Abstract
Semantic non-negative matrix factorization (Semantic NMF) is an unsupervised method for automatic term extraction that jointly factorizes a document-candidate matrix and a candidate-seed similarity matrix, thereby incorporating semantic relations between candidate terms and representative domain “seed” terms. This guidance steers the factorization toward terminology-relevant dimensions. The first application of this approach to term extraction relied on static word2vec embeddings, which limited effectiveness because such representations assign a single, context-independent vector to each word, cannot represent new terminology, and exhibit geometric deficiencies that reduce discriminative power. In this study, we evaluate Semantic NMF on the ACTER benchmark using five modern, large-scale multilingual embedding models. Results show that higher-quality embeddings enable Semantic NMF to consistently outperform a standard NMF baseline across all ACTER languages and domains, with a best-case$\mathbf{F 1}$score rising from 30.3% to 53.6%. Thus, these results establish Semantic NMF with modern multilingual embeddings as a robust, interpretable, cross-lingual solution for automatic term extraction.