Using distant supervision to augment manually annotated data for relation extraction.

Area of Science:

• Biomedical Natural Language Processing
• Machine Learning
• Data Science

Background:

• Deep learning models in natural language processing (NLP) require extensive annotated data.
• Biomedical literature often has limited labeled datasets, making model training difficult and expensive.
• Human annotation requires significant effort and specialized domain expertise.

Purpose of the Study:

• To address the challenge of limited labeled data in biomedical NLP.
• To explore data augmentation techniques using distant supervision for deep learning models.
• To improve the performance of NLP models in the biomedical domain despite noisy data.

Main Methods:

• Augmenting manually annotated data with large datasets obtained through distant supervision.
• Applying heuristic methods to filter and remove noise from distantly supervised data.
• Utilizing transfer learning-inspired techniques to train the deep learning models.

Main Results:

• The developed methods successfully filtered noisy annotations from distant supervision.
• Models trained on augmented, cleaned data demonstrated superior performance compared to those trained solely on original manual annotations.
• The approach effectively mitigated the impact of data scarcity in biomedical NLP tasks.

Conclusions:

• Distant supervision, coupled with noise reduction and transfer learning, is a viable strategy for enhancing deep learning models in biomedical NLP.
• This method offers a cost-effective solution for building robust NLP models when labeled data is scarce.
• The findings suggest a promising direction for future research in automated biomedical text analysis.