Nested event extraction method based on domain pre-training

Abstract

The invention discloses a nested event extraction method based on domain pre-training. The method comprises the following steps: 1, acquiring and preprocessing a field corpus, constructing a field professional word list for the corpus by adopting an adjacent word solidification degree method, and randomly extracting text data for manual labeling to obtain a nested event text data set; 2, by using a domain corpus and a domain professional word list, pretraining domain language model CaseBERT and adding news category classification pre-training task to the universal language model BERT; 3, defining a template for extracting nested event information in a layered manner, and carding the hierarchical relationship of nested events; 4, adopting a CaseBERT model and a predefined nested event extraction template to perform trigger word and event attribute joint extraction on the nested event text data set. The method is suitable for a domain multi-body nested event extraction task, and the accuracy of domain nested event extraction is effectively improved by pre-training the domain language model and pre-defining the nested event hierarchical extraction template.

Description

technical field [0001] The invention belongs to the technical field of natural language processing, specifically relates to the technical field of text information extraction, and more specifically relates to providing a nested event extraction method based on domain pre-training. Background technique [0002] With the explosive growth of Internet information, the need to quickly and accurately obtain the required information from the vast information sources is becoming more and more urgent. Event extraction is an in-depth research task of information extraction. It aims to extract events of interest to users from plain text, and present them to users in a structured form, and then provide them with query, analysis and utilization. Common important downstream applications , such as: building knowledge graphs, intelligent question answering and information retrieval, etc. Event extraction from text, especially multi-agent nested event extraction, has become a research diffi...

AI Technical Summary

Problems solved by technology

First, solve the problem of directly migrating BERT to the vertical domain, which will be limited in specific domain-based scenarios, and its performance will decline

The present invention proposes a method for pre-training a domain language model, taking the news case domain as an example to improve the domain language processing capability; secondly, factors such as overlapping attributes of multi-subject nested events, event attributes scattered in different sentences and other factors cause information extraction The problem of poor performance

Images