The process of continuously keeping up to date with the state-of-the-art on a specific research topic is a challenging task for researchers, not least due to the rapid increase in published research. While existing tools, such as scholarly search engines or citation graphs, can help find related articles, they often fail to highlight the factual differences between papers that address the same research problem. As a result, identifying what is new or different in a given article remains a challenge. This is particularly important for writing surveys, conducting literature reviews, or evaluating the novelty of a contribution. In this thesis, we address this problem by introducing the concept of the Knowledge Delta(K∆), which refers to the factual differences between related scientific articles, and improve informativeness, particularly for long summaries. On the TalkSumm dataset, argumentative structure enhance summary quality, while on CL-SciSumm, abstract- based methods perform best, reflecting the close alignment between human-written short summaries and abstract content. For the second step, we explore different approaches for representing the extracted factual knowledge. We use Large Language Models, including GPT-4o, GPT-4o-mini, LLaMA3.1-8B, and LLaMA3.1-70B, to represent each of the research task, method, dataset, and results of each article. These representations are produced using designed prompts and are guided by either the full article or the AZ-extracted summaries. These representations are evaluated using the downstream task of Narrative Knowledge Delta (NK∆) generation. which focuses on identifying differences between pairs of scientific articles that tackle theFor the final step, we introduce the task of Narrative Knowledge Delta (NK∆) generation, same research problem, presented in a narrative form. We create two small, manually annotated datasets, one based on interviews with researchers and another based on a survey paper, to serve as ground truth. In addition, we construct a large-scale synthetic dataset using existing scientific annotations to serve as examples for LLM prompts in few-shot settings. To evaluate NK∆ outputs automatically, we apply the scientific fact- checking model MultiVerS to assess the LLM-generatedannotated claims. The results show that structured prompting, combined with examplesNK∆ sentences against manually from the synthetic data, improves the factual accuracy of generated comparisons.wepropose a method for automatically identifying and representing these differences. We define a specific form of Knowledge Delta called the Narrative Knowledge Delta(NK∆), which captures how a pair of scientific articles differ in terms of their researchtasks, methods, datasets, and results. We formulate three research questions that guide this work: how to extract factual knowledge from scientific articles, how to represent it in a structured way, and how to compare and evaluate the differences between articles. Based on these questions, we design a three-step methodology. First, we use Argumentative Zoning (AZ) to detect regions within a scientific article that cover the main components of an article (i.e. factual knowledge). Second, we represent the extracted knowledge using Large Language Models (LLMs) through structured prompts and templates to generateNK∆. Third, we evaluate the generated NK∆ using a scientific fact-checking model, MultiVerS, to assess the correctness of the comparisons against the original documents.To support the first step of the pipeline, we develop an AZ annotation platform that allows annotators to label rhetorical zones in full scientific articles. This tool aims to address the lack of existing AZ datasets at the full-paper level and handles the complexity of the annotation task. Using this tool, we conduct two rounds of annotation, both online and offline, with Master’s and Bachelor’s students and created a corpus of AZ- labeled articles focused on NLP and text mining. We also train a BERT-based classifier to automatically identify AZ zones. Our results show that high-quality annotations significantly improve model performance compared to a BERT model trained on a baseline corpus. Additionally, as a downstream task, we use the AZ-trained model in a scientific summarization pipeline. We experiment on two scientific summarization datasets: TalkSumm and Cl-SciSumm. Our experiments reveal that AZ-based approaches
