The Role of Semantic Roles in Disambiguating

Verb Senses

The Role of Semantic Roles in Disambiguating

Verb Senses

Hoa Trang Dang and Martha Palmer

2005. Proceedings of the 43rd Annual Meeting of the ACL, pages 42–49.

Hoa Trang Dang and Martha Palmer

2005. Proceedings of the 43rd Annual Meeting of the ACL, pages 42–49.

Verbs are syntactically complex, and their syntax is thought to be determined by their underlying semantics (Grimshaw, 1990; Levin, 1993).

Disambiguation of verb senses can be further improved with better extraction of semantic roles.

WordNet: word senses.

PropBank: semantic role labels.

Mallet: for learning maximum entropy models with Gaussian priors.

Senseval-2: the system was tested on thousands of the test instances of the 29 verbs from the English lexical sample task for Senseval-2.

Basic Automatic System

Topical features

Local features

1.Collocation features

2.Syntactic features

3.Semantic features

Basic Automatic System

Topical Features

Topical features for a verb in a sentence look for the presence of keywords occurring anywhere in the sentence and any surrounding sentences provided as context.

The set of keywords is specific to each verb lemma to be disambiguated.

Local Features

Collocational features:

unigrams: words w-2, w-1, w0, w+1, w+2

part-of-speech p-2, p-1, p0, p+1, p+2

bigrams: w-2w-1,w-1w+1,w+1w+2; p-2p-1,p-1p+1 p+1p+2

trigrams: w-3w-2w-1,w-2w-1w+1,w-1w+1w+2,w+1w+2w+3;

p-3p-2p-1, p-2p-1p+1, p-1p+1p+2, p+1p+2p+3

Local Features Syntactic features

Is the sentence passive?

Is there a subject, direct object, indirect object , or clausal complement?

What is the word (if any) that is the particle or head of the subject, direct object, or indirect object?

If there is a PP complement, what is the preposition, and what is the object of the preposition?

Local Features

Semantic features:

What is the Named Entity tag (PERSON, ORGANIZATION, LOCATION, UNKNOWN) for each proper noun in the syntactic positions above?

What are the possible WordNet synsets and hypernyms for each noun in the syntactic positions above?

EvaluationFeature Accuracy

co 0.571

co+syn 0.598

co+syn+sem 0.625

Accuracy of system on Senseval-2 verbs using topical features and different subsets of local features.

co=collocational syn=syntactic sem=semantic

This system: 62.5% accuracy.Lee and Ng, 2002: 61.1% accuracy.

EvaluationFeature Accuracyco+syn 0.598

co+syn+ne 0.597co+syn+wn 0.623

co+syn+ne+wn

0.625Accuracy of system on Senseval-2 verbs, using topical features and different subsets of semantic class features.

ne=named entity tags wn=WordNet classes

PropBank

PropBank is a corpus in which verbs are annotated with semantic tags, including coarse-grained sense distinctions and predicate-argument structures.

Example: [ ARG0 Mr. Bush] has [rel called] [ ARG1-for for an agreement by next September at the latest]

Frameset TaggingFeature Accuracybaseline 0.760

co 0.853synsem 0.859

co+synsem 0.883pb 0.901

co+pb 0.908co+synsem+pb 0.907

Accuracy of system on frameset-tagging task for verbs with more than one frameset, using different types of local features. (pb=PropBank role features.)

*The most frequent frameset gives a baseline accuracy of 76.0%.

WordNet Sense-taggingFeature Accuracy

co 0.628synsem 0.638

co+synsem 0.666pb 0.656

co+pb 0.681co+synsem+pb 0.694

Accuracy of system on WordNet sense-tagging for instances in both Senseval-2 and PropBank, using different types of local features.

*PropBank ARGM features are included.

Frameset tags for WordNet sense-tagging

Feature Accuracy

orig 0.564

orig*fset 0.587(orig+pb)*fs

et 0.628

Accuracy of system on WordNet sensetagging of 20 Senseval-2 verbs with more than one frameset.

orig=original local features.

Conclusion

Disambiguation for verbs can be improved through more accurate extraction of features representing information such as that contained in the framesets and predicate argument structures annotated in PropBank.