Building NLP Systems: Pipelines, Models, and Deployment

NLP projects often start with a misconception: "We'll take BERT and it will just work." A month later, latency is too high for production, the model weighs 1.5 GB, and F1 on Russian text is 0.6. We've seen dozens of such projects. The problem isn't the model—it's the lack of a systematic approach to the pipeline. We build production-ready NLP systems for Russian language that actually work in production: with data drift control, metric monitoring, and architecture selection driven by the task, not by the trend. In NLP system development, we combine rigorous Russian morphology analysis with efficient model selection to build cost-effective solutions. Processing natural language in a Russian context requires accounting for morphology, choosing the right pipeline, and applying MLOps practices.

Problems We Solve

• Russian morphology. The word "разработка" has 12 forms. Without lemmatization, TF-IDF loses 40% of meaning. We use pymorphy3 or natasha—they provide lemmas with >95% accuracy for technical texts. pymorphy3 documentation confirms 97% accuracy for literary text.
• Data drift. A month after deployment, the token distribution shifts. We automatically detect drift and trigger a retraining cycle. Without this, F1 drops 10–15% per quarter.
• Architecture choice. 80% of classification tasks are solved by Logistic Regression + TF-IDF with F1 0.92–0.95. Fine-tuning BERT is needed only when data is scarce (<5k examples) or semantic complexity is high (sarcasm, context dependency).

How We Do It: A Case from Our Practice

A fintech startup client needed to classify customer inquiries into 12 categories (complaint, return, consultation). Data: 50k labeled messages. Our approach:

• Analysis: class imbalance (3 classes accounted for 70% of samples).
• Prototype: FastText + TF-IDF. F1 = 0.91. Inference time 2 ms on CPU.
• Comparison: fine-tune BERT-base: F1 = 0.93, but latency 150 ms on GPU and 20× higher inference cost. FastText outperformed BERT in speed by 75× with comparable quality. Moreover, Logistic Regression + TF-IDF is 10× cheaper than BERT with similar accuracy.
• Result: we used FastText, added rule-based correction for rare classes. F1 = 0.93, deployed on 2 CPUs, reducing monthly infrastructure costs from $3000 to $300 for 1M predictions.

Lesson: lightweight solution + smart rules often beat a heavy transformer.

How to Choose NLP Model for Your Task

Why Morphology Is the Main Pain of Russian NLP

In English, tokenization is trivial: split by spaces. In Russian, "разработанный" and "разработана" are distinct tokens that don't look alike. Without lemmatization, the model cannot generalize. Transformers like BERT require careful tokenization; using a SentencePiece tokenizer helps but morphological analysis is still beneficial. We use pymorphy3, which gives lemmas with 97% accuracy for literary text and 93% for technical text. For NER, we use natasha, which considers context and outputs BIO-format tags. Russian morphology analysis is a mandatory step in any NLP pipeline. We also consider precision-recall trade-off and AUC-ROC when evaluating models.

Framework Comparison for Russian

For production, spaCy is usually sufficient. DeBERTa is only needed when maximum quality is critical. We often use model distillation to reduce BERT size by 40% with minimal accuracy loss, and ONNX runtime for efficient CPU inference.

Our Process

1. Analytics — gather requirements, audit data, select metrics (F1, latency, cost).
2. Prototype — MVP in 1–2 weeks: pipeline with lightweight models, establish baseline.
3. Training — if needed: fine-tune transformers, augment data, distill models.
4. Deployment — Docker, FastAPI, Triton inference server (for GPU). CI/CD with data drift tests.
5. Monitoring — log metrics, set alerts when F1 drops >5%.

What's Included

• Pipeline code repository (Python, PyTorch/TensorFlow)
• Architecture and API documentation (OpenAPI)
• Configured CI/CD (GitHub Actions / GitLab CI)
• Monitoring stack (Prometheus + Grafana dashboard)
• Client team training (2–3 workshops)
• 3 months of post-deployment support

Estimated Timelines

• Prototype (basic pipeline): 1–2 weeks
• Production solution (single task): 3–5 weeks
• Comprehensive NLP platform (multiple tasks): 2–4 months

Pricing is determined after analysis—contact us to discuss your project.

Why Choose Us

• Proven track record in AI solutions, with 5+ years of NLP experience, 30+ NLP projects delivered (fintech, e-commerce, healthcare), and 10+ active clients
• Experience with OpenAI, Yandex GPT, Hugging Face
• Certified MLOps specialists (Kubeflow, MLflow) with 5+ years in the field

Get in touch for a free consultation on your project.

from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.linear_model import LogisticRegression
from pymorphy3 import MorphAnalyzer
import re

morph = MorphAnalyzer()

def preprocess(text):
    tokens = re.findall(r'[а-яё]+', text.lower())
    lemmas = [morph.parse(tok)[0].normal_form for tok in tokens]
    return ' '.join(lemmas)

# Example usage
vectorizer = TfidfVectorizer()
X_train = vectorizer.fit_transform([preprocess(t) for t in train_texts])
model = LogisticRegression().fit(X_train, train_labels)

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.