【摘要】 提出了一个基于图结构的依存句法分析模型。使用双向长短期记忆神经网络模型自动提取富文本信息,然后添加约束性条件来保障句法分析的正确性,并结合隐马尔科夫模型得到句法分析树,使算法避免了人工选择特征,同时考虑到了长距离依存问题。另外就语言模型、句子的长度等因素对算法效果的影响做了分析。以中文汉宾树库数据集作为实验语料,与其他3种常见句法分析算法进行对比,依存关系准确率、根节点准确率和头节点占所有词百分比在短句和长句分析上均有所提高,验证了新算法的可行性和有效性。更多还原

【Abstract】 A graph-based dependency parser is proposed.Rich textual information and effective features are obtained by bidirectional long short term memory(LSTM)model,combined with constraints and the hidden Markov model(HMM)model to obtain a parsing tree,avoiding the feature of artificial selection,and considering the problem of long distance dependence.In addition,language model and sentences length have been analyzed.Compared with three other common syntactic parsers in Chinese treebank,the evaluation indicators(e.g.dependency accuracy,root node accuracy and chief match)values increase in short sentences and long sentences.Experimental results show that this model is effective and feasible.更多还原