Learning Subjective Adjectives from Corpora Janyce M Wiebe

Learning Subjective Adjectives from Corpora Janyce M. Wiebe Presenter: Gabriel Nicolae

Introduction n Subjectivity in natural language refers to aspects of language used to express opinions and evaluations (Banfield 1982; Wiebe 1994) n Subjectivity tagging is distinguishing sentences used to present opinions and other forms of subjectivity (subjective sentences) from sentences used to objectively present factual information (objective sentences).

Why do we need subjectivity tagging? Because: Apart from the subject of the document, additional components influence its relevance: n the evidential status of the material presented n attitudes adopted (Hatzivassiloglou) The task is especially relevant for: n News reporting n Internet forums – recognizing flames Applications for which it is relevant: n Information extraction n Information retrieval

Subjectivity – Examples (1/2) Simple subjective sentence: At several different layers, it’s a fascinating tale. Simple objective sentence: Bell Industries Inc. increased its quarterly to 10 cents from 7 cents a share.

Subjectivity – Examples (2/2) Subjective sentence about a speech event: “The cost of health care is eroding our standard of living and sapping industrial strength, ” complains Walter Maher, a Chrysler health-andbenefits specialist. Objective sentence about a speech event: Northwest Airlines settled the remaining lawsuits filed on behalf of 156 people killed in a 1987 crash, but claims against the jetliner’s maker are being pursued, a federal judge said.

Aspects of subjectivity expressions (1/5) n There are expressions subjective in all contexts: n ! n But many are subjective depending on the context: n sapping, eroding n A potential subjective element is a linguistic element that may be used to express subjectivity. n A subjective element is an instance of a potential subjective element, in a particular context, that is indeed subjective in that context. (Wiebe 1994)

Aspects of subjectivity expressions (2/5) n There are different types of subjectivity, and the work focuses on three: positive evaluation (e. g. fascinating) n negative evaluation (e. g. terrible) n speculation (e. g. probably) n

Aspects of subjectivity expressions (3/5) n A subjective element expresses the subjectivity of a source. n Source = writer or someone mentioned in text. At several different layers, it’s a fascinating tale. Source = writer. “The cost of health care is eroding our standard of living and sapping industrial strength, ” complains Walter Maher, a Chrysler health-andbenefits specialist. Source = Maher.

Aspects of subjectivity expressions (4/5) n A subjective element also has a target. n Target = what the subjectivity is all about or directed toward. At several different layers, it’s a fascinating tale. Target = a tale. “The cost of health care is eroding our standard of living and sapping industrial strength, ” complains Walter Maher, a Chrysler health-andbenefits specialist. Target = the cost of health care.

Aspects of subjectivity expressions (5/5) n The former examples have object-centric subjectivity. Other examples: I love this project. The software is horrible. n Subjectivity may also be addressee-oriented (directed towards the listener and reader) You are an idiot.

Experiments n Corpus: n 1, 001 sentences of the Wall Street Journal Treebank Corpus (Marcu et al. 1993) n manually annotated with subjectivity classifications n in addition: subjective elements & strength of elements (on a scale of 1 to 3)

Improving Adjective Features Using Distributional Similarity (1/2) n Intuition: words correlated with many of the same things in text are more similar. n Challenging test: 10 -fold cross validation 1/10 training n 9/10 testing n For each training set i n Extract all adjectives from subjective elements of strength 3 n For each adjective n n Identify top 20 entries in a similarity thesaurus (Lin 1994) n These are the seed sets for fold i. n Evaluate these seed sets on the remaining 9/10 of the corpus.

Improving Adjective Features Using Distributional Similarity (2/2) n Baseline: the precision of a simple adjective feature (= the conditional probability that a sentence is subjective, given that at least one adjective appears). n Average precision: 55. 8% n Above-mentioned process: n Average precision: 61. 2% è Increase: 5. 4% n Repeat experiment with similarities from Word. Net. n Average precision: 62. 0% è Slight increase, but lower coverage.

Refinements with Polarity and Gradability (1/2) n Polarity: presented in previous paper. n Gradability: the semantic property that enables a word to participate in comparative constructs and n to accept modifying expressions that act as intensifiers and diminishers. n Gradable adjectives express properties in varying degrees of strength, relative to a norm (explicitly or implicitly supplied by the modified noun) n a small planet – a large house n (Hatzivassiloglou) list of 73 adverbs and NPs that are frequently used as grading modifiers. a little, exceedingly, somewhat, very

Refinements with Polarity and Gradability (2/2) n The work uses samples of adjectives identified as: having positive polarity n having negative polarity n being gradable n n Samples were determined using a new corpus from the Wall Street Journal.

Results and Discussion n Experiments for automatic/manual identification of polarity +, -, + -, and gradable adjectives. n Promising results: n In all cases, the average improvement over the baseline of the intersection btw. seed sets and gradability/polarity sets is at least 9%. n The gradability/polarity sets and the seed sets are more precise together than alone. n Excellent individual results: gradability/automatic and polarity -/automatic sets intersected with the seed sets. n Future work: n Some of the data that is currently part of the test set could be used to filter the sets (3/10 of data used for training and 1/3 of training data used for seeding, and 2/3 for filtering)