[4G2-GS-6] Language media processing:

[4G2-GS-6-05] Visualization of Text Structure Using Correspondence Analysis for Summary Evaluation

〇Soichiro Morishita¹ (1. CyberAgent, Inc.)

Keywords:LLM, text summarization, visualization

With the advancements in large language models, high-precision text summarization has become achievable without the need for dedicated summarization models, making automated summarization an increasingly accessible technology. At the same time, there is a growing demand for evaluating the validity of automatically generated summaries, emphasizing the importance of evaluations conducted by non-experts. Against this backdrop, this study proposes a framework for visualizing the correspondence between source texts and summaries to make summary evaluation more accessible to non-experts. Specifically, the proposed method involves dividing the text into chunks, calculating the similarity between chunks using sentence embeddings, and constructing a distance matrix. Correspondence analysis is then applied to this distance matrix to perform dimensionality reduction and map the correspondence relationships. Furthermore, the procedure makes it possible to explicitly reveal repetitions and topic development, thereby visualizing the structure of the text. This approach provides a valuable tool for enhancing the objectivity of summary evaluation and supporting efficient evaluations by non-experts.