基于终身朴素贝叶斯方法的情感分类  

Sentiment Classification Based on Lifelong Naive Bayesian

作　　者：王松 买日旦·吾守尔[1] 古兰拜尔·吐尔洪 段淑敏 WANG Song;MAIRIDAN Wushouer;GULANBAIER Tuerhong;DUAN Shumin(School of Information Science and Technology,Xinjiang University,Urumqi,Xinjiang 830046,China)

机构地区：[1]新疆大学信息科学与工程学院,新疆乌鲁木齐830046

出　　处：《中文信息学报》2023年第9期150-160,共11页Journal of Chinese Information Processing

基　　金：自治区自然科学基金(2018D01C075);教育厅高校科研青年项目(61021800032,61021211418);自治区高层次创新人才项目(100400016,042419006)。

摘　　要：面对快速增长和变化的网络信息,现有情感分类模型受灾难遗忘的影响,在训练完成后难以累积和迁移知识。受到人类持续学习过程的启发,尝试用终身学习解决这一问题。该文讨论终身学习概念并为终身学习范式构建多任务情感分类数据集;将终身朴素贝叶斯方法用于中文情感分类任务:把不同种类的商品评论视为不同的情感分类任务,LNB算法依次学习并为每一个任务构建分类器。利用领域注意力机制改进LNB,提出该文方法LNB-DA。实验表明该文方法具备知识累积与迁移的能力,负类F 1值优于对比方法0.15%~23.41%。The existing sentiment classification approach is defected in incapability of knowledge accumulation and transferrin.Lifelong learning(LL),a new learning paradigm,tries to address this problem by simulating the human behavior of continual learning.This paper discusses general knowledge of lifelong learning and builds a new Chinese sentiment classification dataset for LL setting.We apply lifelong naive Bayesian(LNB)to Chinese sentiment classification tasks,where the pre-defined classification tasks are learned sequentially to build a classifier for each category of product reviews,and the LNB is enhanced by the domain attention mechanism.Experimental results show that our approach Lifelong Na ve Bayesian with Domain Attention(LNB-DA)outperforms baselines by 0.15%-23.41%according to F 1-negative score.

关 键 词：终身学习 持续学习 朴素贝叶斯 情感分类 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]