From Technology to Experience: Vivo Machine Translation Deployment Practice

The article begins with an overview of the growing demand for mobile translation, noting that users spend 6‑7 hours daily on phones and desire a translation workflow that can be completed in two steps to improve experience.

1. Understanding Business Needs – Before building translation capabilities, it is essential to identify core user groups and high‑frequency scenarios, such as watching foreign media, chatting with overseas friends, and reading international news, and to tailor translation services accordingly.

2. Algorithm Beyond NMT – While Neural Machine Translation (NMT) models form the core, the author highlights mature open‑source frameworks, the importance of preprocessing (tokenization, BPE), and additional modules like language detection, tokenization, truecasing, detokenization, and sentence splitting.

3. Language Detection – Discusses the multi‑class classification problem of detecting over 100 languages, handling data sparsity, short texts, mixed‑language sentences, and similar language families.

4. Text Pre‑processing – Covers Chinese word segmentation, English tokenization, handling punctuation, truecasing/detruecasing, and recasing for low‑resource scenarios, emphasizing the correct order of operations.

5. Sentence and Mixed‑Language Splitting – Describes splitting long sentences and separating mixed‑language segments to avoid corrupting native language fragments during translation.

6. Model Domain Adaptation – Presents two methods: fine‑tuning with domain‑specific data and adding adapter layers to encoders/decoders, enabling the model to generate more domain‑appropriate translations.

7. Robustness Enhancements – Lists techniques such as random ground‑truth word replacement, label smoothing, adversarial sample generation, and random data perturbations to improve model stability.

8. Data Quality and Quantity – Emphasizes that data determines the upper bound of translation quality; includes analysis of dataset length distribution, domain distribution, and cleaning steps for both monolingual and bilingual corpora.

9. Evaluation – Stresses the need for both objective metrics (multiple BLEU test sets) and subjective human evaluation, especially when BLEU plateaus, to assess fluency, adequacy, and style.

10. Engineering Practices – Covers performance optimization (sentence splitting, caching, model compression, batch inference), rapid online bug fixing via retrieval libraries and translation intervention, and establishing a closed‑loop iteration pipeline that incorporates logs, error analysis, and continuous model updates.

The conclusion reiterates that technology serves product goals, urging practitioners to keep user experience at the forefront when optimizing machine translation systems.