Synergy of Graph-Based Sentence Selection and Transformer Fusion Techniques For Enhanced Text Summarization Performance

Abstract

This paper presents a new method to improve text summarization by combining the strengths of Graph Neural Networks with Transformer-based models. Text summarization is pivotal in natural language processing; it deals with the compression of large documents into smaller summaries while retaining essential information. The work proposed here focuses on enhancing the summarization process through the segmentation of articles into Elementary Discourse Units and further selecting important ones for the summary. In this paper, we propose a multi-stage methodology involving the application of GNN for the selection of salient EDUs from input articles. The process, therefore, aims at arriving at the most important pieces of information requisite for summarization. Following that, we fuse the selected EDUs using BART (Bidirectional and Auto-Regressive Transformers), which is a recent Transformer-based architecture for text generation. In evaluating the performance of our method, experiments are carried out on the CNN/Daily Mail dataset since it is a widely used benchmark for text summarization. We compare our method against the baseline model in whose respect we assess its performance using ROUGE scores, a measure that judges the quality of summaries against human-written references. The experimental results show that the proposed approach can outperform the baseline methods with ROUGE scores. The precision, recall, and F1 scores achieved using our method are higher than those obtained using other baseline methods, implying that our method is suitable for generating informative and coherent summaries. Our study suggests that the performance of GNN, in combination with Transformer-based models for text fusion, holds promise in yielding an effective methodology for text summarization. We present a novel method for improving text summarization, a fundamental task in natural language processing. Summarization is aimed at distilling important information from a given huge amount of text or document into shorter, more comprehensive summaries while retaining the main points. Our approach optimizes the process by segmenting the articles into Elementary Discourse Units (EDUs) and selecting the most important ones for summarization. In the first stage of our approach, we apply Graph Neural Networks (GNN) to identify the most important EDUs from the input articles. This step is a way to bring out information that should be captured for summarization purposes. Next, we use BART to fuse these selected EDUs, following a common Transformer architecture famous for text generation tasks. We validated our method by performing experiments on the CNN/Daily Mail dataset, benchmarked widely in the field. We compared the proposed model against baseline methods for a performance measure based on ROUGE scores, primarily indicating similarity between summaries and human-generated references. As observed, our approach has outperformed the baseline models in terms of precision, recall, and F1 scores. It can be concluded, therefore, that our method yields more informative and coherent summaries. This study establishes that GNNs convey content selection capability that works best with mapping information of EDUs to their source text, and fusing information using Transformer-based models in text summarization presents promising advances.