Multilingual large language models (LLMs) are predominantly trained and evaluated within English-centric pipelines. However, the semantic consequences of English-language mediation at the level of textual representations remain poorly understood beyond surface-level similarity measures. This paper puts forward a metric-based approach to evaluating the cultural and semantic integrity of texts produced using multilingual large language models (LLMs), with a specific focus on low-resource languages.

We set forth a set of complementary embedding-based metrics designed to diagnose how English mediation reshapes textual semantic representations at multiple levels. Using English-mediated back-translation via an LLM as a controlled diagnostic probe, we compare a high-resource language (Russian) with a low-resource language (Lingala). Texts are embedded into a shared semantic space, and semantic integrity is assessed using three metrics: Semantic Self-Similarity (SSI), capturing local semantic recognizability; Neighborhood Preservation Score (NPS), measuring the stability of local semantic relations; and axis-based drift, quantifying directional semantic bias along an interpretable semantic opposition.

The results reveal a pronounced cross-linguistic asymmetry. Russian texts maintain high semantic self-similarity, indicating strong surface-level semantic preservation, but display only moderate neighborhood preservation, reflecting nontrivial structural reorganization. In contrast, Lingala texts show severe degradation in both semantic self-similarity and neighborhood preservation, indicating a collapse of relational semantic structure under English mediation. Additionally, Lingala – but not Russian – exhibits a small yet systematic directional drift along the examined semantic axis. What is of importance is that this directional bias is independent of structural instability, which is indicative of multiple, distinct mechanisms of English-centric effect.

These findings indicate that surface similarity metrics considerably underestimate semantic disruption, particularly for low-resource languages. The suggested framework provides a scalable diagnostic toolkit for assessing semantic integrity in multilingual LLM representations and is directly applicable to the analysis and evaluation of LLM-generated texts beyond translation-based scenarios. Although we are validating the framework using Russian and Lingala, the proposed metrics are intended for use with other low-resource languages and in multilingual settings.