Context plays a crucial role in translation, enhancing both accuracy and fluency. With the advancement of machine translation (MT), the concept of context is now considered across an increasingly broader range of phenomena. Despite its importance, however, systematic definitions of context provided by communication studies and translation studies remain fragmented, and the concept of context remains elusive in MT research. To the best of our knowledge, no dataset currently exists that comprehensively evaluates MT’s sensitivity to context. In this study, we propose a systematic taxonomy of context and introduce Context-8, an evaluation dataset designed to assess context sensitivity in MT for English-to-Japanese translation. The initial release includes 130 groups comprising 533 English-to-Japanese translation examples, each requiring different context categories to produce accurate and fluent translations. The data are taken from both hand-crafted and online materials. We release Context-8 to support the evaluation and benchmarking of MT systems with respect to context sensitivity.

Industrial recommendation training increasingly relies on decentralized multi-GPU infrastructure, yet sparse embedding tables remain a major bottleneck due to significant storage, lookup, and communication overhead. Most existing works typically optimize only one of these bottlenecks or are designed for parameter server architectures, making them less effective in industrial decentralized training stacks. To tackle these challenges, we propose MixEmb, a practical framework for efficient decentralized embedding training through importance-aware quantization and caching. We first design a novel importance metric, i.e., embedding score, which can distinguish the contribution of embedding vectors based on spatial heterogeneity and temporal dynamics. Guided by embedding scores, MixEmb employs mixed precision quantization and hierarchical caching at fine-grained vector level. Besides, our key finding is that quantization and caching decisions are tightly coupled. Instead of naive combination, we delicately integrate them into a unified training framework through coupled co-design. A theoretical convergence analysis is also provided for MixEmb. We implement and evaluate MixEmb in real-world training scenarios. Across three workloads, MixEmb attains up to 1.55× speedup and 3.54× compression ratio compared to state-of-the-art methods.