Project 3: Mood Journal Sentiment Tracker¶
Mentors: To be announced
Problem: Mood-tracking tools that route personal journal text through a third-party API force a privacy trade-off students shouldn't have to accept just to get sentiment trends over time.
Context: Fine-tunes an emotion classifier starting from bhadresh-savani/distilbert-base-uncased-emotion, based on dair-ai/emotion (public, on Hugging Face) and optionally extended with a handful of self-labeled journal entries. Training is required here, not just prompting.
Goals: Classify each journal entry's dominant emotion and chart the trend over time, without sending personal journal text to a third-party API.
Deliverables: A daily journal entry form + trend chart (e.g. stacked area of emotion scores per day), running the classifier locally so entries never leave the machine. Students label ~30 of their own entries, re-tune the base checkpoint on them, and compare classifications before/after tuning to see the effect of a small amount of domain-specific training data.
Showcase: TBD
References:
- CARER: Contextualized Affect Representations for Emotion Recognition (EMNLP 2018): the paper behind the
dair-ai/emotiondataset this project fine-tunes on, describing how the six emotion labels were derived. bhadresh-savani/distilbert-base-uncased-emotion: the exact base checkpoint this project starts from, already fine-tuned once on the same dataset, this project fine-tunes it again on a student's own entries.