Applying Graph Neural Networks and Reinforcement Learning for Abstractive Summarization of Long Texts

Text data analysis tasks have attracted significant attention from researchers in recent years, with notable advancements achieved by models such as ChatGPT in various applications, including dialogue generation and question answering. Reinforcement Learning, and, specifically, Reinforcement Learning from Human Feedback has emerged as a key component contributing to their success. However, certain tasks, particularly the processing of long text sequences, have received comparatively less focus. The prevalent Transformer model, a cornerstone of SOTA models, exhibits quadratic complexity with respect to sequence length, making it inadequate for directly handling lengthy sequences. In this paper we propose an approach addressing the challenge of handling long text sequences in the summarization task. Leveraging pre-trained SOTA models fine-tuned specifically for text summarization, in conjunction with a Graph Neural Network model, this approach demonstrates the capability to process long sequences with lower computational complexity with respect to the length of the input sequence than other Natural Language Processing models. To evaluate results of our approach, we have fine-tuned several models using publicly available datasets with texts and their summaries (ArXiv, Gazeta), and, also, we have collected our own dataset of long texts in Russian, containing sets of all news in the past 24 hours and summaries with main news per this period. The proposed approach, applied to pre-trained models allows to improve their quality in terms of metrics, used by authors of models in the corresponding articles. Since application of this approach to already trained model does not require significant computational resources, researchers and developers can easily modify their existing models, increasing maximal length of processed texts, and improving quality of processing.