Supervised Learning for Enhanced Text Summarization in NLP Systems

This research develops a supervised learning framework for improved text summarization in Natural Language Processing (NLP) systems, including the aspects of text relevance, coherence and abstraction. A structured pipeline starting from text normalization and sentence tokenization is proposed to ensure the clean and structured input to the model. Rapid Automatic Keyword Extraction (RAKE) is used to efficiently select the features and thus, help in underscoring the terms which are contextually significant during the summarization process. To identify summary worthy sentences, a deep contextual understanding based on a fine tuned BERT sentence classifier is leveraged. All the system is built with the Hugging Face Transformers library with PyTorch as the backend which allows for fast model training and deployment. Experimental results on benchmark datasets show the baseline models are unable to achieve the results as good as or better than the advanced models in terms of ROUGE scores and sentence classification accuracy. However, our approach is able to achieve extractive precision along with abstractive flexibility and thus provides a scalable and domain adaptive solution to real world NLP applications that generate high quality, human like summaries.