Adaptive Convolution for Semantic Role Labeling

Semantic role labeling (SRL) aims at elaborating the meaning of a sentence by forming predicate-argument structure. Recent researches depicted that effective use syntax can improve SRL performance. However, is complicated linguistic clue and hard to be effectively applied in downstream task like SRL. This work encodes using adaptive convolution which endows strong flexibility existing convolutional networks. The CNNs may help encoding structure for SRL, but it still has shortcomings. Contrary traditional networks same filters different inputs, uses adaptively generated conditioned on syntactically-informed inputs. We achieve this with integration filter generation network generates input specific filters. helps model focus important syntactic features present inside input, thus enlarging gap between syntax-aware syntax-agnostic systems. further study hashing technique compress size terms trainable parameters. Experiments CoNLL-2009 dataset confirm proposed substantially outperforms most previous systems both English Chinese languages.