02 / NATURAL LANGUAGE PROCESSING
Sentiment analysis for Taglish.
An XLM-R sentiment model trained on synthetic Tagalog-English intake responses.
TAGLISH SENTIMENT
“Okay naman ako,
pero medyo pagod.”
pero medyo pagod.”
TagalogEnglishContext
Question + answerXLM-RSentiment
EXAMPLE INPUT
Problem
Classify sentiment in Tagalog-English intake responses. Limited access to patient notes required synthetic training data.
My contribution
- Led the NLP pipeline and generated synthetic data, with validation involving a licensed psychologist.
- Implemented training and evaluation with class-weighted cross-entropy loss.
- Compared transformer models and selected XLM-R.
Method
01
Synthetic dataset
Generate Taglish question-answer pairs from psychological-profiler questionnaires.
02
Model comparison
Fine-tune and compare mBERT, TaglishBERT, and XLM-RoBERTa using question-aware inputs.
03
Class imbalance
Use class-weighted loss for positive, negative, and neutral examples.
Result
An XLM-R model that supplies positive, negative, or neutral sentiment labels to the project’s RAG pipeline.
LIMITATIONS
Outputs sentiment, not diagnoses. Trained on synthetic data; real intake data needs separate evaluation.
NEXT PROJECT