Real-Time Text Summarization Using Event Compression Memory in a Rolling-Window Transformer

View at publisher Summarise with VIRAT
Authors: Santhosh Kumar R, Gunavathi S
Cited by: 0
References: 13

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

Specifically, in environments where there's a need for recap systems in real-time, as live sports, breaking news, or emergency broadcasts, there is an increasing need to develop systems that can create summaries from a continuous text stream. In the typical formulations, transformer-based models rely on sliding windows over the input and treat every part of the input equally; however, this approach can backfire by absorbing irrelevant content, resulting in wasted memory and, worse yet, degraded performance. To enhance both performance and focus, we present a memory-efficient procedure based on an Event Compression Memory (ECM) framework. In this manner, the model learns to extract only the most essential event-focused information by employing an in-house developed score that considers the semantic importance of the fragments. Compressed representations of informative texts are retained in a running window, enabling the model to generate concise, context-sensitive summaries as new data becomes available. The system was implemented in Visual Studio Code and compiled as a lightweight text model. It was then trained and tested on a CPU-based Asus TUF Gaming A15, which features an AMD Ryzen 7 4800H processor with 16 GB of RAM, without utilizing GPU acceleration. The method achieved high performance, with an ROUGE-L F1 score of 54.6, 89.2% event recall, and an average latency of 238 ms per chunk. The model is also lightweight enough to run in real time on edge or mobile devices. Future extensions could include multi-lingual adaptation, multimodal integration, and on-device deployment for field applications.

Fields of Study & Subjects

Topic Modeling Advanced Text Analysis Techniques Natural Language Processing Techniques

Publication details

Journal / source
International Conference on Cognitive Systems and Computer Interaction (ICoSCI)
Pages
1–5
Publication date
15 January 2026
Language
EN