Research On Social Media Summarization

Development of social media platforms has enabled convenient information sharing.More and more users post and exchange ideas and opinions on online social media sites,such as Twitter and Facebook.The amount of data that Internet users transmit via the social media has reached an astronomical number every minute.The social media has brought us opportunities in knowledge management,but the high volume,complicated structure and redundancy in social textual data also pose challenges.Automated text summarization system is one of the important technology to alleviate the information overload problem.Traditional text summarization methods fall into two categories:(1)extractive summarization;(2)abstractive summarization.Extractive summarization systems extract representative textual units to normalized text.With the development of social media,a large number of texts with social attributes on the social web has emerged.Different from ordinary text,social media text is short,semi-structured,contains a lot of redundant information with emotions and opinions.A summary of social media is demanded for more applications such as public opinion monitoring and so on.Therefore,traditional text summarization systems are not applicable in summarizing social texts.This paper carries out research on three aspects:(1)To address the problem of semi-structured social media text,we propose a summary method based on the Dirichlet process mixture model.The model assumes that for each unit text segment,one and only one topic is expressed,which is either a background topic,or a topic drawn from a tag.It can automatically estimate the number of topics,and is capable to detect new topics.We present applications to tag-driven summarization,comparative summarization and update summarization.We conduct experiments on real world data sets,and present the model can well adapt to social media text.(2)For the problem of redundant emotions and opinions on social media,we study a novel problem that automatically generates a polling questionnaire,it assembles the task of summarization from a question bank collected from online debate forums.We propose a framework which is based on a bi-level topic sensitive question graph.Given a topic,select reasonable questions to generate a questionnaire.We conduct compre-hensive quantitative and qualitative experimental to verify that our system satisfies the general requirements for opinion polling composition.(3)We study a new problem summarization from microblog based on public opinion evolution.We use a summarization framework based on dynamic topic model to model public opinion and detect the evolution of topics,then sort the social media texts by keywords,emotions,and other features,finally extract representative microblog to construct the public opinion evolution summary.The experimental results show that the algorithm can find the reason of the sudden change in the public opinion and generate a summary that meets the public's information requirements.