Semantic relation extraction aware of N-gram features from unstructured biomedical text

Area of Science:

• Biomedical Natural Language Processing
• Knowledge Graph Construction
• Machine Learning

Background:

• Semantic relation extraction is vital for building knowledge graphs from biomedical texts.
• Unsupervised generative models like Rel-LDA and Type-LDA are popular but ignore n-gram features.
• The bag-of-word assumption limits their ability to capture richer linguistic patterns.

Purpose of the Study:

• To propose novel unsupervised relation extraction models, Rel-TNG and Type-TNG, that incorporate Topic N-Grams (TNG).
• To overcome the limitations of existing models by leveraging n-gram features for improved semantic understanding.
• To evaluate the performance of the proposed TNG models against established methods.

Main Methods:

• Development of Rel-TNG and Type-TNG models utilizing the Topic N-Grams (TNG) framework.
• Application of collapsed Gibbs sampling algorithm for model inference.
• Experimental validation on the GENIA and EPI biomedical corpora.

Main Results:

• Rel-TNG and Type-TNG models demonstrate comparable performance to their unigram counterparts.
• The proposed TNG models show superior performance compared to Rel-LDA and Type-LDA when prior knowledge is utilized.
• The effectiveness of incorporating n-gram features in unsupervised relation extraction is highlighted.

Conclusions:

• Rel-TNG and Type-TNG models offer an effective enhancement over traditional LDA-based approaches for biomedical relation extraction.
• Leveraging Topic N-Grams is a promising direction for improving knowledge graph construction from unstructured biomedical data.
• The proposed models provide a valuable tool for advancing automated knowledge discovery in the biomedical domain.