Developing an AI System for Automatic Product Categorization

Imagine you have 50,000 SKUs, with hundreds of new products added weekly from suppliers using different names, descriptions, and images. Manual categorization can't keep up—errors multiply, the content team is overwhelmed, and buyers can't find what they need. We solve this with an AI system that automatically categorizes products by analyzing text, images, and attributes. Our experience spans years in machine learning and numerous successful projects for retail and marketplaces. We offer a turnkey solution: from data audit to a ready-to-use API. Contact us for a preliminary assessment of your catalog.

According to Wikipedia, hierarchical classification reduces computational complexity by stepwise category refinement.

How automatic product categorization works

Product catalogs have a tree structure: "Clothing → Outerwear → Jackets." The classifier first determines the top level, then refines—this reduces complexity and improves accuracy. Example hierarchy:

Clothing and footwear
    ├── Men's clothing
    │   ├── Outerwear
    │   │   ├── Jackets
    │   │   └── Coats
    │   └── Pants
    └── Women's clothing

For each level, we train a separate model or use a unified hierarchical architecture. We choose the approach based on your catalog—from a simple BERT classifier to a retrieval-based system using embeddings. The hierarchical approach is 30% more efficient than flat classification for catalogs with depth greater than 3 levels.

Input data and features—developing the AI system

To ensure accurate model performance, we use as many available features as possible:

class ProductFeatures(BaseModel):
    title: str                    # "Men's winter Nike jacket blue"
    description: str | None       # full description
    attributes: dict              # characteristics: material, size, color
    images_url: list[str] | None  # for multimodal classification
    brand: str | None
    price: float | None           # price range hints at category
    supplier_category: str | None # supplier category (noisy but useful)

The multimodal approach (text + image) provides a 5–10% accuracy boost over text-only. This is confirmed in our projects: in one case (10,000-product catalog), accuracy rose from 89% to 96% after adding images.

Classifier implementation

The architecture choice depends on catalog size and availability of labeled data:

Example code for a hybrid approach:

def categorize_product(product: ProductFeatures) -> CategoryPrediction:
    text = f"{product.title}\n{product.description or ''}\n{format_attributes(product.attributes)}"

    # Fast classifier
    top_categories = fast_classifier.predict_top_k(text, k=5)

    if top_categories[0].score > 0.85:
        return top_categories[0]  # high confidence → immediate

    # Low confidence → LLM for refinement
    return llm_classify(product, top_categories)

Below is a comparison of accuracy with different feature combinations:

Why multimodal improves accuracy

Images contain information not present in text: style, color, visible material. For categories like "Dresses" or "Sneakers", visual features are critical. We use pre-trained vision encoders (e.g., CLIP) that output embeddings combined with text embeddings. This is especially effective for products with sparse descriptions.

Handling tricky cases

• Multi-category products: "Book-style phone case"—accessory or case? Both. We use multi-label classification.
• Mismatch between title and content: "Crafting kit"—what's inside? Need description. If missing, the model flags the product for manual review.
• New categories: Automatically create an "Unknown category" cluster for review. After confirmation, retrain the model.

Metrics: Top-1 accuracy, Top-3 accuracy (product in one of 3 predicted categories). Typical results: Top-1 90–95%, Top-3 97–99% for standard catalogs.

What our work includes

1. Data audit: analyze catalog structure, labeling quality, available features.
2. Model design: choose architecture (BERT, hierarchical, retrieval-based) for your case.
3. Training and validation: on your data with tracking via MLflow.
4. Integration: REST API or gRPC, documentation, code examples.
5. Deployment and monitoring: containerization, A/B testing, logging.
6. Team training: how to update the model, add new categories.

Timeline and cost

Basic solution: from 2 to 6 weeks. Cost is calculated individually after auditing your data and requirements. We guarantee transparent pricing and fixed timelines. Certified engineers (TensorFlow, AWS, GCP) ensure stable operation.

Request a consultation: we evaluate your catalog and propose the optimal solution. Get demo access to a working system.

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.