Saraiki Poetry Sentiment Analysis (NLP)

Overview

This project explores sentiment analysis of Saraiki poetry, a low-resource regional language, using classical NLP preprocessing and a lightweight deep learning model.

The focus is on:

• Manual data collection and labeling
• Emotion normalization and sentiment abstraction
• End-to-end NLP pipeline using Python and TensorFlow
• Preparing the model for potential mobile (Android) deployment via TensorFlow Lite

Dataset

• Language: Saraiki
• Domain: Poetry
• Total samples: 958 (full dataset kept private)
• Public release: Partial sample only

Annotation Strategy

Each poem was manually labeled with an emotion:

• happy, sad, surprise, fear, anger, disgust

Emotions were mapped to binary sentiment:

• Positive (p): happy, surprise
• Negative (n): sad, fear, anger, disgust

Only a subset of the dataset is uploaded to illustrate:

• Data format
• Labeling approach
• Reproducibility
• The complete dataset is retained for possible future academic use.

Data Cleaning & Analysis

Before training, the dataset was cleaned and analyzed to ensure consistency:

• Normalized inconsistent emotion labels (case differences, spelling variants)
• Verified emotion and sentiment distributions
• Analyzed emotion-to-sentiment relationships

Dataset Analysis (Visual)

The following visualizations summarize the dataset structure:

• Emotion distribution:

• Sentiment (positive / negative) distribution:

• Emotion → sentiment mapping:

Methodology

Preprocessing

• Data loaded from Excel files using Pandas
• Text tokenized using Keras Tokenizer
• Sequences padded/truncated to fixed length
• Out-of-vocabulary handling enabled

Model

A lightweight neural network was used due to dataset size:

• Embedding layer
• Global Average Pooling
• Dense (ReLU)
• Output layer (Sigmoid)

The model performs binary sentiment classification, not multi-class emotion prediction.

Training

• Train / test split: 80% / 20%
• Loss function: Binary Crossentropy
• Optimizer: Adam
• Epochs: 30

The model learns general sentiment patterns from poetic text rather than fine-grained emotional nuance.

Mobile Deployment

After training, the model is converted to TensorFlow Lite (TFLite).
This enables:

• On-device inference
• Low latency predictions
• Future integration with Android applications

Limitations

• Small dataset size
• Possible class imbalance
• Binary sentiment only
• Simple model architecture

Results should be interpreted as exploratory and baseline-level, not production-ready.

Technologies Used

• Python
• Pandas
• Matplotlib
• NLTK
• TensorFlow / Keras
• TensorFlow Lite

Note

This project emphasizes data preparation, language challenges, and applied ML workflow in a low-resource setting rather than optimized model performance.