EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Exploring Embedding Interpretability by Correspondences Between Topic Models and Text Embeddings

Meng Yuan, Lida Rashidi, Justin Zobel

Information Retrieval Research · 2025

DOI: 10.54195/irrj.23703 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Information Retrieval Research, 2025. Verfasst von Meng Yuan und 2 weiteren Autor:innen. Die Arbeit nennt 63 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Text embeddings have become essential for representing documents in Information Retrieval (IR), yet their high-dimensional nature often limits interpretability. To bridge this gap, we introduce a novel mapping framework that aligns embedding dimensions with topics derived from both probabilistic and neural models. Using three standard collections and three embedding methods, we demonstrate that embedding features consistently map to a subset of coherent topics, even as the total number of topics varies. We further quantify this correspondence with a Mean Mapping Specificity Improvement Rate, showing that mapped topics exhibit significantly higher specificity than the global topic set if the embedding dimensions are set properly. A stability analysis over varying embedding dimensions confirms the stability of the mapping across random feature samples. Our contributions are three-fold: A general-purpose mapping method that visualizes and formalizes correspondences between embedding features and topic representations; Empirical evidence that text embeddings and topic models are not independent descriptors but can mutually validate each other’s semantic structures; A numeric indicator that captures the degree to which embedding features correspond to high-quality topics, providing a new tool for evaluating embedding interpretability and guiding dimensionality reduction choices. These findings suggest that topic-embedding mapping can serve both as a diagnostic for embedding quality and as a means to visualise embedding dimensions more human-interpretable, advancing the practice of collection description in IR.

Zitieren

APA

Yuan, M., Rashidi, L., & Zobel, J. (2025). Exploring Embedding Interpretability by Correspondences Between Topic Models and Text Embeddings. Information Retrieval Research. https://doi.org/10.54195/irrj.23703

Vancouver

Yuan M, Rashidi L, Zobel J. Exploring Embedding Interpretability by Correspondences Between Topic Models and Text Embeddings. Information Retrieval Research. 2025. doi:10.54195/irrj.23703

BibTeX
@article{Yuan2025,
  author = {Yuan, Meng and Rashidi, Lida and Zobel, Justin},
  title = {Exploring Embedding Interpretability by Correspondences Between Topic Models and Text Embeddings},
  journal = {Information Retrieval Research},
  year = {2025},
  doi = {10.54195/irrj.23703}
}
Kostenlos starten →