CSC525 · Applied NLP · COURSEWORK
Supportive Conversation Assistant
An educational, non-clinical conversation prototype for positive, neutral, and negative sentiment. The project combines a reproducible data-preparation pipeline, a trained classifier, curated responses, limited conversation state, and a command-line interface.
How it works
Text is normalized and deduplicated, then represented with TF-IDF features. A logistic-regression model predicts sentiment. The response layer selects an affirmation from a reviewed catalog and avoids repeating the same response consecutively.
A design decision
A curated response catalog makes the available responses directly inspectable. It also limits variety and personalization. The system uses a keyword-triggered resource response for certain high-risk phrases; that layer has limited coverage.
REPORTED FINDINGS
What the experiment showed
The coursework report records approximately 97.3% validation accuracy and 0.974 weighted F1 for three-class sentiment classification. The dataset combines curated movie-dialogue and customer-support text, with labels produced using heuristics or TextBlob.
Limits of the result
These scores describe agreement with labels in that dataset. They do not establish mental-health benefit, reliable understanding of unfamiliar language, or comprehensive crisis detection. Sarcasm, figurative language, and the limited response catalog remain important failure cases.
The next iteration
Evaluate a separately reviewed test set, publish results for each sentiment class, and document successful and unsuccessful conversations. A training or SOP assistant would be a separate extension with its own reviewed source material.
These are planned improvements, not current capabilities.