Xuri Tang

Automatic Metaphor Recognition Based on Semantic Relation Patterns

Focusing on Chinese subject-predicate constructions, this paper analyzes the limitations of Selectional-Preference based metaphor recognition and proposes a new metaphor recognition model which is based on Semantic Relation Patterns. The model constructs Semantic Relation Pattern by integrating six types of semantic relations between a subject head and other subject heads in a subject-predicate cluster which share the same predicate head, and then employs a SVM classifier for metaphor recognition. Experiments show that the model outperforms the Selectional-Preference based metaphor recognition model to a great extent, achieving an F-1 of 89% in metaphor recognition, about 37% higher than Selectional-Preference based model. Analysis shows that the model is able to account for lexicalized metaphors, truth-condition literality and other types of literality and metaphor failed in Selectional-Preference based models. More importantly, the model can be generalized to unknown predicate heads. Theoretically, the semantic-relation-pattern model can also be applied in all endocentric constructions such as verb-objects and adjective-nouns.

Computing Word Similarity on Large-Scale Corpus

This paper proposes a novel approach for word similarity computation based on word sense vectors. The word sense vector is built using HIT-IR Tongyici Cilin (Extended) for concept generalization and is further modified by the use of relative and absolute frequency filters. Experiments show that the approach not only overcomes the problem of similarity computation of unseen words but also yields a result closer to human judgment when compared to word similarity computation approaches based on dictionaries.

Polyphonic Word Disambiguation with Machine Learning Approaches

Five different classification models, namely RFR_SUM, CRFs, Maximum Entropy, SVM and Semantic Similarity Model, are employed for polyphonic disambiguation. Based on observation of the experiment outcome of these models, an additional ensemble method based on majority voting is proposed. The ensemble method obtains an average precision of 96.78%, which is much better than the results obtained in previous literatures.