We participated in the constrained (open-weight) track of the WMT 2026 General MT Task for the English-to-Japanese and Simplified Chinese-to-Japanese directions. Our system builds on Marco-MT-Algharb, a strong open-weight translation model from the previous WMT General MT task. We further fine-tuned this model to generate more natural Japanese translations. To this end, we constructed a Context-Aware JApanese Linguistic Acceptability dataset (CAJALA). During annotation, human annotators were shown multiple paraphrased variants of a segment within its document-level context and asked to select the most natural one. We then back-translated the CAJALA dataset to create parallel data for Direct Preference Optimization (DPO). We also trained the model to perform fill-in-the-middle (FIM) translation, in which the model reconstructs target-side segments from their surrounding document. At inference, we use Minimum Bayes Risk (MBR) decoding, followed by selective FIM post-editing of low-quality segments.

This paper presents a novel hybrid generative/discriminative model of word segmentation based on nonparametric Bayesian methods. Unlike ordinary discriminative word segmentation which relies only on labeled data, our semi-supervised model also leverages a huge amounts of unlabeled text to automatically learn new “words”, and further constrains them by using a labeled data to segment non-standard texts such as those found in social networking services. Specifically, our hybrid model combines a discriminative classifier (CRF; Lafferty et al. (2001) and unsupervised word segmentation (NPYLM; Mochihashi et al. (2009)), with a transparent exchange of information between these two model structures within the semi-supervised framework (JESS-CM; Suzuki and Isozaki (2008)). We confirmed that it can appropriately segment non-standard texts like those in Twitter and Weibo and has nearly state-of-the-art accuracy on standard datasets in Japanese, Chinese, and Thai.