RezaulKarim03/Joint_IDSF_Bangla

Joint_IDFS_Bangla

Dataset Description

The Joint_IDFS_Bangla dataset is an intent detection and slot filling dataset in the Bangla language, derived from the SNIPS dataset originally in English. It includes Bangla sentences annotated for intent classification and slot filling tasks, commonly used in natural language understanding and named entity recognition applications.

The training and development (validation) datasets were translated from English using the Llama 3 language model with some rule-based approach for post-processing and manually reviewed for consistency. Instances identified as problematic were removed. The test set was manually translated and annotated by two bilingual PhD students proficient in both English and Bangla. Verification was subsequently conducted by two additional bilingual PhD students, achieving an inter-annotator agreement (Cohen’s Kappa) of 0.83.

Tasks Supported

• Intent Detection (Text Classification)
• Slot Filling (Named Entity Recognition)

Languages

Bangla (bn)

Dataset Structure

The dataset consists of three separate files per split:

• seq.in: Contains sentences in Bangla.
• seq.out: Corresponding slot labels for each token in the sentences.
• label: Intent labels for each sentence.

Each line across these files corresponds to a single data instance.

Dataset Statistics

Split	Intents	Slots
Train	12,850	54,747
Validation	685	2,865
Test	694	3,105

Dataset Splits

• Training: 12,850 instances
• Validation: 685 instances
• Test: 694 instances

Source Data and Annotation

• Source: Converted from the English SNIPS dataset to Bangla using Llama 3 (meta-llama/Llama-3.3-70B-Instruct).
• Training/Validation Annotation: Automated translation (Llama 3) with rule-based consistency checks.
• Test Annotation: Fully manual annotation and verification by 4 annotators with a Cohen’s Kappa of 0.83.

Known Limitations

• The training and validation datasets are translated using an LLM (Llama 3). Although manually reviewed, some translation inaccuracies may persist.

Licensing and Use

Dataset can be used freely for research purposes, provided the original paper is cited:

@inproceedings{karim2025leveraging,
  title={Leveraging machine-generated data for joint intent detection and slot filling in bangla: A resource-efficient approach},
  author={Karim, AHM Rezaul and Uzuner, Ozlem},
  booktitle={Proceedings of the First Workshop on Challenges in Processing South Asian Languages (CHiPSAL 2025)},
  pages={208--216},
  year={2025}
}

Dataset Release Date

January 2025