
], sentence level [
23
,
33
,
43
,
54
] or even phrase
level [
49
]. Often, sentiment analysis work on reviews also aims at extracting product
properties and the opinions toward these properties, which is called aspect-oriented
sentiment analysis [
21
].
Sentiment analysis has also been applied to different text types. A great part of
the work examines customer reviews, like product [
21
,
49
]ormovie[
37
]reviews.
Since reviews are meant to share experiences and report opinions, they contain many
subjective text parts and are therefore predestined for sentiment analysis. Yet, reviews
can also contain objective parts summarizing the properties of the reviewed entities.
Regarding movie reviews, one challenge is to separate plot information, which itself
may be characterized as positive or negative, from opinions toward the movie. All
work treating customer reviews must handle challenges arising from user-generated
content such as potential spelling mistakes and grammatical errors.
Early work in classifying product reviews used lexicon-based techniques together
with natural language processing algorithms in order to create opinion summariza-
tion. Hu and Liu [
21
] propose a three-stage approach to aspect-based opinion sum-
marization. They first search for product features in customer reviews by applying
association mining with some pruning. Then, the authors determine the polarity of
sentences mentioning the features. Whether a sentence has to be classified as posi-
tive or negative results from the orientation of the individual opinions words (adjec-
tives) in the sentence that is summed up to an overall orientation. The orientation of
opinion words is pre-calculated based on a list of seed adjectives and the applica-
tion of WordNet's information on synonyms and antonyms. Similar to Hu and Liu,
Turney [
49
] categorizes product reviews in either 'recommended' or not 'recom-
mended' by calculating the average sentiment orientation of the review's phrases.
Turney calculates the orientation of phrases containing adjectives and adverbs by
determining the mutual information between a phrase and the words “excellent” and
“poor” and subtracting both values to obtain a final sentiment orientation score.
Search WWH ::
Custom Search