Aspect-Based Persian Sentiment Analysis

Context

Spring 2023 · Academic transformer prototype · Maturity: Research.

Problem

Classify sentiment across multiple aspects rather than assign one global sentiment to a text.

Users

Academic research context.

My Role

Academic NLP implementation.

Product Decisions & Responsibilities

• Structured labels around 8 aspects and 7 sentiment levels.

Technical Approach

• Prepared train/dev/test records and transformer classification experiments.

Constraints

• Current evidence does not support a trustworthy final test-set metric.

Product Flow

1. Prepare labeled data
2. Split train/dev/test
3. Train transformer
4. Evaluate by aspect/sentiment

Outcome

• 4,048 train/dev/test records; 8 aspects; 7 sentiment levels.

What I Learned

• Evidence quality matters more than publishing a flattering metric.

Evidence Boundary

Next Iteration

Extend the product only after the next decision can be tied to stronger user, operational, or business evidence rather than adding features for their own sake.