Project 06 / 10 · RAG & NLP
Sentiment Analysis — DistilBERT + BiLSTM
Transformer vs. recurrent architectures on 204K HuffPost headlines, with a live RSS pipeline surfacing hour-by-hour sentiment from 6 major outlets.
01 / Context
The problem
The project compares transformer and recurrent architectures under one three-class news-labeling setup, then carries the stronger model into a live multi-source news monitor.
02 / System workflow
From input to output
- 01Headline / RSS
- 02Tokenization
- 03DistilBERT inference
- 04Class probabilities
- 05Trend dashboard
03 / Architecture
Inside the system
The deployed path tokenizes headlines to 64 WordPiece tokens, passes them through six DistilBERT layers, and classifies the pooled representation into negative, neutral, or positive sentiment.
A three-layer BiLSTM with additive self-attention provides the comparison architecture on the same dataset and label scheme.
FastAPI exposes single and batch inference while Streamlit provides manual analysis, batch review, source heatmaps, and hourly or daily trends.
04 / Engineering decisions
What shaped the build
Decision 01
Fair model comparison
DistilBERT and BiLSTM are evaluated on the same HuffPost dataset and label construction rather than unrelated benchmarks.
Decision 02
Operational input
RSS ingestion turns static model evaluation into continuous analysis across six news publishers.
Decision 03
Model delivery
The deployed checkpoint is downloaded and cached on first Streamlit boot, keeping the repository and hosted runtime manageable.
05 / Evidence
Measured, not implied
80.4%
Accuracy
Fine-tuned DistilBERT on the three-class task.
0.919
ROC-AUC
Deployed transformer performance.
+4.4pp
Model lift
Accuracy improvement over BiLSTM with self-attention.
06 / Delivery
What shipped
- 01Streamlit Community Cloud hosts the interactive application.
- 02Docker Compose runs the API and interface together for local or container deployment.
- 03GitHub Actions packages and ships updates on each push.
Constraint documented
Neutral headlines are the hardest class, with F1 0.60, because broad category-derived labels cannot cleanly represent mixed sentiment.