A General Multi-Decoder Framework for Frame Semantic Role Labeling

Given the high proportion of errors in argument identification in the task of Frame Semantic Role Labeling (FSRL), we utilize multiple argument identification decoders, including tagging, boundary prediction and argument enumeration to construct a general FSRL framework. The tagging decoder predicts a tag sequence for a given sentence based on BIO scheme. The boundary prediction decoder provides boundary information of arguments within sentences, thereby enhancing the tagging decoder's attention on boundary positions of arguments. The argument enumeration decoder produces all possible candidate arguments to constrain and guide the predictions made by the tagging decoder. We conduct experiments on two datasets and compare our model with the previous state-of-the-art model. The F1-score is improved up to 1.35% and 0.24% on two datasets, respectively. Furthermore, the results of ablation experiment further demonstrates the effectiveness of the model.