From Benchmark to Better Embeddings: Leveraging Synonym Substitution to Enhance Multimodal Models in Ukrainian

AbstractWe study the robustness of text–image retrieval for Ukrainian under synonym-substitution attacks (SSA). On Multi30K with OpenCLIP, we evaluate two SSA methods: dictionary-based and LLM-based, and find Ukrainian degrades far more than English (e.g., GPT-4o SSA drops HIT@1 from 32.1 → 10.9 vs. 41.6 → 30.4). We introduce a Hybrid method that filters dictionary candidates with an LLM to preserve sense and grammar, yielding higher-quality perturbations (Ukrainian HIT@1 16.8 vs. 7.6/10.9). To mitigate this problem, we propose synonym-augmented fine-tuning, injecting one-word substitutions into training; it boosts robustness (Hybrid 28.1, GPT-4o 25.1) without harming original performance. This is the first systematic SSA evaluation for Ukrainian multimodal retrieval and a practical recipe for improving models in low-resource, morphologically rich languages. We release code, prompts, and trained checkpoints at https://github.com/YuriiLaba/UA-B2BE.