Problem: BLEU misses style and terminology errors

A translator spent two hours post-editing a translation that the system rated 0.95 BLEU. The client rejected the project due to style guide violations and incorrect terminology. BLEU compares against a reference, but the reference often ignores context, tone, and domain-specific vocabulary. We solve this with Quality Estimation (QE) — an AI system that evaluates translations without a reference, like a human reviewer. Budget savings on review can reach 60%.

How Quality Estimation Without a Reference Translation Works

Quality Estimation analyzes the source text and translation, producing a score of 0–1 at the segment, word, and document levels. Segment-level scores indicate which sentences need review. Word-level QE tags each word as OK or BAD — the reviewer sees errors immediately. Document-level QE evaluates coherence, terminology consistency, and stylistic unity.

How QE Saves 40–60% of Reviewer Time

Assume you process 10,000 segments per day. Without QE, a reviewer checks every segment. With QE, only segments with score < 0.7 (typically 20–30%). At a threshold of 0.9 for auto-publishing, 10–15% of segments skip review entirely. We implemented such a pipeline in a fintech application localization project: the reviewer handled 3,000 segments instead of 10,000, and style errors dropped by 80%. This resulted in estimated annual savings of $120,000 in reviewer costs.

Why CometKiwi Is Better Than BLEU for QE

CometKiwi (Unbabel/wmt22-cometkiwi-da) is a transformer-based model trained on thousands of human judgments. It outperforms BLEU and traditional metrics in correlation with human evaluation. Here’s a comparison of key metrics:

CometKiwi needs no reference and provides word-level errors through MQM taxonomy.

Error Types Classified by QE (MQM Taxonomy)

We use the MQM taxonomy, dividing errors into four classes:

• Accuracy — mistranslation, omissions, additions.
• Fluency — grammar, spelling, punctuation.
• Terminology — glossary violations, inconsistent term usage.
• Style — tone of voice mismatches, stylistic inconsistencies.

Common Mistakes When Implementing QE

Typical errors include: using a model without fine-tuning for the language pair (lower precision), choosing the wrong score threshold (missing errors or overburdening the reviewer), ignoring word-level QE (losing context for individual words), and not integrating MQM (difficult to improve the process).

How We Integrate QE Into Your Pipeline

1. Audit the current process: measure volume, latency, and existing metrics.
2. Model selection: CometKiwi, OpenKiwi, or fine-tuning for your language pair.
3. Integration: REST API or gRPC — wrapped in a microservice.
4. MQM taxonomy setup: connect error type classification via an LLM (GPT-4o or LLaMA 3).
5. Testing: measure precision/recall on your dataset.
6. Deployment: Kubernetes + GPU (T4 or A10).

from transformers import AutoModelForSequenceClassification, AutoTokenizer

class QualityEstimator:
    def __init__(self, model_name: str = "Unbabel/wmt22-cometkiwi-da"):
        self.model = load_comet_model(model_name)

    def estimate_segment(self, source: str, hypothesis: str) -> QEScore:
        score = self.model.predict(
            [{"src": source, "mt": hypothesis}],
            batch_size=8
        ).scores[0]

        return QEScore(
            score=score,              # 0-1, where 1 = excellent quality
            requires_review=score < 0.7,
            error_probability=1 - score
        )

    def estimate_batch(
        self,
        segments: list[tuple[str, str]]
    ) -> list[QEScore]:
        data = [{"src": src, "mt": mt} for src, mt in segments]
        scores = self.model.predict(data, batch_size=32).scores
        return [QEScore(score=s, requires_review=s < 0.7) for s in scores]

QE Model Comparison

What's Included in the Work

• Audit of the current translation pipeline with metric measurement.
• Selection and configuration of a QE model (CometKiwi, OpenKiwi, fine-tuning).
• Integration via REST API or gRPC with documentation.
• Error classification training for your MQM taxonomy.
• Deployment on infrastructure (Kubernetes, GPU).
• Team training and 1 month of support.

Timeline and Pricing

Timelines range from 2 to 6 weeks depending on complexity (volume, number of languages, need for fine-tuning). Pricing is calculated individually. For a typical project with 5 language pairs and 50,000 segments, pricing starts at $15,000. Get a consultation — send a description of your current pipeline, and we’ll evaluate your project within 2 days.

Why Work With Us

• Over 5 years of experience in NLP and machine translation.
• Completed 15+ quality evaluation projects for fintech, e-commerce, and software localization.
• Certified engineers in PyTorch and MLOps.
• We guarantee reviewer time savings of at least 40% or your money back.

Learn more about Quality Estimation on Wikipedia. Typical use cases include e-commerce product descriptions, financial reports, and software UI localization.

Contact us to get a consultation on implementing QE into your translation process. Order a project evaluation in 2 days — send a description of your current pipeline.

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.