A Robust Approach to Fine-tune Pre-trained Transformer-based models for Text Summarization through Latent Space Compression.