We introduces MuCAL (Multilingual Contrastive Alignment Learning), a multilingual model for aligning knowledge graphs (KG) with text. MuCAL enables robust KG–text retrieval across languages and difficulty levels, and is used to automatically generate preference data by ranking outputs from multiple LLMs. With this data, we further apply Direct Preference Optimization (DPO) to directly align generation with KG semantics, avoiding reward modeling. Experiments on KG-to-English text generation show that MuCAL-based similarity signals improve DPO training and achieve better out-of-domain generalization than standard instruction tuning, demonstrating MuCAL’s effectiveness for preference learning in KG-to-text tasks.

In this work, We investigate multilingual Knowledge Graph (KG)-to-Text generation across 9 languages, covering both high-resource (English, Chinese, French, Spanish, Russian) and low-resource languages (Breton, Irish, Maltese, Welsh). We construct silver multilingual training data and new gold out-of-domain test sets for the high-resource languages, and use these along with existing in-domain test sets to evaluate three approaches: (1) NLG+MT—a KG-to-English model followed by machine translation, (2) FTMT—fine-tuning multilingual MT models on the silver data, and (3) FewShot—LLM prompting with different strategies. We find that the best prompting strategy consistently outperforms the other methods across all nine languages, and we provide an analysis of performance differences between high- and low-resource languages as well as in- vs out-of-domain data.