基于混合神经网络的个性化自然语言情感识别系统  

Personalized Natural Language Emotion Recognition System Based on Hybrid Neural Network

作　　者：韦灵 卢光云 唐爱龙 WEI Ling;LU Guangyun;TANG Ailong(College of Information Science and Engineering,Liuzhou Institute of Technology,Liuzhou 545616,China)

机构地区：[1]柳州工学院信息科学与工程学院,柳州545616

出　　处：《自动化与仪表》2024年第9期26-28,65,共4页Automation & Instrumentation

基　　金：2021年度柳州工学院科学基金项目(2021KXJJO8);2021年中国高校产学研创新基金——新一代信息技术创新项目(2021ITA02008)。

摘　　要：为在面临输入自然语言信息中存在微小变化或噪声干扰时,提升系统的鲁棒性与泛化能力,设计基于混合神经网络的个性化自然语言情感识别系统。以现场可编程门阵列为核心,设计自然语言采集模块,用于采集个性化自然语言信息;通过预处理模块,处理采集的个性化自然语言信息中的语音信息,得到对数梅尔谱图;情感识别模块利用动态卷积神经网络和长短时记忆网络,组建混合神经网络,其中,通过动态卷积神经网络,在对数梅尔谱图内提取语音信息特征,采用长短时记忆神经网络,在自然语言文本信息内提取文本信息特征,通过全连接神经网络,融合语音与文本信息特征,输出个性化情感识别结果。实验证明,该系统可有效采集个性化自然语言信息,并提取语音信息特征,完成个性化自然语言情感识别;在噪声干扰下,该系统的情感识别精度较高。In order to improve the robustness and generalization ability of the system when there are slight changes or noise interference in the input natural language information,a personalized natural language emotion recognition system based on hybrid neural network is designed.Taking field programmable gate array as the core,a natural language acquisition module is designed to collect personalized natural language information.Processing the voice information in the collected personalized natural language information through a preprocessing module to obtain a logarithmic Mel spectrogram.Emotion recognition module uses dynamic convolution neural network and long-term and short-term memory network to build a hybrid neural network,in which,through dynamic convolution neural network,voice information features are extracted from logarithmic Mel spectrogram,and through long-term and short-term memory network,text information features are extracted from natural language text information,and through fully connected neural network,voice and text information features are fused to output personalized emotion recognition results.Experiments show that the system can effectively collect personalized natural language information,extract phonetic information features and complete personalized natural language emotion recognition.Under the interference of noise,the emotion recognition accuracy of the system is high.

关 键 词：混合神经网络 个性化 自然语言 情感识别 动态卷积 长短时记忆网络 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]