NLP Product Categorization

Our models support multi-label classification, meaning a product can be assigned to multiple categories with independent confidence scores. For example, a 'yoga mat bag with water bottle holder' might be classified as both 'Sports > Yoga Accessories' (0.92 confidence) and 'Bags > Gym Bags' (0.78 confidence). You can configure whether to use the top-1 category, top-N categories, or a confidence threshold for your specific use case.

Yes. We ingest your custom taxonomy as a target classification scheme, including any number of levels, nodes, and naming conventions. Our models learn the mapping from product content to your specific category structure. For initial setup, we need your taxonomy tree and a sample of products already categorized in your system (ideally 50-100 examples per leaf category). The model trains on your examples and generalizes to your full catalog.

Accuracy varies by attribute type. For well-defined attributes like brand, color, and size, accuracy exceeds 98%. For more complex attributes like material composition, compatibility lists, and technical specifications, accuracy is typically 93-96%. Every extracted attribute includes a confidence score. We recommend setting a confidence threshold (e.g., 0.85) and routing low-confidence extractions for human review to maintain your quality standards.

We use multilingual transformer models (XLM-RoBERTa) that were pre-trained on text in 100+ languages simultaneously. These models learn shared semantic representations across languages, meaning they understand that 'Kopfh_rer' (German), 'headphones' (English), and '????' (Japanese) all refer to the same product type. This approach avoids the errors and latency of translation APIs while providing native-quality understanding in every supported language.

Our aspect-based sentiment analysis goes far beyond a simple positive/negative score. It identifies: (1) specific features mentioned (battery life, build quality, ease of use), (2) sentiment per feature (battery life: positive, build quality: negative), (3) frequency of each mention, (4) comparison statements (better than X, worse than Y), (5) purchase intent signals, (6) complaint patterns that repeat across reviews, and (7) temporal trends showing how sentiment changes over time — for example, detecting a quality issue emerging in recent reviews.

Our models handle novel products through a combination of semantic generalization and active learning. If a genuinely new product type appears (e.g., a new gadget category), the model classifies it to the most semantically similar existing category and flags it with a lower confidence score. Our active learning pipeline prioritizes these low-confidence products for human review, and the corrections are incorporated into the next model training cycle — typically within one week.

Yes. Our tagging output is designed for direct integration with search platforms including Elasticsearch, Algolia, Solr, and proprietary search engines. Tags are delivered as structured arrays that can be indexed as searchable facets. We also support tag hierarchy (parent-child relationships), synonym mapping, and weighted tags that signal relevance strength. Integration is typically via API, webhook, or direct database write.

For sparse listings with minimal text, our system uses multiple fallback strategies: (1) image-based classification using computer vision to identify the product type from photos, (2) category inference from price range and seller category context, (3) cross-reference matching against known products in our database using partial title matching, and (4) hierarchical classification that at minimum assigns a parent category even when leaf-level confidence is low. The confidence score transparently reflects data scarcity.

Our NLP pipeline processes 10,000+ products per minute for standard categorization (title + description classification). For full enrichment jobs including attribute extraction, sentiment analysis, tagging, and cross-taxonomy mapping, throughput is approximately 3,000-5,000 products per minute. Bulk jobs of millions of products are parallelized across GPU clusters and typically complete within hours. Real-time API classification returns results in under 200ms per product.