Causal relationship extraction from biomedical text using deep neural models: A comprehensive survey.

Area of Science:

• Biomedical Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Causal relationship identification is vital for biomedical knowledge bases.
• This task is a fundamental yet challenging Natural Language Processing (NLP) problem.
• Few studies have addressed causal relation extraction despite its importance.

Purpose of the Study:

• To survey state-of-the-art research in causal relationship identification from biomedical texts.
• To evaluate various machine learning techniques, including deep neural networks.
• To investigate the impact of class imbalance on model performance.

Main Methods:

• Implemented and evaluated Multiview CNN (MVC), attention-based BiLSTM, and graph LSTM models.
• Utilized advanced word embedding models like ELMo and BioBERT.
• Assessed a baseline rule-based system and investigated data augmentation via random oversampling to address class imbalance.

Main Results:

• Deep learning models, particularly those leveraging transformer architectures (BioBERT), show promise.
• A simple random oversampling technique significantly improves performance by mitigating class imbalance.
• The study demonstrates considerable improvements over existing state-of-the-art systems.

Conclusions:

• Deep learning approaches, enhanced by oversampling for class imbalance, offer a powerful solution for biomedical causal relationship extraction.
• Further research can build upon these findings to develop more robust and accurate NLP systems for scientific literature.
• Addressing data challenges like class imbalance is key to advancing information extraction in the biomedical domain.