Bhavik-Ardeshna/Question-Answering-for-Low-Resource-Languages

Cascading Adaptors to Leverage English Data to Improve Performance ofQuestion Answering for Low-Resource Languages # Cascading Adaptors to Leverage English Data to Improve Performance of Question Answering for Low-Resource Languages Our work contributes by evaluat-ing cross-lingual performance in seven languages- Hindi, Arabic, German, Spanish, English, Viet-namese and Simplified Chinese. Our models areevaluated on the combination of XQuAD and datasets which are similar to SQuAD. For more details on how the models were created, please refer to our paper, Cascading Adaptors to Leverage English Data to Improve Performance ofQuestion Answering for Low-Resource Languages This repository contains both links to models at Huggin Face 🤗, and Langauge/Task Adapter in Task and Language Adapter with all configurations. ## Fine-Tuned Model at Hugging Face 🤗 | Language | mBERT | XLM-RoBERTa | |:----------:|:-------------:|:-------------:| | Arabic (ar) | multilingual-bert-base-cased-arabic | xlm-roberta-base-arabic | | German (de) | multilingual-bert-base-cased-german | xlm-roberta-base-german | | Spanish (es) | multilingual-bert-base-cased-spanish | xlm-roberta-base-spanish | | Arabic (ar) | multilingual-bert-base-cased-arabic | xlm-roberta-base-arabic | | Chinese (zh) | multilingual-bert-base-cased-chinese | xlm-roberta-base-chinese | | Vietnamese (vi) | multilingual-bert-base-cased-vietnamese | xlm-roberta-base-vietnamese | | English (en) | multilingual-bert-base-cased-english | - | ## Dataset Size The following table shows how much data is in each language: Split | en | de | es | ar | zh| vi | hi | |:---: |:---: |:---: | :---: |:---: | :---: | :---: | :---: | train | 12780 | 5707 | 6443 | 6525 | 6327 | 6685 | 6854 | test | 1148 | 512 | 500 | 517 | 504 | 511 | 507 | ## Conclusion We have investigated the efficacy of cascading adapters with transformer models to leverage high-resource language to improve the performance of low-resource languages on the question answering task. We trained four variants of adapter combina …