Research And Implementation Of Event Detection And Summarization For Social Media

With the rapid development of Internet era,social media has become the main way for government agencies and media to release information,and it has also become an important way for users to obtain and disseminate information in real time,research on social media events has attracted a lot of attention from researchers.By leveraging the knowledge gained from these events,we have the potential to understand user interests and help governments and organizations make better decisions.This paper focuses on social media event detection and event summarization,which have great practical significance for the public opinion analysis and comprehensive governance of cyberspace.Most of the existing burst-based social media event detection models do not pay much attention to the impact of event elements,which have a greater effect on the iden-tification,screening and clustering of event tweets.Existing social media event summa-rization models are mostly Focuses on the extraction of keywords and key tweets,and due to the sparsity of social media data,the summaries obtained by such models are usually poor in information content,not smooth and concise,while the partial generation-based summarization models also do not focus on events,but simply reuse other text generation models.For the social media event detection task,this paper proposes an event detection model based on bursty elements and heterogeneous information network.The entire model first processes the input social media data stream by time window slices,and uses the burst properties of the When,Where,Who,What(4W)event elements to obtain event tweets.Then,the obtained event tweets are connected to form a heterogeneous infor-mation network through the connection of event elements,and the embedding of event tweets is regenerated through the heterogeneous information network,so that the tweets discussing the same events are closer in the vector space; finally Through the embed-ding of obtained event tweets,the Jarvis-Patrick clustering algorithm is used to cluster the tweets,and the obtained tweet event clusters are the detected events.Through exper-iment on public datasets,the model proposed in this paper achieves higher accuracy and less repetition rate,and detects more events than the baseline model.For the social media event summarization task,This paper proposes a social media event summarization model based on an element-aware extraction-generative paradigm.The extractor is a sequence labeling task whose purpose is to obtain saliency tweets con-taining more event information.The extractor mainly uses heterogeneous graph network and dilated gate convolution to mark whether the tweets is salient.These salient tweets with more event information are then fed into the pre-trained BART model and fine-tuned to get the event summary.Through experiment on public datasets,the model in this paper can obtain better performance summaries,and the results of manual evaluation show that the obtained summaries contain more event information,and are more fluent and concise.