社交媒体多模态品牌追踪与事件检测

项目名称： 社交媒体多模态品牌追踪与事件检测

项目编号： No.61472059

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 李豪杰

作者单位： 大连理工大学

项目金额： 84万元

中文摘要： 社交媒体特别是微博作为用户信息共享平台近年来得到了空前发展，成为网络舆情的重要载体。如何高效地从海量微博数据中获取用户对产品的观点和态度成为企业的迫切需求和面临的挑战性问题。现有研究大都集中于使用单模态即文本对品牌进行跟踪分析，而忽视了微博中内含的图像模态信息，因而获取信息的覆盖性和准确性有待提高。本课题致力于研究基于多模态的品牌追踪和事件检测技术，从图像检索、聚类及多模态融合角度对其中的关键问题和技术展开研究。具体内容包括：结合文本和图像检索的多源多模态微博信息采集及重排序技术；用于微博滤噪的高效准确的商标检测方法；基于文本和图像联合谱聚类进行品牌感兴趣事件检测，并对事件进行多模态呈现；深入分析并构建微博热点事件特征模型，进而研究有效的热点事件预警方法。本课题将有力推动微博等社交媒体舆情监控的理论发展和应用，为企业决策和网络传播环境监管提供核心算法和技术。

中文关键词： 计算机视觉；图像检索；图像分析

英文摘要： As a new information sharing plateform, social media, especially microblogs has got explosive growth in recent years and has become an important carrier for public opinions. How to effectively mine users' opinions and attitudes towards their products from such massive microblogs data is becoming an urgent and challenging issue for companies and manufacturers. Current research on this problem usually focuses on the use of text information for brand tracking and analysis, while neglecting the visual information contained in microblogs, which makes the coverage and precision of the obtained knowledge is not satisfactory.In this project,we propose a multi-modality framework for brand tracking and event detection in social media and address this issue by exploiting the techniques of image matching,clustering and multi-modality fusion. More specifically, we first introduce a multi-faceted microblogs data gathering strategy to collect high-coverage and diverse data by integarting textual and visual retrieval. Then we filter out the unrelated microblogs with a proposed efficient and effective logo detection algorithm.To detect the brand events of interests from the multi-modal microblog data, we propose to use a hyper-graph approach to clustering the textual and visual data jointly to sufficiently utilize the underlying correlations between the two kinds of data.We also develop a multi-modal summarization method for the mined events.For hot event alarming, we will first thouroughly investigate the various features of hot event and then use them to study semi-supervised alarming algorithms. Our research will surely make an important advance in social media monitoring and will display an important role in business decision-making and network information dissemination environment monitoring.

英文关键词： computer vision;image retrieval;image analysis