Exsumm-VIZ: Visual Interpretation of Attention Model In Text Summarizations
Abdulhaq Adetunji Salako, Jiansu Pu, Zhang Jinlun, LI Ting LIU Guanqun, Bless Lord Y. Agbley · 2021 18th International Computer Conference on Wavelet Active Media Technology and Information Processing (ICCWAMTIP) · 2021
Deep learning has enhanced the state of the art in several domains, including text summarizing, with pointer generator summarization one of the dominant approaches in recent years. However, several obstacles remain for the summarizing model, such as over-and-under summarization and processing large phrases, requiring human evaluation. Furthermore, summarization raises issues of explainability, and user interaction, compelling the development of better analytics systems. To address these limitations, we introduce EXSUMM, a tool for visualizing analysis of attention weights, data, and evaluation metrics associated with text summarization. Through its linguistic and semantic filtering, the tools help users understand what their models are “focusing on.” such as abstractiveness and factual consistency. We validated our visual system's effectiveness through various use cases and user studies and discovered a favorite for EXSUMM over typical text-based summarization systems, particularly for lengthy texts.