Improving Keyword-Based Topic Classification in Cancer Patient Forums with Multilingual Transformers.

Area of Science:

• Computational linguistics
• Medical informatics
• Social network analysis

Background:

• Online patient communities facilitate crucial support networks for cancer patients.
• These forums contain rich, heterogeneous data valuable for research.
• Effective analysis of this data can improve patient support and information management.

Purpose of the Study:

• To classify discussion topics within an Italian cancer patient online forum.
• To enhance message reviewing and filtering through improved topic identification.
• To assess the effectiveness of combined text representation methods for this task.

Main Methods:

• A case study was conducted on user posts from an Italian cancer patient community.
• Text classification was performed by combining count-based (bag-of-words) and prediction-based (contextual embeddings) representations.
• Pre-trained multilingual models, such as BERT, were utilized to investigate their reusability in lower data regimes.

Main Results:

• Pairing bag-of-words representations with pre-trained contextual embeddings significantly improved prediction quality.
• The combined approach demonstrated effectiveness in handling linguistic ambiguities and misspellings.
• The study confirmed the reusability of pre-trained multilingual models on non-English data, even with limited datasets.

Conclusions:

• Combining traditional and advanced text analysis methods enhances topic identification in online cancer forums.
• This approach offers a robust solution for managing and understanding patient-generated content.
• Findings support the use of multilingual AI models in diverse, lower-resource medical data environments.