Research On Feature-based Semantic Role Labeling

During the last decade, there has been an increasing interest in shallow semantic parsing of natural language. In particular, given a sentence and a predicate, semantic role labeling (SRL) identifies its semantic arguments and classifies their semantic roles. This paper systematically studies feature-based SRL with focus on local modeling and joint modeling. Evaluation on the PropBank benchmark corpus shows that our system works well and achieves higher performance than comparable ones.Firstly, a local model-based baseline system was constructed with brief investigation on the major components, including pruning, identification, classification and post-processing. Moreover, the impact of diffferent basic features was evaluated. Secondly, proper feature engineering was systematically explored for SRL. With a greedy strategy, the commonly-used forty-nine features were fully investigated with gold parsing and automatic parsing, respectively. Evaluation shows that the predicate,head word and related lexical features play an important role in SRL. It also shows that more features don't necessarily mean higher performance and that proper feature engineering is critical for high performance in SRL.Thirdly, a head-driven pruning algorithm was proposed improve over Xue's prunning algorithm by keeping the children of the predicate's siblings as the candidate argument and discarding those with the same head word with its parental node. Evalation shows that our pruning algorithm outperforms Xue's pruning algorithm.Finally, a joint model was used to consider the soft constraints in SRL by re-ranking the N-best non-overlapping joint assignments of semantic roles in a parse tree according to the local model, using a dynamic programming algorithm. Evaluation shows that the joint model is promising and improves the performance.