Joint Cultural and Grammatical Correction of Japanese Translations via a GAN-BiLSTM Framework with Policy Gradient Optimization

Abstract

Japanese translations often exhibit grammatical deviations and cultural inconsistencies. This study proposes an automatic correction framework that integrates a Generative Adversarial Network with a Bi-LSTM encoder–decoder augmented by embedded cultural vectors to generate correction candidates. A convolutional discriminator jointly evaluates grammaticality and cultural appropriateness under a Wasserstein adversarial objective, while policy gradient optimization refines the generator using discriminator-based rewards. Semantic consistency is preserved through back-translation constraints. Experiments on 60,000 culturally annotated Japanese–Chinese sentence pairs from news, business, and academic domains, split into training, validation, and testing sets, compare the proposed method with RNN, Transformer, and GAN-only baselines. The model achieves 86.2 percent grammatical accuracy, a cultural alignment score of 4.52, a fluency score of 4.10, and a perplexity of 45.6, consistently outperforming baselines across sentence lengths and cultural load levels. The results demonstrate unified optimization of grammatical correction, cultural alignment, and semantic preservation in Japanese translation post-editing.