90 likes | 172 Views
A classification-based, probabilistic model integrating constituent-based methodology by Daniel Gildea & Martha Palmer in ACL 2002. Current implementation involves Multi-class SVM using libsvm package from CSIE, NTU with a radial basis kernel at degree 3. Feature sets include verb, phrasal type, path, voice, position, and sub_cat frame. Challenges include model selection and the trade-off between lexical richness and data sparseness in clustering slot-verb role nouns. Verb clustering utilizes a revised model of Gildea 2002. Future work involves automatic constituent boundary detection, adding more features, and exploring other machine learning models.
E N D
Automatic Semantic Role Tagger Szu-Ting Yi
Overview • Classification-based • Probabilistic model • Constituent-based
Current Implementation • Daniel Gildea & Martha Palmer (ACL 2002) • Multi-class SVM (using libsvm package from CSIE, NTU) • Kernel: radial basis • Degree: 3
Feature Sets • Verb – the predicate • Phrasal type • Path • Voice • Position • Sub_cat frame
Problems • Model selection • Trade-off between more lexical information & data sparseness
cluster slot verb role noun Verb Clustering • (revised model of Gildea 2002)
Conclusion & Future work • Automatic constituent boundary detection • Incorporating more features • Other machine learning models