基于候选主题词与话题分类的人物行为研究-《计算机技术与发展》

文章信息/Info

Title:: Research on Character Behavior Based on Candidate Keywords and Topics Classification

作者:: 刘晓芳¹ ; 欧荣安²; 罗欢³ ; 刘芳婷⁴; 张辉极¹ ; 韩冰²; 赵建强¹; 5*; 1. 厦门市美亚柏科信息股份有限公司,福建厦门 361008;
2. 广州市刑事科学技术研究所,广东广州 510030;
3. 福州大学,福建福州 350108;
4. 厦门市人民检察院检察技术信息部,福建厦门 361008;
5. 西安电子科技大学,陕西西安 710071

Author(s):: LIU Xiao-fang1 ; OU Rong-an2 ; LUO Huan3 ; LIU Fang-ting4 ; ZHANG Hui-ji1 ; HAN Bing2 ; ZHAO Jian-qiang1; 5*; 1. Xiamen Meiya Pico Information Co. ,Ltd. ,Xiamen 361008,China;
2. Guangzhou Institute of Criminal Science and Technology,Guangzhou 510030,China;
3. Fuzhou University,Fuzhou 350108,China;
4. Procuratorial Technology Department of the People’s Procuratorate of Xiamen,Xiamen 361008,China;
5. Xidian University,Xi’an 710071,China

Keywords:: group chat topic; candidate keywords; topic classification; personalities relationship; interaction intensity; group memberactivity; social label

摘要:: 如何从海量聊天数据获取聊天主题和聊天人物行为是案件智能化分析的热点问题之一。传统词嵌入方法, 将文本中的所有词汇映射到向量空间,存在词汇特征冗余的问题。为了缓解这一问题,该文提出一种基于候选主题词的话题分类算法—CTW ( candidate topic words) 。该算法使用 LDA 主题模型抽取聊天文本中的关键词,使用预训练词向量得到显著的语义特征,同时为增强特征,将字符特征与获取的词汇特征进行融合。传统方法同时还存在只关注话题无法更精确地刻画人物行为的问题。针对该问题,该文提出了同时获取聊天话题和人物行为的方案:针对已归类的话题,该方案使用群成员互动强度、群成员活跃度作为人物行为网络中的权值,构建话题参与人的行为网络图;最后通过成员在群中的备注给人物赋予不同的社会标签,以丰富人物行为。实验表明,提出的话题分类算法,在实际搜集的数据集上比基线模型拥有更佳的性能,在获取群聊话题的同时得到了更丰富的人物行为描述。

Abstract:: For the high incidence of new types of cyber crime,criminals use various chat groups to induce victims and commit crimes. The effective description and portrayal of the chat content and the character relationships will provide strong support for the case investigation.Generally traditional methods only focus on one aspect of chat topics or character behavior,which is unable to extract behaviors and chattopics at the same time quickly and effectively in forensics. Therefore,we propose a method- CTW ( candidate topic words) to portray character relationships based on candidate keywords and group chat topics. Firstly,LDA model is used to extract key words,using the embedding of? char and extracted key words for Text CNN to categorize topics. For categorized topics,we treat interaction intensity and group member activity as the weight value on the directed edges,not only create a network diagram of the character behavior of the participants in the? ? group chat and get the members’ remarks in the group as social relationship labels which are put on the edges of the relationship diagrams. Experiments show that the method proposed shows good results on actual data sets,which can effectively portray character relationships based on the topics of group chats.