Research On Event Representation Learning Based On Deep Learning

An event indicates the fact that specific participants and things interact at a specific time and place.As an important information carrier,event has become an important part of the mas-sive data on the Internet and has a profound impact on people’s work and life.Therefore,it is particularly urgent and important to accurately analyze and utilize massive event data.Event representation learning refers to data transformation and feature extraction of data containing events to obtain event representations to improve the performance of event-related downstream tasks,and plays an important role in downstream applications such as script event prediction,stock market prediction,question answering,sentiment analysis,fact checking,hot event de-tection and opinion analysis.Existing event representation learning approaches mostly take event triples（i.e.<subject,predicate,object>）as input.On the one hand,the interdependence between event extraction and event representation learning is ignored.On the other hand,the input event template is fixed,lack of handling for events that may have multiple arguments.In addition,existing training methods of event representation learning are relatively simple.To address the above issues,this thesis focuses on the study of event representation learning based on deep learning.The main contributions are summarized below:1.To address the problem of the fixed input and losing the sight of the interdependence between event extraction and event representation learning,this thesis proposes an event rep-resentation learning framework for joint event extraction.Under this framework,an event en-coder for flexible input（i.e.<trigger,argument₁,argument₂,...,argument_k>）is proposed while event extraction and event representation learning is connected.On the one hand,as the event components are continuously extracted,the model updates the event representation to help the extraction of subsequent event arguments.On the other hand,the model learns to capture event semantics in the process of event argument extraction.The experimental results on the ACE2005 dataset indicate that the proposed learning framework is better than other baselines.2.To address the problem that the existing training methods of event representation learn-ing are relatively simple,this thesis proposes an event representation learning framework based on event similarity ranking.Under this framework,five event augmentation strategies,includ-ing event argument synonym replacement,contextual events,events of the same type,events of different types,and trigger word antonym replacement,are proposed to provide richer infor-mation for the training of models.Meanwhile,an objective function based on event similarity ranking is proposed.The proposed objective function learns the representations for multiple events at the same time to reflect the ranking of the similarity between anchor event and target events.The experimental results on the ACE 2005 dataset indicate that the proposed learning method outperforms the traditional training methods.