Resolving passage ambiguity in machine reading comprehension using lightweight transformer architectures.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Machine Learning

Background:

• Online content complexity hinders efficient information retrieval.
• Existing deep learning models for Machine Reading Comprehension (MRC) struggle with ambiguity and high computational demands.
• Transformer models like BERT offer solutions but are resource-intensive.

Purpose of the Study:

• To develop a resource-efficient model for MRC that addresses passage ambiguities.
• To fine-tune the DistilBERT model for improved text comprehension and reduced computational costs.

Main Methods:

• Fine-tuning the DistilBERT model for the MRC task, creating Distil-BERT-MRC.
• Evaluating the model's performance on WikiQA, SQuAD 2.0, NewsQA, and Natural Questions datasets.
• Comparing Distil-BERT-MRC against other transformer models like RoBERTa and XLNet.

Main Results:

• Distil-BERT-MRC achieved 90.23% exact match and 91.42% F1 score on the WikiQA dataset.
• The model demonstrated competitive performance while utilizing reduced computational resources.
• Experiments confirmed the generalizability and resource efficiency of distilled transformer models for MRC.

Conclusions:

• Distilled transformer models, like Distil-BERT-MRC, offer an effective and resource-efficient solution for Machine Reading Comprehension.
• The proposed approach balances performance with reduced computational requirements, making MRC more accessible.
• This work contributes to advancing NLP by providing a practical alternative to complex transformer architectures.