Predicate-attention neural model for Chinese semantic role labeling

Abstract

Semantic role labeling functions to convey the meaning of a sentence through forming a predicate-argument structure directed at the specific predicate. In recent years, end-to-end semantic role labeling methods associated with the deep neural network have received significant attention in the field of computational linguistics. Moreover, end-to-end semantic role labeling methods have demonstrated a beneficial capacity to reduce the incompleteness caused by handcrafted features, which is an observed short-coming of traditional Chinese role labeling methods. However, the critical focus of sentences are frequently lost as a result of existing semantic role labeling structures attributing equal importance to every single word, instead of the overall concept denoted by particular terms. Hence, the performance and function ability of deep neural network models is reduced. In this paper, we introduce a specific attention mechanism based on the established predicate. This mechanism would automatically calculate the weighted contributions of each word, and the corresponding Part-of-Speech, in order to accurately represent the general fundamental ideas of the sentence. In addition, we extended the Bidirectional LSTM using two different semantic role constraint methods, to effectively utilize the dependency and constraint relationships among different semantic role tags, hereby further improving the performance of the whole neural Chinese semantic role labeling model. Experimental results demonstrate the efficacy of our proposed model through providing a baseline that allows for meaningful comparisons, inferring that both weighted contributions of the predicate, and semantic role constraints can help significantly refine the overall model function.