Analysis of joint inference strategies for the semantic role labeling of Spanish and Catalan
This paper analyzes two joint inference approaches for semantic role labeling: re-ranking of candidate semantic frames generated by one local model and combination of two distinct models at argument-level using meta learning. We perform an empirical analysis on two recently released corpora of annotated semantic roles in Spanish and Catalan. This work yields several novel conclusions: (a) the proposed joint inference strategies yield good results even under adverse conditions: small training corpora, only two individual models available for combination, minimal output available from the individual models; (b) stacking of the two joint inference approaches is successful, which indicates that the two inference models provide complementary benefits. Our results are currently the best for the identification of semantic role for Spanish and Catalan.