Automate UI Localization with Context-Aware AI

Automate UI Localization with Context-Aware AI

When your application scales to 10+ languages, manual localization becomes a nightmare: strings lose context, placeholders break, plural forms are ignored. We build an AI system that integrates into your CI/CD pipeline and translates interface strings with full awareness of context: element type, maximum length, neighboring screens, and state.

On one project with 50,000 strings, manual translation took three weeks and contained 12% placeholder errors. After implementing the AI pipeline, time dropped to 2 hours, and context-aware translation accuracy reached 94% with p99 latency 3.2 seconds. The system automatically checks length limits, ICU MessageFormat correctness, and placeholder integrity.

Problems We Solve

Length constraints: A button with text "Save" (4 chars) becomes "Сохранить" (9 chars). If the UI element has a fixed width, the translation must fit. Our system checks character limits and warns of overruns before commit.

Element context: The word "Back" in a navigation button translates to "Назад", but in "back to top" it's different. The AI receives metadata: element type, parent screen, neighboring strings, and a screenshot with highlighting.

Placeholders: A string like "Hello, {name}!" must keep the placeholder format in translation. We escape placeholders before translation and validate after.

Plural forms: Russian uses 4 declension forms. ICU MessageFormat is a mandatory requirement. The AI generates correct forms based on a sample number.

Why AI Translation Outperforms Manual for UI?

Manual translation takes days and often loses context for short strings (e.g., "Save" translated as "Спасать" instead of "Сохранить"). AI with prompt engineering and few-shot examples delivers consistent results in seconds. In our test on 15,000 strings, context-aware translation accuracy was 94%, with p99 latency 3.2 seconds. That's 20x faster than a team of 3 translators.

How to Automate Localization in GitHub Actions?

Integration with TMS (Crowdin, Phrase) is standard:

1. GitHub Action triggers on push with new or changed strings.
2. Strings are uploaded to Crowdin via API with metadata (screenshot, element type).
3. AI translation runs: the model gets context from vector DB (ChromaDB) and translation history.
4. Strings with confidence > 0.9 are auto-approved; the rest go to review queue.
5. After approval, a PR is created with updated .json or .arb files.

Example Code: Translation Function with Context

async def translate_ui_string(
    key: str,
    source: str,
    target_lang: str,
    context: UIStringContext
) -> UITranslation:

    prompt = f"""Translate the UI string to {target_lang}.

String: {source}
Element type: {context.element_type}
Screen: {context.screen_name}
Max characters: {context.max_length or 'no limit'}
Context: {context.description or 'not specified'}

Requirements:
- Fit within the character limit
- Preserve all placeholders {{...}} unchanged
- Style: {context.style_guide}"""

    translation = await llm.generate_async(prompt, max_tokens=200)

    # Validation
    if context.max_length and len(translation) > context.max_length:
        translation = shorten_translation(translation, context.max_length)

    placeholders_ok = verify_placeholders(source, translation)

    return UITranslation(
        key=key,
        translation=translation,
        placeholders_valid=placeholders_ok,
        length_ok=not context.max_length or len(translation) <= context.max_length,
        confidence=estimate_confidence(source, translation, context)
    )

Stack and Experience

We use GPT-4o and Claude 3.5 for translation, LangChain for orchestration, ChromaDB for context storage. Deployment uses Docker + Kubernetes with Triton Inference Server. We have 5+ years of experience in AI localization, with over 15 projects for products with 1M+ users. We guarantee 100% placeholder accuracy and length limit compliance.

What's Included in the Deliverable

• Fully automated translation pipeline integrated into your CI/CD.
• Vector context database for continuous quality improvement.
• API and model documentation.
• Team training (2 hours).
• 2 weeks of technical support.

Our Process

1. Analysis – audit of current strings, identification of context issues.
2. Design – architecture selection, LLM tuning, TMS configuration.
3. Implementation – pipeline development, testing on 200 strings.
4. Testing – A/B comparison with manual translation, validation across all languages.
5. Deployment – production install, monitoring, prompt fine-tuning.

Timeline and How to Start

Project estimating takes 2 to 4 weeks depending on string volume and number of languages. Contact us for a free audit of your localization files. Order a consultation on UI translation automation.

According to our project data, AI localization reduces time-to-market for a new language from 2 weeks to 2 days.

NLP Development: Text Classification, NER, Embeddings, and Information Extraction

We often receive a task: process 50,000 support tickets — currently all manual. Dataset — 3,000 labeled examples, 12 categories, imbalance: one category occupies 40% of the sample, three at 1-2% each. Baseline accuracy — 78%. Sounds decent until you look at recall for rare classes: 0.31, 0.44, 0.28. These classes — complaints and churn threats — are most important to the business.

This is a typical NLP development project. The problem is not the algorithm but that accuracy is the wrong metric. Our experience across 30+ projects shows: we start by analyzing business metrics and only then choose the model.

Why accuracy is not the right metric for rare classes?

Accuracy ignores imbalance. If the "churn" class appears in 2% of cases, the model can predict "all good" and get 98% accuracy — but the business loses clients. Solution: F1 macro (averaged over all classes) or weighted F1. For NER — strict entity F1 (exact matches only). We guarantee: after choosing the correct metric, model quality becomes measurable and predictable.

Text Classification: From BERT to Distillation

BERT-like models are the standard for classification. ruBERT-base or ruBERT-large from DeepPavlov for Russian. multilingual-e5-large — for multiple languages in one pipeline. XLM-RoBERTa-large — a strong multilingual backbone.

Fine-tuning for classification: add a classification head on top of the [CLS] token, train for 3-5 epochs with lr=2e-5, weight decay=0.01. For imbalance — weighted CrossEntropyLoss or focal loss with gamma=2.0. Contact us — we will show a code snippet.

Imbalance case study. Dataset — 3,000 examples, imbalance 1:20. Solution: class_weight via sklearn + CrossEntropyLoss. Additionally — augmentation of rare classes via backtranslation (ru→en→ru through MarianMT). Recall for rare classes rose from 0.31 to 0.67 with a slight drop in accuracy (76%→74%). Full NLP development end-to-end took 3 weeks.

Distillation for production. BERT-large gives F1 0.89, but inference on CPU — 180ms. Distillation into DistilBERT or ruBERT-tiny2 reduces latency to 25ms with F1 0.84. Export to ONNX Runtime provides an additional 1.5-2x speedup. DistilBERT achieves 7x lower latency than BERT-large with only a 5% drop in macro F1 – a typical production trade-off.

How to choose between BERT and LLM for your task?

For most classification and extraction tasks, BERT-sized models offer the best trade-off between cost and performance. Shift to LLMs only when the task demands generation, complex reasoning, or zero-shot generalization.

NER: Named Entity Recognition

NER — extracting persons, organizations, locations, dates, amounts, document numbers. For general categories (PER, ORG, LOC), pre-trained models work well. For specialized ones (medical terms, legal concepts) — fine-tuning is needed.

Data annotation. The main cost of an NER project. For a quality model — 500-2,000 labeled sentences per entity type. Tools: Label Studio (open source) or Prodigy (by spaCy creators). IOB2 format — standard.

Architecture. Token classification on top of BERT: each token gets a label (B-PER, I-PER, O). spaCy 3.x with transformer pipeline — a convenient production choice.

Nested entities. Standard IOB models cannot handle nested entities (organization inside an address). For such tasks — span-based NER: SpanBERT or SpERT. More complex but correct.

Post-processing is mandatory. The model predicts tokens — normalized entities are needed. Date — dateparser. Amounts — regex + validation. Names — deduplication via rapidfuzz. Included in our standard delivery.

Sentiment Analysis and Opinion Mining

Binary classification positive/negative works out of the box with BERT. Complexity — aspect-based sentiment analysis (ABSA): "the restaurant has good food but terrible service." For ABSA: aspect extraction (NER) + sentiment per aspect. Joint models BERT-for-ABSA — quality on Russian data is lower due to dataset scarcity. RuSentiment, SentiRuEval — main resources.

For production with simple positive/negative/neutral: distil models are enough. Three classes, balanced dataset, 2,000+ examples — F1 macro 0.82-0.87 in 1-2 days.

Text Summarization

Extractive summarization (select sentences) — TextRank or BM25 without training. Fast, no hallucinations. Good for long documents.

Abstractive (generates new text) — seq2seq: mT5, mBART, FRED-T5, ruT5-large. For production via LLM API (GPT-4, Claude) — often the best cost/quality/speed trade-off.

Embeddings: Vector Representations of Text

Embeddings are the foundation of semantic search, deduplication, clustering, RAG. Quality critically affects downstream tasks.

Models. E5-large-v2, BGE-M3, multilingual-e5-large — strong multilingual embedders. sentence-transformers/paraphrase-multilingual-mpnet-base-v2 — fast option. For Russian: ru-en-RoSBERTa (Skoltech) performs well on semantic textual similarity.

Embedding quality evaluation uses the MTEB benchmark as standard. But top results on MTEB don't guarantee success on a domain dataset — we build domain-specific eval.

Fine-tuning embeddings. If standard models don't give the required Recall@k — contrastive learning on domain pairs with MultipleNegativesRankingLoss. How to perform this for domain data:

1. Collect 500–2,000 semantically similar pairs from your domain.
2. Apply MultipleNegativesRankingLoss with a batch size of 32–64.
3. Train for 1–3 epochs using AdamW (lr=2e-5).
4. Evaluate Recall@k on a held-out domain test set.

This approach yields a 5–15% improvement in Recall@k in practice.

Dimensionality and storage. E5-large: 1024 dim, float32 — 4KB per vector. For 10M documents — 40GB. Quantization int8 reduces to 10GB. FAISS IVF_PQ — more compact but with losses. Included in our deployment recommendations.

Information Extraction

Structured extraction is a frequent task. Examples: key contract terms, technical characteristics, dates and amounts from invoices.

1. Regex + rule-based. For INN, OGRN, amounts, dates — more reliable than neural networks. No data required.
2. NER + post-processing. For variable formats.
3. LLM with structured output. GPT‑4 / Claude with JSON schema — for complex documents. Cost: minimal per document. For 10k+ documents/day — we calculate the economics.

We guarantee a hybrid: regex/NER for typical fields + LLM for edge cases. Our guarantee is backed by years of production experience and more than 30 projects.

Work Stages

Prototype on existing data — 1-3 weeks. Production system with CI/CD — 1.5–2.5 months. Cost is calculated individually — get a consultation for a project estimate.

What's Included

• Model and pipeline architecture documentation
• Access to the model via REST API (FastAPI + ONNX)
• Client team training (2-hour webinar + Q&A)
• Accuracy guarantee on the agreed test set
• Months of post-delivery support (bug fixes, adaptation to new data)

Our Experience

Years of NLP projects from classification to RAG systems. The team includes ML engineers experienced with Hugging Face, spaCy, LangChain, MLOps. We use vLLM, Kubeflow, Weights & Biases — a production stack, not toys. Contact us to evaluate your NLP project within two days — request a free consultation on your text processing pipeline.