Event Detection Based On Deep Models

Nowadays,many real world events take place at every moment in the world,e.g.,public safely incidents(demonstrations,fire accidents,etc.)and sporting and recreational events(sports comperirions,performances,etc.).Most people know real world events through the Internet platform,e.g.,online news media and social media.It is very important to detect events through the Internet platform.This thesis mainly studies the event detection problem of two common scenarios,that is,the multi-domain event detection under single data source and the isomorphic event detection under different data sources based on transfer learning.For the multi-domain event detection under single data source,this thesis proposes a layer-wise deep stacking model to solve the problem of the scarcity of calibration samples in real-world scenarios by making full use of multi-domain features.This model consists of a Dropout module,a Block module and a Detector module.The Dropout module can avoid the problem of over-fitting,the Block module uses multiple claddifiers to fuse multi-domain data and the Detector module automaticially determines the depth of the model.For isomorphic event detection under different data sources based on transfer learning,this thesis proposes a transfer learning method based on the maximum classifier discrepancy model by considering joint distribution.On one hand,it gets close the joint distribution between different source data by minimizing the difference distance of maximum mean value.On the other hand,using the discrepancy between two classifiers to adversarially learn makes the features between different source data as close as possible,which can achieve transfer learning after training.In order to verify the validity of two models proposed in this paper,two datasets for different event detection tasks are collected,i.e.,a single-source Multi-domain event detection DataSet with 412 events and a multi-source event detection dataset with total 68 events of 3 data sources,and experiments are carried out in both datasets.Experiments show that the two models proposed in this paper have higher performance performance than other algorithms.In addition,through some feature analysis,visualization and other experiments,it can also show the superiority of the models.Therefore,the model proposed in this paper can deal with the corresponding event detection tasks well.