Beyond Selectional Preference: Metaphor Recognition with Semantic Relation Patterns

This paper analyzes the limitations of selectional-preference based metaphor recognition, and proposes a new model for metaphor recognition, using Chinese subject-predicate construction as illustration. After showing with experiments that selectional-preference based metaphor recognition has difficulty in recognizing conventional metaphors and literal expressions with low frequency, the paper presents a metaphor recognition model which is based on Semantic Relation Pattern, a distribution pattern integrating six types of semantic relations between a subject head and other subject heads within a subject-predicate cluster, and employs a SVM classifier for metaphor recognition. Contrastive Experiments show that the proposed model achieves an F1 of 89% in metaphor recognition, about 37% higher than the selectional-preference based model. Further analysis shows that the proposed 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 proposed model possesses the ability to generalize to unknown predicate heads. Theoretically, the semantic-relation-pattern model can also be applied for metaphor recognition in other endocentric constructions such as verb-objects and adjective-nouns.