The most crucial part of a search engine is ranking. Since the majority of search engines work with a significant amount of natural language data, an efficient ranking system requires an in-depth understanding of text meaning. Deep learning-based natural language processing models are now yielding encouraging results for ranking problems. In this paper, we propose an innovative approach of fusing a transformer-based model BERT with various listwise loss functions for ranking answers for a given user question. This approach uses BERT to learn contextual embedding, which has been applied to capture complex question-answer relations for ranking. The architecture employs a list-wise technique for training a ranking function by employing a question and its related answers as one instance and minimizing a loss function defined on both predicted and actual labels. We validated our approach on publically available datasets SemEval (2016) and SemEval (2017) using NDCG as evaluation metrics for ranking results.