Research On Personal Event Extraction For Social Media Text

With the rapid development of the Internet and social media,more and more people use social platforms to share their views and describe personal events.These personal events contain rich information such as the motivation and results of user behavior,implicit emotions,etc.,and are of great value.In the face of massive personal event information,without the help of automated processing tools,it is difficult for enterprises or individuals to effectively obtain critical and effective information.Personal Event Extraction aims to automatically extract structured personal event information from massive event text,so it has important practical significance and application value.Currently,social media-oriented event extraction faces two challenges:1)open event types,and it is impossible to define all event types in advance;2)lack of corresponding annotation corpus.Existing event extraction methods either need to design corresponding extraction templates based on event types,which lacks flexibility and cannot extract new event types;or based on traditional machine learning models,which need to prepare sufficient labeled corpus for various event types.To this end,we plan to build a personal event dataset,and on this basis,study the problem of fewer training samples under certain event types in the personal event field.Based on Few-Shot learning,research on social media-oriented personal event extraction tasks.The main work of this thesis are:(1)Based on the FrameNet external knowledge base,we constructed a personal event ontology database.Based on Twitter data,we constructed a personal event dataset.(2)Aiming at the problem of scarce corpus under certain event types in the field of personal events,we have studied the event detection task based on a Few-Shot Learning framework and proposed a Multi-Attention-Based Prototypical Network(MAPN).The model enhances the feature representation of words in text by introducing external information such as part-ofspeech tagging,dependency parsing,a Bi-directional LSTM network and self-attention mechanism.In addition,we also introduce sample-level attention to enhance the calculation of prototype representations in the Prototypical Networks,which effectively improving the model's ability to respond to diverse text representations.The experimental results show that on the dataset constructed in this thesis,the model we proposed is better than the comparison model.(3)For the extraction of personal events,we propose a Graph Information Augmented Prototype Networks(GIAPN).Based on the MAPN model,for the characteristics of personal event text representation is more abstract and semantic expression is deeper,we constructed the dependency analysis result as a graph structure representation.By constructing the dependency analysis result as a graph representation,we introducing a Graph Convolutional Neural Network to model the dependency graph and integrating it into the word feature representation to enhance the grammatical structure information contained in the word feature.In addition,the model also realizes the joint extraction of event types and event arguments.The experimental results show that our proposed model has a greater performance improvement than the comparative models.