AmashaHw/sinhala_intent_commands_dataset

language:

si
tags:
nlp
intent-classification
sinhala
voice-commands
accessibility
license: apache-2.0

# Sinhala Intent Commands Dataset

## 📌 Overview

This dataset contains Sinhala natural language command variations designed for training an intent classification model in a voice-assisted smart reading system.

The dataset enables users to interact with documents using flexible Sinhala voice or text commands, supporting accessibility-focused applications such as assistive reading for visually impaired users.

---

## 🎯 Purpose

This dataset was created as part of the research project:

**"Intelligent Segmentation and RAG-based Summarization with Non-Linear Voice Navigation for Sinhala OCR Text"**

It supports:

* Sinhala voice command understanding
* Non-linear document navigation
* Context-aware content retrieval
* Accessibility-focused smart reading systems

---

## 📊 Dataset Structure

| Column | Description |
| ------ | -------------------------- |
| text | Sinhala command input |
| intent | Corresponding intent label |

---

## 🧠 Intent Classes

The dataset includes multiple intent categories such as:

**Navigation**

* NEXT
* PREV
* READ
* REPEAT

**Structure-Based Navigation**

* SECTIONS
* JUMP_SECTION
* READ_SECTION

**Search**

* SEARCH

**Summarization**

* SUMMARIZE_SHORT
* SUMMARIZE_MEDIUM
* SUMMARIZE_LONG
* SUMMARIZE_DOC_SHORT
* SUMMARIZE_DOC_MEDIUM
* SUMMARIZE_DOC_LONG

**Question Answering**

* ANSWER_CHUNK_SHORT
* ANSWER_CHUNK_MEDIUM
* ANSWER_CHUNK_LONG
* ANSWER_DOC_SHORT
* ANSWER_DOC_MEDIUM
* ANSWER_DOC_LONG

Each intent contains multiple Sinhala paraphrases to improve model generalization.

---

## 📈 Dataset Size

* Total samples:1245
* Balanced dataset: approximately **50 samples per intent class**

---

## ⚙️ Usage

### Load dataset in Python

```python
import pandas as pd

df = pd.read_csv("sinhala_intent_commands_dataset.csv")
print(df.head())
```

---

## 🧪 Model Training

This dataset was used to train an intent classification model using:

* TF-IDF with character n-grams (3–5)
* Linear SVM (LinearSVC)
* Calibrated classifier for confidence estimation

This approach helps handle:

* Sinhala language variations
* ASR (Automatic Speech Recognition) noise
* Short command-based inputs

---

## ⚠️ Limitations

* May not cover all Sinhala dialects and linguistic variations
* Contains ASR-like variations but may not fully represent real-world speech errors
* Some noise may exist due to OCR-based preprocessing
* Limited to command-style inputs rather than full conversational language

---

## 🔮 Future Improvements

* Expand dataset with real voice-transcribed Sinhala data
* Include more dialect variations
* Improve robustness to ASR errors
* Extend to conversational intent understanding

---

## 📜 License

This dataset is released under the **Apache 2.0 License**.

---

## 👤 Author

**Amasha Hewagama**
Sri Lanka Institute of Information Technology (SLIIT)

---

## 🔗 Citation

If you use this dataset, please cite:

Author: Amasha Hewagama
Title: Sinhala Intent Commands Dataset
Year: 2026

---

## 🙌 Acknowledgements

This dataset was developed as part of an undergraduate research project in Data Science, focusing on accessibility and intelligent document interaction using Sinhala language technologies.