Performance comparison of transformer-based models for multi-reasoning in machine reading comprehension

Abstract

Machine Reading Comprehension (MRC) is an artificial intelligence task that ex amines a given passage or text and answers queries regarding it. The objective is to make an intelligent support system that has the ability to understand the contex tual information of the passage and give correct answers for multi-reasoning ques tions, commonsense based questions and multiple-choice questions, etc. One of the main challenges faced by MRC models in commonsense based and multi-reasoning questions is the need for understanding and reasoning beyond explicit textual infor mation. To enhance the capabilities of MRC systems in these areas, the research focuses on the comparative analysis of state-of-the-art transformer-based models in cluding BERT, ALBERT, RoBERTa, DistilBERT, MobileBERT, and ELECTRA. Our investigation specifically targets the enhancement of commonsense reasoning within MRC frameworks. In regards to this, we have used a binary decision mak ing approach in our algorithm, in order to achieve a better outcome from these transformer-based models. To evaluate the performance, the experiments were con ducted using CosmosQA dataset, which consists of narrative-driven questions that necessitate commonsense understanding to resolve.