Combining Deep Learning With Multiple Features For Event Detection And Summarization

In recent years,with the development of Internet,online social platform has found extensive application and can broadcast news with far greater speed and spread than traditional social media,especially when something huge happens.As a well-known online social platform which owns enormous users,event detection based on information on Twitter has received great research and attention.It seems more meaningful to detect what exactly has happened on Twitter and thus to react correspondingly.The whole task can be split into two important subtasks: event detection and event summarization.The former detects what happened on the platform while the latter summarize the event.We focus event detection and summarization on tweets of earthquake domain,of which the core thing is detection of what happened by clustering and extraction of representational tweets as summaries by calculating similarities.Based on the investigation,statistical-feature based methods have dominated the field of event detection and summarization.With the development of Deep Learnnig,event detection and summarization is facing new opportunities to reach a breakthrough.The contents of this paper are organized as follows:1.We extract lexical features with linguistic tools,and topic features with the Biterm Topic Model,and weight features with the graph-of-words and k-degeneracy,which will all append to the BLSTM-based joint model.2.We construct the deep learning based event detection and summarization framework with representations learned by CNN and BLSTM respectively.Experiment shows that the BLSTM performs better than CNN in the whole task when no external features are added.3.We then combine the extracted multiple features with BLSTM-based joint-trained framework.Lexical features are added to the input of LSTM,and topic features are combined with the input of clustering,and weight features in are used facilitate the scoring of tweets in clusters.Experiments show that both topic features and weight features in joint model have significant influence on the improvement of performance on our task.4.After combining statistical features with LSTM-based joint model,the Cminand ROUGE-1 on 47 earthquake event data reached to 45.11 and 21.62 respectively,which performs better than previous works in the field of event detection and summarization.Thus,experiment results demonstrate the rationality and effectiveness of our deep learning based joint model with multiple features.