Domain Translation with Monolingual Lexical Distribution

Abstract

Neural machine translation (NMT) often demands a large amount of high-quality training data when adapting to a new domain with a carefully designed fine-tuning strategy. However, constructing a sufficient amount of parallel data for training poses challenges even for fine-tuning. This work proposes to fine-tune a generic NMT model using only the monolingual lexical distribution estimated from a small amount of in-domain data in the target language. Word frequency plays a critical role in analyzing the differences among corpora in various fields, e.g., psycholinguistic and language education, and our challenge lies in whether we can fit a model using the naive statistics collected from a target language domain in NMT. We leverage a variant of energy-based models (EBMs) based on Conditional Distributional Policy Gradients (CDPG) with a large number of EBMs to constrain the fine-tuning process with lexical distribution. We conduct experiments across four translation directions and four domain datasets, totaling 16 domain adaptation scenarios. The results demonstrate that our method enables robust domain shift while mitigating catastrophic forgetting, achieving effective domain adaptation using only a small amount of monolingual resources.