A Joint Event Extraction System Based on Fusion of Feature and Attention Mechanism

Abstract

The present invention provides a joint event extraction system based on the fusion of features and attention mechanisms, including: an input layer, a feature extraction layer, a trigger word and argument recognition layer, and an event post-processing layer; the information output by the pre-training layer is combined with speech and speech The location information is input into BiGRU and CNN to obtain the contextual semantic feature matrix and local feature matrix of the sentence, and then use the attention mechanism to fuse these two features, and finally combine the three representations for event extraction; the word embedding layer uses a pre-trained model BioBERT, and fine-tune it to improve the effect of the model; event joint extraction avoids cascading errors and improves the effect of the model.

Description

technical field [0001] The invention belongs to the field of event extraction, and in particular relates to an event joint extraction system, equipment and storage medium based on the fusion of features and attention mechanisms. Background technique [0002] The definition of biological event consists of four parts, namely event type, event trigger, event argument and argument role. For example, the following sentence S contains two events. The first event type is Development, the trigger word is formation, the event argument is capillary tubes, and the argument role is Theme type; the second event type is Negative Regulation, triggering The word is inhibited, the event argument is Thalidomide, and the argument role is of type Theme. [0003] S: Thalidomide inhabited the formation of capillary tubes. [0004] According to the above definition of biological events, biomedical event extraction is divided into two sub-tasks, trigger word recognition and argument classificatio...

AI Technical Summary

Problems solved by technology

[0009] However, the current research on event extraction models only uses a part of shallow semantic information, and most of the event extraction models are pipeline models, which use pipeline processing methods, which can easily lead to errors generated in the trigger word recognition stage being transmitted to arguments classification stage

Images