Font Size: a A A

Enriching Biomedical Relationship Extraction With Dependency Syntactic Information

Posted on:2022-02-01Degree:MasterType:Thesis
Country:ChinaCandidate:J Y FanFull Text:PDF
GTID:2494306569980839Subject:Computer technology
Abstract/Summary:PDF Full Text Request
Automatic extraction of the relationship between entities included in biomedical literature plays an important role in many biomedical applications such as biomedical knowledge graph construction,biomedical knowledge discovery,etc.With the development of deep learning and natural language processing,automatic relationship extraction has made great progress.However,the domain-specific vocabulary,the long semantic distance between entities,and a large amount of useless information,still bring great challenges for biomedical relationship extraction.This paper aims to improve the performance of biomedical relationship extraction,utilizing the sequence information and the dependency syntactic information.Firstly,a GCN based approach is proposed to extract biomedical relationship.With the improved multi-head GCN structure and the root-based syntactic information extraction method,our approach is more appropriate for biomedical data.The evaluation on CPR,DDI,GAD corpus demonstrates that our method performs better than previous GCN based methods and many other methods.Besides,a pre-trained language model based approach with dependency syntactic information is proposed.This approach optimizes the preprocessing module,the model structure module,and the feature extraction module.In the part of preprocessing,entity mask,biomedical abbreviation replacement,and syntactic structure transformation are adopted and a syntactic information extraction method with chi-square keywords is introduced.Then the entity contextual representation and the syntactic representation would be extracted.In the part of model structure,a transformer encoder with syntactic information is applied to the Pub Med BERT model.Our experiments verify the effectiveness of the optimization of each part.Compared with state-of-the-art pre-trained model based approaches,our method can achieve better performance.
Keywords/Search Tags:relationship extraction, GNN, pre-trained model, syntactic information, biomedicine
PDF Full Text Request
Related items