Entity Extraction — Definition, How It Works in Chatbots (2026)

What it is

When a user types "Cancel my October 15 order for the blue dress in size M", a chatbot needs to know:

• Intent: cancel_order
• Date: October 15
• Product: blue dress
• Size: M

The dates, product, and size are entities. Entity extraction is the NLU task of finding and labeling them.

Standard entity types in most platforms:

• Person — names
• Location — cities, addresses, countries
• Organization — company names
• Date / Time — calendar references ("tomorrow", "next Tuesday", "3 PM")
• Money / Number — currency amounts, quantities
• Email / Phone / URL — structured strings

Plus custom entities specific to your domain — product names, SKUs, account IDs, plan tiers, etc.

How it works

Two main approaches:

1. Rule-based extraction

Regular expressions + dictionaries: a regex catches phone numbers; a dictionary of product names matches against the message text. Cheap, predictable, language-specific.

2. ML-based NER

Neural network trained on labeled examples. Modern platforms use either pretrained NER models (spaCy, Stanford NLP) or LLM-based extraction (prompt the LLM to "extract dates, products, and sizes from this message").

LLM-based extraction handles phrasing variation and multilingual input gracefully. Rule-based extraction is faster and cheaper but brittle.

Entity extraction in chatbot platforms

• Dialogflow — first-class entity definitions, system entities (date, location, currency, etc.) + custom entities.
• Botpress — visual entity definitions, LLM-extraction available.
• Rasa, Microsoft Bot Framework — similar explicit entity approach.

Marketing-focused platforms (Manychat, Chatfuel) typically capture entities through explicit form fields or button menus rather than NER on free text. Customer-service and AI-agent platforms make heavier use of NER.

When entity extraction matters

• Transactional flows. Booking, cancellation, lookup — these need structured data extracted from free text.
• CRM enrichment. Pulling names, companies, emails from conversation to populate records.
• Search and filtering. "Show me red shoes under $100" needs entity extraction (color, max_price) to translate to a database query.
• Multi-step conversations. Slot-filling — asking the user for each missing entity until all are collected.

Common pitfalls

• Ambiguous references. "Last Monday" depends on the current date. "Sales rep John" depends on which John. Production systems disambiguate carefully.
• Custom entities require maintenance. Product catalogs change; entity dictionaries need updates.
• Multi-language gotchas. Date formats differ (US: month/day; LATAM/EU: day/month). Currency symbols vary. Language-specific NER training data is uneven.

Worked example — slot-filling a booking flow

A salon chatbot needs to capture three entities to book an appointment: service, date, and time. Here is how the bot handles realistic user messages:

Cases 3 and 5 illustrate why fuzzy time and date handling matters in production. Case 4 shows how entity extraction alone is insufficient when references are anaphoric — the bot must combine NER with CRM lookup. Case 6 demonstrates the multilingual edge: a PT-BR-trained NER understands "amanhã" (tomorrow) and "às 14h" (at 2 PM), while an English-only NER would fail entirely.

Entity types in production

Most platforms ship a standard library of pre-built entity types you can use without custom training:

Custom entities (@product, @plan_tier, @account_id) cover business-specific vocabulary that no platform ships by default. These require ongoing maintenance as product catalogs evolve.

Related terms

• Natural Language Understanding — the broader NLU category.
• Intent recognition — the complementary NLU task.
• Natural Language Processing — the parent field.

FAQ

Is entity extraction the same as NER?

Yes — "Named Entity Recognition" is the academic / technical term; "entity extraction" is more common in product documentation. Both refer to the same task.

Can LLMs do entity extraction without training data?

Yes. Prompt a modern LLM with "Extract the date, product, and size from this message: [user message here]" and it works zero-shot for most common entity types in major languages. Dedicated NER models are still cheaper to run at scale.

What's the difference between entity extraction and intent recognition?

Intent = what the user wants (action). Entity = what specific data points the action needs (parameters). Both run on the same message; both fill in slots in the chatbot's response logic.

How accurate is modern entity extraction?

For well-defined entity types (date, time, currency, email) in major languages, modern LLM-based extraction reaches 90-97% precision. Custom domain entities (specific product SKUs, account IDs) depend heavily on training data quality and gazetteer coverage — well-maintained custom extractors hit 85-95%, neglected ones drop to 60-70%.

Can I extract entities from voice input?

Yes, but the pipeline adds latency and error compounding: speech-to-text transcription (95-99% accuracy in clean audio) feeds into entity extraction, so any transcription error propagates. Voice-specific gotchas include digit homophone confusion (15 vs 50), proper-noun mishearings, and dialect-specific number formats. Pair voice entity extraction with explicit user confirmation for high-stakes data.

Sources

• Dialogflow documentation. Entities concepts. cloud.google.com/dialogflow (verified 26 May 2026).
• Stanford NLP. NER course materials. nlp.stanford.edu/ner (verified 26 May 2026).