Towards Robust Semantic Role Labeling

Most research on semantic role labeling (SRL) has been focused on training and evaluating on the same corpus in order to develop the technology. This strategy, while appropriate for initiating research, can lead to over-training to the particular corpus. The work presented in this paper focuses on analyzing the robustness of an SRL system when trained on one genre of data and used to label a different genre. Our state-of-the-art semantic role labeling system, while performing well on WSJ test data, shows significant performance degradation when applied to data from the Brown corpus. We present a series of experiments designed to investigate the source of this lack of portability. These experiments are based on comparisons of performance using PropBanked WSJ data and PropBanked Brown corpus data. Our results indicate that while syntactic parses and argument identification port relatively well to a new genre, argument classification does not. Our analysis of the reasons for this is presented and generally point to the nature of the more lexical/semantic features dominating the classification task and general structural features dominating the argument identification task.