Semantic Role Labeling Via Syntax-Free Methods

How to do semantic parsing on a large amount of unstructured texts,which is used to extract information and discover knowledge,is increasingly the focus of research in recent years.Semantic role labeling is an important way of shallow semantic parsing,and it is also a basic technology used for text analysis and information extraction.Traditional semantic role labeling methods have limitations that they highly rely on artificial features and dependency path.Thus,this dissertation studies semantic role labeling as follows:After analyzing the drawbacks of traditional methods,a feature-minimized and syntax-free model is proposed.A semantic role labeling model is constructed based on bidirectional long short-term memory,word2vec and part of speech embedding.The senses of words are well expressed by word vector and part of speech.Long short-term memory has remarkable ability to model the sentence as well as sematic role features.In order to solve the problem that long short-term memory has difficulty on parallelizing state computations,a method based on iterated dilated convolutional neural network is proposed.This method helps convolutional neural network to learn long term dependency and it also accelerates training process.Attention mechanism is applied to proposed model.It would help the model to learn long term dependency and structural features between words.On public dataset,a series of experiments that compare the proposed models with traditional models and baseline models are conducted.The experimental results show that the proposed methods outperform other contrast models,which demonstrated the effectiveness of proposed methods.