Generalization of GLM improvements across languages, domains, and tasks

Confirm whether the improved text and graph reasoning capabilities attributed to Graph Language Models extend beyond English knowledge graphs to other languages, domains, and tasks.

Background

The evaluation in the paper is conducted on English commonsense (ConceptNet) and factual (Wikidata) subgraphs and associated text. While results indicate advantages of GLMs over LM and GNN baselines, the scope is limited to English.

The authors explicitly defer verifying whether GLM improvements in text and graph reasoning hold across different languages, domains, and task settings.

References

Confirming GLMs improved text and graph reasoning skills for different languages, domains and tasks is left for future work.

— Graph Language Models  (2401.07105 - Plenz et al., 2024) in Limitations

Although DARTopic shows consistent gains across general, biomedical, and legal benchmarks, our experiments are limited to English corpora. Because the framework relies on token-level semantic graph construction rather than domain-specific resources, it is potentially applicable to broader settings, but multilingual validation remains for future work.

— Domain-Agnostic Neural Topic Modeling with Contextual Token-Level Semantic Graph Representation  (2608.16269 - Seo et al., 17 Aug 2026) in Limitations section