Graph-Based Multi-Tweet Summarization using Social Signals

Proceedings of COLING 2012: Technical Papers, pages 1699–1714, COLING 2012, Mumbai, December 2012. Graph-based Multi-tweet Summarization Using Social Signals LIU X iaoHua1,2 LI Y iTong3 W EI FuRu2 ZHOU Ming2 (1) School of Computer Science and Technology, Harbin Institute of Technology, Harbin, 150001, China (2) Microsoft Research Asia, Beijing, 100190, China (3) School of Electronic and Information Engineering, Beihang University, Beijing, 100191, China {xiaoliu,v-yitli,fuwei,mingzhou}@microsoft.com Abstract We study the multi-tweet summarization task, which aims to find representative tweets from a given set of tweets. Multi-tweet summarization allows people to quickly grasp the essential meaning of a large number of tweets. It can also be used as a pre-processing component for information extraction tasks on tweets. The challenge of this task lies in computing a tweet’s salience score with little information in a single tweet. We propose a graph-based multi-tweet summarization system that incorporates social network features, which make up for the information shortage in a tweet. Another distinguished feature of our system is that tweet readability and user diversity are considered. We evaluate our system on a manually annotated dataset, and show that our system outperforms the state- of-the-art baseline. We further evaluate our method in a real scenario of summarization of Twitter search results and demonstrate its effectiveness. Title and Abstract in another language (Chinese) 基于图模型和社会关系特征的推特消息摘要方法 本文考察多推特消息摘要任务。其目的是帮助用户快速了解大量推特消息的基本意思，或 在信息抽取前对推特消息做预处理。 该任务的主要挑战是：单条推特消息往往不能提供足 够的信息来计算它的显著度。本文提出基于图模型的方法，考虑社会关系网相关的特征、 可读性及用户的多样性来克服单条推特消息的不足。 在人工标注的数据集以及推特搜索 上，该方法均展示了其有效性。 Keywords: Graph, Summarization, Social signals, Tweets. Keywords in Chinese: 图模型，摘要，社会关系网特征，推特消息 . 1699 1 Introduction Twitter1 is a microblogging and social networking service with a huge number of users and is continuously growing at a tremendous rate. Tweets, short messages shared through Twitter with less than 140 characters, have become an important repository of real-time information. However, it is often inefficient for people to consume a large number of tweets, owing to the noise prone nature of tweets2. We propose the task of multi-tweet summarization to overcome this obstacle: eliminating redundancy and noise while keeping the essential information for a given set of tweets. As an illustrative example, consider the following tweets, returned from Twitter search related to the query “obama": • “@morrowchris: Christmas card from the Obama’s :) http://pic.twitter.com/F3VU52io" thts special. • Christmas card from the Obama’s :) http://pic.twitter.com/7xqBQdBV • RT @liberalease: RT if you like President Obama better than ALL OF GOP COM- BINED. / That was easy :) • Obama did it! The war is OFFICIALLY OVER! #WelcomeHomeTroops! :) • obama really brought the troops home :) The first and the second tweet are almost the same; the third tweet is a private conversation, thus not meaningful for the general audience; and the last two talk about similar things. After summarization, the expected outputs are as follows, which keep the main information with noises removed. • Christmas card from the Obama’s :) http://pic.twitter.com/7xqBQdBV • obama really brought the troops home :) We advocate that multi-tweet summarization can play a critical role for many tweet related studies, which have attracted increasing research interests in recent days, such as named entity recognition (NER) and sentiment analysis (SA) …

Content truncated to 500 words