Task reformulation and data-centric approach for Twitter medication name extraction.

Area of Science:

• Natural Language Processing
• Biomedical Informatics
• Computational Linguistics

Background:

• Extracting medication names from social media is difficult due to imbalanced datasets and limited context in short tweets.
• Existing methods struggle with the nuances of user-generated content for medication identification.

Purpose of the Study:

• To develop and evaluate a data-centric approach for automatic medication name extraction from tweets.
• To improve upon existing methods in the BioCreative VII Track 3 challenge for medication name recognition.

Main Methods:

• Formulating the sequence labeling problem as text entailment and question-answer tasks.
• Developing a single model and an ensemble method incorporating dictionary filtering.
• Investigating the use of domain-specific and task-specific pretrained language models.

Main Results:

• A single model achieved a Strict F1 score of 0.77, outperforming the official baseline (0.758).
• The ensemble method with dictionary filtering reached a Strict F1 score of 0.804, achieving the highest performance among participants.
• The study highlights the effectiveness of data-centric strategies in this task.

Conclusions:

• The proposed data-centric approach significantly enhances medication name extraction from tweets.
• Further improvements can be achieved through specialized pretrained language models and continued data-centric refinement.
• This work contributes to more accurate фармаконадзор (pharmacovigilance) and public health monitoring using social media data.