This repository documents a manual sentiment annotation case study created using Label Studio.
The project demonstrates a structured annotation workflow, explicit labeling rules, and quality-oriented decision making suitable for AI training and evaluation.

All annotations were performed manually and exported in a machine-readable JSON format.

• 60 short texts
• Simulated user reviews and feedback
• Mix of clearly polarized, neutral, and ambiguous statements
• Each text annotated with one overall sentiment label

Sentiment labels

• Positive
• Neutral
• Negative

The annotation interface was configured in Label Studio using a single-choice sentiment schema with an optional comment field.
Comments were added selectively to document ambiguity or rule application.

Labeling template configuration:

The annotation process followed clearly defined rules to ensure consistency and reproducibility:

• Single-label classification per text
• Dominant sentiment rule applied to mixed statements
• Explicit handling of contrast markers such as “but”, “although”, “however”
• Clear distinction between neutral statements and weak positives
• Annotation comments used only when necessary to explain decisions

The example below illustrates a labeled entry with a comment explaining the applied rule in a mixed-sentiment case.

In this case, the initial positive sentiment is overridden by an explicitly stated negative experience, resulting in a Negative label.

• Annotation tool: Label Studio
• Input format: CSV
• Output format: JSON
• Annotation type: Manual (human-in-the-loop)

Each task includes:

• One completed annotation
• Optional annotation comment (when applicable)
• Lead time metadata reflecting realistic annotation speed

.
├── data/
│   ├── sentiment_portfolio.csv
│   └── sentiment_portfolio.json
├── images/
│   ├── sentiment-annotation-labelling.png
│   └── sentiment-annotation-entry.png
├── docs/
│   ├── Guidelines_Sentiment_Annotation.pdf
│   └── Case_Study_Sentiment_Annotation.pdf
└── README.md