Toward a Geometric View of Multilingual Intent Detection: Linguistic Packaging vs. Semantic Resolution

Resumen

Trustworthy intent detection is hindered by lack of interpretability in deep learning models. While recent research indicates Recurrent Neural Networks (RNNs) solve high-cardinality intent detection tasks via low-dimensional manifolds quantified by a Functional Dimensionality (FD), it remains unclear if these findings represent general principles or linguistic artifacts. We investigate FD stability using MASSIVE dataset across a diverse set of languages: English, Spanish, Finnish, and Turkish. Our results suggest a dual geometric regularity. First, we observe inter-lingual stability within a fixed embedding paradigm, RNNs converge to statistically stable Semantic Resolution subspaces despite differences in Linguistic Packaging. Second, we observe inter-paradigm robustness of distillation process: while transitioning from static to contextualized embeddings induces massive expansion of geometric volume (ID), Distillation Efficiency (Omega=FD/ID) remains statistically stable. These findings provide preliminary evidence for a geometric view of multilingual intent detection, where semantic resolution appears to be governed by task-mandated constraints rather than Linguistic Packaging or representational schema.