What is entity extraction?
Entity extraction is the process of identifying specific pieces of information in unstructured text or speech, such as names, dates, amounts, addresses and policy or account numbers, and returning them as structured, labeled values that software can use, for example to fill in a form field or look up a record.
Researchers call the core task named entity recognition, dialogue designers call it slot filling, and the wider field that also captures relationships between facts is known as information extraction. In a customer conversation it works alongside intent: the intent says what the person wants, and the entities supply the details needed to do it.
How entity extraction works
A request such as "I'd like to borrow twelve thousand for a car and pay it off over five years" passes through five steps before its details can be used.
| Step | What happens | Example |
|---|---|---|
| Detect | Find the words that carry information | "twelve thousand", "a car", "five years" |
| Label | Assign each one an entity type | Loan amount, loan purpose, term |
| Normalize | Convert each value to a standard format | 12000 in US dollars; auto; 60 months |
| Validate | Check values against rules and records | Amount within product limits; term actually offered |
| Confirm | Read back values that matter before using them | "That's twelve thousand over sixty months?" |
Different entities suit different methods. Values with a fixed shape, such as account numbers with check digits or postcodes, are handled well by pattern rules. Names, addresses and free-form descriptions need trained models or language models that understand context. Many systems combine the two and pass anything uncertain to a person or back to the customer to confirm.
Why entity extraction matters
General models are flexible extractors but not yet the most precise. In a 2023 evaluation of a general-purpose chat model on information extraction, it scored an F1 of 60.1 on a standard named entity recognition benchmark without examples, against 94.6 for the best specialized system. A later benchmark on medical text, published in Nature Communications in 2025, likewise found that fine-tuned models outperformed general models used without task training in most tasks.
An error in extraction is often more costly than an error in understanding intent. A misread intent usually produces a visibly wrong answer that the customer corrects. A misread amount, date or account number can pass quietly into an application, a payment instruction or a customer record, where it surfaces later as a declined transaction, a complaint or a correction that someone has to trace back.
Where entity extraction is used
- Applications from conversation. Filling application fields from what a customer says or types, then confirming them.
- Document processing. Reading identity documents, pay slips, invoices and claim forms.
- Record updates. Capturing new addresses, contact preferences and promised payment dates in the right systems.
- Redaction. Finding personal and card details in transcripts and documents so they can be masked.
- Search and analytics. Indexing conversations by product, amount, location or date.
Where a value is already held in a system or can be entered exactly, such as choosing a product from a short list or keying in a card number, extraction from free speech adds risk without adding much. The strongest designs extract where conversation is natural and switch to structured entry where precision is critical.
What good entity extraction looks like
- Confirmation of critical values. Amounts, dates and identifiers are read back before anything is submitted.
- Normalization rules. Every value is stored in one agreed format.
- Validation against systems. Extracted values are checked against records and product rules.
- Confidence-based review. Low-confidence values go to a review queue instead of straight into records.
- Masking at capture. Sensitive values are redacted from transcripts and logs as soon as they are recognized.
- Traceability. Each stored value can be linked to the words it came from.
How to measure entity extraction
- Field-level accuracy. Precision and recall for each entity type, measured on a labeled sample.
- Exact-match rate on critical fields. The share of amounts, dates and identifiers captured exactly right.
- Correction rate. How often customers or staff change a value after it was captured.
- Confirmation turns. How many exchanges it takes to settle each value, a proxy for friction.
- Redaction recall. The share of sensitive values found and masked in a test set.
Data collection rules for entity extraction
Extraction decides which pieces of a conversation become records, so the rules about what may be collected and how it must be protected apply directly to it. The main US points are listed below, and lenders, insurers and other regulated businesses face further requirements. Lending is used as the example here because its rules on what may be asked are among the strictest.
- Details a lender may not ask for. Regulation B generally prohibits creditors from inquiring about an applicant's race, color, religion, national origin or sex, with exceptions such as government monitoring on mortgage applications. Extraction should not record those details just because an applicant mentions them.
- Social Security numbers. California restricts how businesses display and transmit Social Security numbers. Pipelines that copy them into transcripts, summaries and logs multiply the places they are exposed.
- Card security codes. Payment card industry standards prohibit storing card verification codes after authorization, even in encrypted form, so they must be masked wherever extraction picks them up.
- Accuracy of reported data. Under Regulation V, companies that furnish information to credit bureaus must keep written policies for its accuracy and integrity, including details captured automatically.
Frequently asked questions
What is the difference between entity extraction and intent classification?
Intent classification decides what the person wants. Entity extraction finds the details needed to act on it, such as the amount, the date or the account.
What is named entity recognition?
The research name for identifying and labeling names of people, organizations, places, dates, amounts and similar items in text. It is the core of entity extraction.
Can language models do entity extraction?
Yes, and they handle varied phrasing well, especially when given a defined output format. For high-volume, fixed-format fields, specialized models or pattern rules are often more accurate and cheaper.
Why is extraction harder on voice?
Speech recognition errors come first: "fifteen" heard as "fifty", a name spelled two ways, an email address that is hard to say. Voice systems confirm critical values more often for that reason.
What is slot filling?
The dialogue-design term for collecting the details a task needs, one slot per detail, asking for any that are still missing before the task can proceed.
Related terms
Every detail captured, every step completed.
Encore AI's agents collect what an application needs in the course of a natural conversation, confirm it and submit it, using the questioning your best people already rely on.
Scale your best to be the best.
Book a Demo