Research On Chapter-level Event Extraction Technology Based On Multi-granularity Fusion And Contrastive Learning

Nowadays,massive data have flooded every corner of people's lives.On the one hand,these information have promoted the rapid development of various information technologies,but on the other hand,they have also brought about the problem of information overload.It becomes more difficult to find exactly what you need in the abundance of information.Information extraction technology is oriented to text data,trying to accurately extract important information from semi-structured or unstructured data,which brings a lot of convenience to the processing of text.Event extraction is an important type of information extraction technology.Its goal is to extract the required event elements from the text.To this end,it is necessary to mine the deep semantics of the text.It is not only necessary to automate word,grammar,and context analysis,but also to identify and judge complete event information in combination with text-level knowledge concepts or higher-level semantic common sense.The current event extraction technology is still insufficient in the classification of multi-granularity and hierarchical event elements.In this regard,this paper studies the problem of document-level event extraction,and mainly carries out the following work.1)Multi-granularity gated fusion document-level event detection: Event elements will be distributed in different sentences,and the analysis of trigger words and single sentences alone is not enough to solve the task of event extraction at the document level.For this reason,this paper designs a document-level event detection model with multi-granularity gate fusion,and constructs two types of single sentence and paragraph levels.The features are combined with the forward and reverse,single sentence and paragraph features through the gating mechanism,and event elements are extracted across sentences.The experimental results show that the proposed model improves the F1 value by nearly 1.5% compared with the existing work.2)Hierarchical event argument role classification using comparative learning: Aiming at the inconsistency between training and inference phase of the current event argument role classification model,a hierarchical event argument role classification model incorporating contrastive learning is designed.The experimental results show that the proposed model improves the F1 value by nearly 0.6% compared with the existing work.3)Implementation of a document-level event extraction system: Based on the above two technologies,a document-level event extraction system is constructed,which can convert document-level content into structured event information.After inputting the document information according to the specific template format,the corresponding event information can be extracted by the system.