Graph convolutional network biomedical information extraction method based on multi-head attention mechanism

Abstract

The invention discloses a graph convolutional network biomedical information extraction method based on a multi-head attention mechanism. The method comprises the following steps of: firstly, establishing a graph convolutional network model based on a multi-head attention mechanism; then training a hybrid model; optimizing the model and adjusting parameters to obtain a final model; and finally, performing relation extraction in the biomedical text and feature extraction of a pathological report by using the final model. According to the method, the influence of noise data can be effectively reduced while the performance of relation extraction is improved, and valuable content in long-distance information in the biomedical text is effectively kept. A transfer learning method is used for processing pathological reports of different formats and writing styles, and good universality and reusability are achieved. The method is applied to information and relation extraction with cancer pathology reports as data sources, the recognition effect is good, universality is high, and pathology detection efficiency is effectively improved.

Description

technical field [0001] The invention belongs to the technical field of biomedical information extraction, and relates to a text information extraction method using a graph convolutional network, in particular to a graph convolutional network biomedical information extraction method based on a multi-head attention mechanism. Background technique [0002] Cancer is the number one killer of human health. As an important and easy-to-observe morphological feature of cancer, phenotype provides a valuable window for understanding this complex thing. It is necessary to automatically extract the relationship between phenotype and diagnosis, and perform information extraction and semantic understanding on massive patient pathology reports. However, the complexity of pathological phenotypes and the complexity of diagnostic logic make the corresponding descriptions exist in pathology reports in the form of long and difficult sentences, which poses a great challenge to the task of relati...

AI Technical Summary

Problems solved by technology

Existing methods rely on traditional neural networks to extract relationships between long and difficult sentences, mainly using convolutional neural networks or recurrent neural networks. However, traditional neural networks have always had bottlenecks in processing overly long sentences. As the length of sentences continues to increase, the performance continues to decline. , it is difficult to meet the requirements of relation extraction tasks in pathological scenarios

[0003] Existing technologies are mainly based on recurrent neural networks and convolutional neural networks for medical text information extraction. However, pathology reports include long texts and complex descriptions, so entities across sentences are difficult to detect, which will lead to key information being easily extracted. was left out

Compared with traditional machine learning and neural network methods, graph neural networks can capture semantic and grammatical information in non-adjacent sentences by relying on the dependency structure between sentences. However, it is still difficult to distinguish the correlation of text features using existing methods.

Images