EUVIMEDEUVIMED entdecken
Crossref

Sensitivity analysis of text vectorization techniques for failure analysis: a Latent Dirichlet Allocation and generalized variational autoencoder approach

Abbas Rammal, Kenneth Ezukwoke, Anis Hoayek, Mireille Batton-Hubert

The Journal of Supercomputing · 17.07.2026 · 82 (11)

DOI: 10.1007/s11227-026-08693-1 ↗

Kurz gefasst

Fachartikel in der Zeitschrift The Journal of Supercomputing, online erschienen am 17.07.2026. Verfasst von Abbas Rammal und 3 weiteren Autor:innen. Die Arbeit nennt 43 Literaturangaben.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Industrial Failure Reporting, Analysis, and Corrective Action System (FRACAS) databases contain large volumes of unstructured textual reports describing failure events, diagnostic investigations, and corrective actions. Extracting semantic knowledge from these heterogeneous industrial reports remains challenging due to high-dimensional sparse representations, domain-specific terminology, and contextual semantic variability. This study proposes a unified industrial Natural Language Processing framework integrating Latent Dirichlet Allocation (LDA) topic modeling, Variational Autoencoder (VAE)-based latent representation learning, and supervised classification for semantic failure-conclusion prediction. The proposed framework investigates the sensitivity of predictive performance with respect to different text vectorization strategies, namely TF-IDF, Word2Vec, and Doc2Vec. The experimental study was conducted on more than 12,000 industrial failure-analysis reports collected from a semiconductor FRACAS environment. After semantic topic extraction using LDA, vectorized failure-description data were transformed into compact nonlinear latent representations through a VAE architecture and subsequently classified using Linear Discriminant Analysis (LDA classifier), Naïve Bayes, Support Vector Machines, and Decision Trees. Performance evaluation was performed using repeated k-fold cross-validation, classification accuracy, and F1 score metrics. Experimental results demonstrate that contextual embedding representations significantly improve predictive semantic topic classification compared with sparse lexical representations. Among the evaluated methods, Word2Vec achieved the best overall performance. The combination of Word2Vec embeddings, VAE latent-space learning, and Linear Discriminant Analysis yielded the highest predictive performance, reaching 62.4% classification accuracy and an F1 score of 0.55 under fivefold cross-validation. The results further show that nonlinear latent-space learning improves semantic compactness and enhances discriminative capability for industrial textual data. The proposed framework provides a systematic methodology for evaluating semantic representation sensitivity in industrial NLP applications and demonstrates the importance of contextual semantic embeddings for predictive failure-analysis systems.

Zitieren

APA

Rammal, A., Ezukwoke, K., Hoayek, A., & Batton-Hubert, M. (2026). Sensitivity analysis of text vectorization techniques for failure analysis: a Latent Dirichlet Allocation and generalized variational autoencoder approach. The Journal of Supercomputing, 82(11). https://doi.org/10.1007/s11227-026-08693-1

Vancouver

Rammal A, Ezukwoke K, Hoayek A, Batton-Hubert M. Sensitivity analysis of text vectorization techniques for failure analysis: a Latent Dirichlet Allocation and generalized variational autoencoder approach. The Journal of Supercomputing. 2026;82(11). doi:10.1007/s11227-026-08693-1

BibTeX
@article{Rammal2026,
  author = {Rammal, Abbas and Ezukwoke, Kenneth and Hoayek, Anis and Batton-Hubert, Mireille},
  title = {Sensitivity analysis of text vectorization techniques for failure analysis: a Latent Dirichlet Allocation and generalized variational autoencoder approach},
  journal = {The Journal of Supercomputing},
  year = {2026},
  volume = {82},
  number = {11},
  doi = {10.1007/s11227-026-08693-1}
}
Kostenlos starten →