基于Bi-LSTM的多层面隐喻识别方法  

Multi-level metaphor detection method based on Bi-LSTM

作　　者：朱嘉莹 王荣波[1] 黄孝喜[1] 谌志群 ZHU Jiaying;WANG Rongbo;HUANG Xiaoxi;CHEN Zhiqun(Institute of Cognitive and Intelligent Computing,Hangzhou Dianzi University,Hangzhou 310018,China)

机构地区：[1]杭州电子科技大学认知与智能计算研究所,浙江杭州310018

出　　处：《大连理工大学学报》2020年第2期209-215,共7页Journal of Dalian University of Technology

基　　金：国家自然科学基金青年基金资助项目(61202281);教育部人文社科规划青年基金资助项目(12YJCZH201);国家社会科学基金资助项目(18ZDA290);教育部人文社科项目规划基金资助项目(18YJA740016).

摘　　要：以双向长短期记忆网络(Bi-LSTM)为核心,结合多层卷积神经网络以及单向长短期记忆网络构建了多层面隐喻识别模型.基于多特征协同作用的思想,利用依存关系特征、语义特征、词性特征等多特征融合输入方法,丰富了模型的学习信息.为降低信息干扰,利用基于统计学的规范化文本输入方法提升模型识别效果.在英文语料词层面和句层面实验中,各个特征均表现出明显的正向作用.裁剪和填充处理及多特征协调作用在英文语料词层面研究中使F1值分别提升2.5%和5.1%,在句层面研究中F1值分别提升3.1%和1.9%.在中文语料句层面实验中,最优效果的F1值可达88.8%.Taking Bi-directional long short-term memory network(Bi-LSTM) as the core, combining with multi convolutional neural network layers and unidirectional long short-term memory network, a multi-level metaphor recognition model is built. Based on the idea of multi-feature synergism, the learning information of the model is enriched by using the methods of inputting multi-feature such as dependency feature, semantic feature and part-of-speech feature in parallel. In order to reduce information interference, a standardized text input method based on statistics is used to improve the recognition effect of the model. In the experiments on word level and sentence level of English corpus, each feature has obvious positive effect. In the word level research of English corpus, the F1-scores of croping and filling treatment, multi-feature synergism increase by 2.5% and 5.1% respectively, while in the sentence level research, the F1-scores increase by 3.1% and 1.9% respectively. In the sentence level experiment of Chinese corpus, the F1-score of the optimal effect can reach 88.8%.

关 键 词：自然语言理解 隐喻识别 CNN Bi-LSTM 依存关系 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]