Dual Cross-Lingual Alignment for Multilingual Dialogue Generation

Existing pre-training methods for dialogue generation are typically limited to a single language, while multilingual pre-training, though effective for understanding tasks, struggles to transfer directly to generation tasks. This work identifies the core challenge of multilingual dialogue generation as achieving simultaneous cross-lingual representation and generation alignment. To address this, we propose two novel pre-training tasks: Cross-lingual Dialogue Generation (cDialog) and Masked Parallel Contexts Dialogue Generation (pDialog). cDialog enforces semantic alignment across languages by constructing cross-lingual context-response pairs (e.g., generating an English response from a Chinese context), while pDialog strengthens representation consistency by masking and recovering fragments of parallel contexts. Experiments on a Chinese-English parallel dialogue dataset demonstrate that our approach significantly improves performance in both monolingual and multilingual dialogue generation, achieving state-of-the-art results in BLEU, DISTINCT, coherence, and knowledge accuracy. Furthermore, the method exhibits enhanced robustness in low-resource scenarios, effectively improving generation quality for under-resourced languages through cross-lingual transfer.