APP下载

一种循环神经网络的词义消歧方法

2020-05-21张春祥周雪松高雪瑶

哈尔滨理工大学学报 2020年1期

张春祥 周雪松 高雪瑶

摘 要:词义消歧是自然语言处理领域中的一个重要研究问题。针对汉语一词多义现象,结合上下文语境,采用循环神经网络(Recurrent Netural Networks,RNN)来确定歧义词汇的真实含义。以目标歧义词汇为中心,提取其左右4个邻接的词汇单元。从中抽取词形、词性和语义类作为消歧特征。以消歧特征为基础,结合循环神经网络来构建词义消歧分类器。利用SemEval-2007: Task#5的训练语料和哈尔滨工业大学语义标注语料来优化RNN的参数。使用SemEval-2007: Task#5的测试语料对词义消歧分类器进行测试。实验结果表明:所提出的方法能够提高词义消歧的准确率。

关键词:词义消歧;自然语言处理;循环神经网络;消歧特征

DOI:10.15938/j.jhust.2020.01.012

中图分类号: TP391.2

文献标志码: A

文章编号: 1007-2683(2020)01-0080-06

Abstract:Word sense disambiguation is an important research problem in natural language processing field. For the phenomenon that a Chinese word has many senses, recurrent neural network(RNN) is used to determine true meaning of ambiguous word with its context. Target ambiguous word is viewed as center and its four adjacent word units are extracted. Word, part-of-speech and semantic categories are extracted as disambiguation features. Based on disambiguation features, recurrent neural network is used to construct word sense disambiguation classifier. Training corpus in SemEval-2007: Task#5 and semantic annotation corpus in Harbin Institute of Technology are used to optimize parameters of RNN. Test corpus in SemEval-2007: Task#5 is applied to test word sense disambiguation classifier. Experimental results show that the proposed method can improve accuracy of word sense disambiguation.

Keywords:word sense disambiguation; natural language processing; recurrent neural network; disambiguation features

0 引 言

词义消歧是自然语言处理领域中的基础性研究问题。同时,也是机器翻译和机器理解中的核心步骤。鹿文鹏根据依存句法分析结果来提取歧义词汇的依存约束集合。从WordNet中找出歧义词汇各个词义的代表词。根据词义代表词在依存约束集合中的依存适配度来选择正确的词义[1]。鹿文鹏等[2]利用文本领域关联词和句子上下文词来构建消歧图。利用领域知识来调整消歧图。同时,使用改进的图评分标准对消歧图中的各个词义结点的重要程度进行打分。杨陟卓等[3]运用语言模型来优化有监督消歧模型。充分利用有监督消歧模型和语言模型来确定歧义词汇的真实含义。在训练语料不足的情况下,这种方法可以有效地改善词义消歧的效果。……

登录APP查看全文