@techreport{oai:ipsj.ixsq.nii.ac.jp:00232718, author = {内山, 俊郎 and 甫喜本, 司 and Toshio, Uchiyama and Tsukasa, Hokimoto}, issue = {27}, month = {Feb}, note = {潜在的ディリクレ配分法は,文書などのデータ解析に有用なトピックモデルとして知られる.階層的トピックモデルは,トピックに階層構造(木構造)があるとする方法であり,トピック数が異なる各段階でのトピックを知ることができ,文書集合に潜在するトピックを大局から詳細に至るまで網羅的に把握することが可能である.しかし,木構造はトピックを「1対多」で詳細なトピックへと分割する制約であるため,トピック数増加の影響が 1つのトピックに閉じている.一般に,トピック数を増やすと複数のトピックに影響を与えると考えられ,この意味において可能性を狭めた分析になっている.そこで本研究では,この問題を解決するため,多様な解をトピック数横断で生成し,「多様な解の分析手法」を利用して互いに類似する解を結びつけるというアプローチにより,トピック数増加の影響を多面的に捉える分析方法の確立を目指す.提案手法と実験結果を示す., The Latent Dirichlet analysis is known as a topic model that is useful for analyzing data such as documents. The hierarchical topic model is a method that assumes that there is a hierarchical structure (tree structure) of topics, and it is possible to know the topics at each stage where the number of topics differs, and to comprehensively understand the potential topics in a set of documents from the big picture to the details. However, because the tree structure is a constraint that divides topics into detailed topics on a "one-to-many" basis, the effect of increasing the number of topics is closed to a single topic. In general, increasing the number of topics is considered to affect multiple topics, and in this sense, the analysis narrows the possibilities. To solve this problem, this study aims to establish an analysis method that captures the impact of an increase in the number of topics from multiple perspectives by generating a variety of solutions across the number of topics and linking solutions that are similar to each other using the "analysis method for a diversity of solutions" approach. A proposal of the analysis method and experimental results are presented.}, title = {トピックモデルにおけるトピック数の異なる多様な解の可視化}, year = {2024} }