Zero-shot Sentiment Intelligence Pipeline

Problem

Teams that want sentiment signal on public feedback usually face a choice: label a training set and fine-tune a classifier, or accept whatever a generic off-the-shelf sentiment model gives them. Both are slow to start and awkward to adapt when the domain or label set changes. This pipeline was built to skip that trade-off entirely for exploratory and monitoring use cases.

Approach

The pipeline is built around facebook/bart-large-mnli used in zero-shot classification mode — candidate labels are passed at inference time instead of baked into a fine-tuned head, so the same pipeline can classify against a new label set without retraining anything.

• Config-driven design: swapping the input source, candidate labels, or reporting cadence is a config change, not a code change.
• Generates sentiment time-series, label share, and confidence diagnostics automatically from each run.
• Structured for rapid extension into a scheduled batch job or a lightweight API, without redesigning the core classification logic.

Architecture

Because the model runs zero-shot, there’s no training loop or labeled dataset to maintain — the orchestration layer focuses instead on batching inputs, tracking confidence, and exporting dashboard-ready artifacts (daily counts, trend charts) that a non-technical stakeholder can read directly.

Result

The pipeline turns raw public feedback into a daily sentiment trend and label-share breakdown with no fine-tuning step, and the config-driven structure means it’s already set up to be pointed at a different domain or a multilingual dataset without a rebuild.