|
--- |
|
license: apache-2.0 |
|
task_categories: |
|
- question-answering |
|
language: |
|
- am |
|
pretty_name: AmaSquad |
|
--- |
|
# AmaSQuAD - Amharic Question Answering Dataset |
|
|
|
## Dataset Overview |
|
AmaSQuAD is a synthetic dataset created by translating the SQuAD 2.0 dataset into Amharic using a novel translation framework. The dataset addresses key challenges, including: |
|
- Misalignment between translated questions and answers. |
|
- Presence of multiple answers in the translated context. |
|
|
|
Techniques such as cosine similarity (using embeddings from a fine-tuned Amharic BERT model) and Longest Common Subsequence (LCS) were used to ensure high-quality alignment between questions and answers. |
|
|
|
## Key Features |
|
- **Language**: Amharic, a widely spoken Semitic language with limited NLP resources. |
|
- **Data Size**: Includes training and development sets based on SQuAD 2.0, tailored for extractive machine reading comprehension. |
|
- **Use Case**: Designed for training and evaluating Amharic Question Answering systems, particularly extractive QA models. |
|
|
|
## Applications |
|
- Developing and benchmarking machine reading comprehension models for Amharic. |
|
- Bridging the resource gap in low-resource language NLP research. |
|
|
|
## Caveats |
|
- As a synthetic dataset, some translation-induced artifacts may be present. |
|
- The dataset complements but does not replace the need for human-curated Amharic QA datasets. |
|
|
|
## Citation |
|
If you use this dataset, please cite: |
|
Hailemariam, N. D., Guda, B., & Tefferi, T. *XLM-R Based Extractive Amharic Question Answering with AmaSQuAD*. Carnegie Mellon University. |
|
|
