Event Extraction:Algorithms And Applications

Event extraction,an important technology of information extraction traced back to 1980s,has been popular due to the advent of big data and the rapid development in natural language processing technology.The event extraction system can help people quickly find the information they need from the growing mass of information,and automatically extract,classify and reconstruct the information.At present,the event extraction system has been widely used in various fields.The enterprise can extract the user's evaluation of the product from the comment data to improve their product;The government can respond to emergencies by extracting daily hot spots from news data.In the early stage of event extraction task,the rule-based method was usually adopted due to the limited computing resources and lack of(manually)annotated data sets.With the increasing of computing power and the availability of(manually)annotated data sets,the rule-based methods begin to transfer to machine learning methods.In recent years,with the rapid development of deep learning,researchers have begun to use neural networks to extract events.Targets extraction and opinion words extraction in review event extraction are interdependent and related,and triggers extraction and arguments extraction in the news event extraction are also interdependent and related,but the traditional neural networks approaches treated them separately.This thesis does some research about how to make full use of feature information of the subtasks and treats them jointly with neural networks.on this basis,the Chinese military news event extraction system was developed.Paper work including:1.This thesis defines the event extraction task as a sequence labelling task,and designs a model which extracts features by bidirectional long short term memory networks and uses conditional random field to inference the best sequence.The experimental results show that this model achieves better performance than the methods which not use neural networks.2.In order to make full use of the feature information of the targets and the feature information of the opinion words,this thesis proposes a multi-task interactive learning model based on attention mechanism.By means of attention mechanism,this thesis hopes that the targets and opinion words can pay attention to each other,so as to accurately utilize the feature information of each other.The experimental re-sults on three benchmark datasets show that the multi-task interactive model has achieved the state-of-art or comparable performance.3.This thesis defines a new event extraction task in the military news domain,and designs the trigger-based templates and then extracts events based on the traditional rule-based method.In order to prove that multi-task interactive model is reasonable and effective,this thesis applies multi-task interactive model to this new task.At the same time,this thesis introduces two domain features in order to mitigate the unknown words in the triggers and arguments extraction tasks.The experimental results show that the multi-task interactive model proposed is much better than rule-based method,and the two domain features proposed also make the model better.Combining the above two methods of event extraction,a Chinese military news event extraction system was independently developed and adopted by a well-known domestic military research institution.The data and source code is available at https://github.com/qfzxhy/EventExtraction.