A sentiment label is only useful if the product can act on it
A 37,000-plus review study opens a harder question than positive or negative: what should a product team change next?
A score is not a diagnosis
A public Thread review project works with more than 37,000 Android and iOS reviews and explores multi-label sentiment with Word2Vec. The dataset gives the work scale. The useful product question begins after the label is assigned.
A negative review can mean a confusing setup, a broken flow, a moderation concern or a feature users cannot find. If every complaint collapses into one bucket, the model can summarize dissatisfaction while leaving the team unable to choose a fix.
Keep feeling, topic and action separate
Sentiment describes tone. Topic describes what the person is talking about. Action describes what the product team can test. Those are related signals, but they answer different questions. A useful analysis preserves the distinction and lets a reviewer move from a theme back to the reviews that formed it.
Evaluation should do the same. Pair aggregate performance with per-class precision and recall, inspect examples the model gets wrong, and note which kinds of reviews are missing. One headline number cannot show whether a small but consequential group disappears inside the average.
Make the handoff useful
The deliverable is not a sentiment chart. It is a short, reviewable path from comment to theme to product question: what changed, for whom, and what would count as evidence that the change helped? That is how text analysis becomes a decision surface instead of a report that ends at the model.
Explore the project repository · Multi-Label-Text-Sentiment-Thread-Mobile-App ↗