B651: Sentiment Analysis

A subjective text

Components of an opinion: opinion holder, object, opinion itself

Object

Set of features F, each with an associated synonym sent
Synonym set for each feature W
F and W may be unknown

Structure of the text: opinion holder selects subset of features, refers to each with a word/phrase from the associated synset, expresses a positive, negative, or neutral opinion of each

Unsupervised document-level sentiment classification (Turney)

Extract two-word phrases containing adjectives or adverbs from POS-tagged corpus

Calculate semantic orientation of phrases

Based on pointwise mutual information of phrase and the words "excellent" and "poor"
PMI(w₁, w₂) = log₂ [p(w₁ & w₂) / p(w₁) p(w₂)]
SO(phrase) = PMI(phrase, "excellent") - PMI(phrase, "poor")
PMI estimated using AltaVista NEAR (within 10 words) queries
SO(phrase) = log₂ [(hits(phrase NEAR "excellent") hits("poor")) / (hits(phrase NEAR "poor") hits("excellent"))]
SO examples: online service: 2.78, small part: 0.053, inconveniently located: -1.54 unethical practices: -8.48

Review is positive ("recommended") if average SO of phrases in it is positive, negative ("not recommended") otherwise

Accuracy of 74% across four review domains: cars, banks, movies (lowest), travel destinations

Supervised document-level sentiment classification (Pang, Lee, and Vaithyanathan)

Starred movie reviews from IMDb

Document represented as bag of features

Three classifiers tried: naive Bayes, maximum entropy, support vector machines

Features: unigrams (with NOT tag added where negation found), bigrams

Some results

SVM superior (83% for unigrams)
Unigrams superior to bigrams
Worse results when frequency of features taken into account
No improvement with POS attached, restriction to adjectives, positional indication
Difficulty of "thwarted expectations" narrative

Sentence-level sentiment classification (Ding and Liu)

Assign positive, negative, or neutral evaluation to an object feature

Context-independent "opinion words": poor, great, etc., each with an assigned orientation (1, -1, 0)

Split sentences at BUT words

In each sentence segment mentioning feature, sum opinion word orientations weighted by distance from feature word

Context-independent opinion words (the battery life is very long)

Infer orientation based on co-occurrence with context-independent opinion words: this camera takes great pictures and has a long battery life
Unless there is a BUT word: this camera takes poor pictures but has a long battery life