Split extraction from sentiment
One fine-tuned DistilBERT checkpoint extracts aspect spans; another classifies sentiment. Deterministic rules normalize known aliases, generate tags and aggregate saved predictions into product insights.
NLP · Product engineering / Hosted demo
Review Tag Generator
Two fine-tuned models turn reviews into feature-level sentiment. Every insight keeps its source text, with separate reviewer corrections and annotations.
01 / THE PROBLEM
A star rating hides which part of a product worked or failed. Feature-level analysis needs both an aspect span and its sentiment, while keeping the original evidence inspectable. A useful product also needs to distinguish model predictions from human corrections.
02 / FOLLOW THE WORKFLOW
A simplified example of the documented flow. Sample outputs are authored illustrations, not live model predictions.
Synthetic review
“The battery lasts all day.”
Start with a product review.
Aspect span: battery
The [battery] lasts all day.
Identify the feature mentioned in the text.
Illustrative label, not live inference
Battery → Positive
Attach sentiment to the extracted aspect.
Preview first → explicitly save
Battery life · “lasts all day”
Keep the supporting text beside the feature.
Step 1 of 4 · Review
03 / ENGINEERING
One fine-tuned DistilBERT checkpoint extracts aspect spans; another classifies sentiment. Deterministic rules normalize known aliases, generate tags and aggregate saved predictions into product insights.
A review can be analyzed without immediately contributing to saved product counts. Explicit saving, durable bulk-import jobs and row-error exports make data changes visible.
Accept, edit or reject a suggestion, correct sentiment, or annotate missed spans. Original predictions and decision history remain traceable. These edits do not automatically retrain the model or change product totals.
The full-review sentiment policy was selected on validation before the held-out test. Reported extraction and sentiment scores have different denominators and do not establish complete tag accuracy.
04 / EVIDENCE & LIMITS
The saved checkpoints reproduce the documented frozen-test results: extraction precision 64.85%, recall 77.51%, and sentiment accuracy 82.49%. Sentiment evaluation uses gold aspects. A separate ten-request warm local smoke measurement reported 566.45 ms median preview time on Apple M4, including scheduling and polling; it is not a production latency benchmark.
End-to-end canonical-tag quality is not yet established. Full-review context can confuse contrasting aspect sentiments. The hosted backend may need time to wake; optional cloud AI explanations are disabled according to the README. The retailer adapter is experimental.
05 / READ THE SOURCE
Based on public repository documentation reviewed on 7 November 2026. Results are attributed to the project’s recorded evaluations; they were not rerun for this portfolio.
TOOLS BEHIND THE WORK
Explore the stack. Drag the ribbon or select a tool.