TextAPI

← All articles

Named entity recognition without training a model

· 6 min read

Pull people, companies, money, and dates out of raw text with one HTTP call — including the character offsets you need to highlight them.

Named entity recognition is the step where unstructured text becomes something you can query: which companies are mentioned in this filing, and how much money? Training a model for it is a project. Calling one is a line of code.

One call, structured output

curl -G "https://use.textapi.com/ner" \
  --data-urlencode "text=Satya Nadella said Microsoft will invest \$10 billion in OpenAI next January"

{
  "spans": [
    {"text": "Satya Nadella", "start": 0,  "end": 13, "label": "PERSON"},
    {"text": "Microsoft",     "start": 19, "end": 28, "label": "ORG"},
    {"text": "$10 billion",   "start": 41, "end": 52, "label": "MONEY"},
    {"text": "next January",  "start": 63, "end": 75, "label": "DATE"}
  ]
}

Four facts from one sentence, each with a type and a position.

The offsets are the useful part

start and end are character indices into the exact string you sent, with end exclusive — the same convention as Python slicing. That means text[span["start"]:span["end"]] always reproduces span["text"], which is what makes highlighting reliable:

import requests

text = "Satya Nadella said Microsoft will invest $10 billion in OpenAI next January"
spans = requests.get(
    "https://use.textapi.com/ner", params={"text": text}
).json()["spans"]

# Rebuild the string with <mark> tags around each entity.
out, cursor = [], 0
for span in sorted(spans, key=lambda s: s["start"]):
    out.append(text[cursor:span["start"]])
    out.append(f'<mark data-label="{span["label"]}">{text[span["start"]:span["end"]]}</mark>')
    cursor = span["end"]
out.append(text[cursor:])

print("".join(out))

Walk the spans in order and slice — never search for the entity text with str.replace. If “Microsoft” appears three times, replace hits the wrong one.

The labels you’ll actually see

The common ones, in rough order of how often they show up in business text:

Label What it catches
PERSON People, real or fictional
ORG Companies, agencies, institutions
GPE Countries, cities, states
LOC Non-political locations — mountains, bodies of water
DATE Absolute and relative dates, including “next January”
MONEY Monetary values with units
PRODUCT Objects, vehicles, foods — not services
CARDINAL Numerals that don’t fall under another type

Also present: TIME, PERCENT, QUANTITY, ORDINAL, EVENT, WORK_OF_ART, LAW, LANGUAGE, NORP (nationalities and political groups), and FAC (buildings, airports, highways).

Small model or large model

The default is the small model. Pass model=en_core_web_lg for the large one:

curl -G "https://use.textapi.com/ner" \
  --data-urlencode "text=Satya Nadella said Microsoft will invest \$10 billion in OpenAI next January" \
  --data-urlencode "model=en_core_web_lg"

On that sentence, the large model finds a fifth entity the small one misses entirely — OpenAI — which is the general pattern: the large model has better recall on names it hasn’t seen in an obvious context.

It also labels OpenAI as GPE — a geopolitical entity — which is simply wrong. That’s the honest state of off-the-shelf NER: recall improves with model size, but labels on unusual or newer proper nouns are unreliable in both models. If your downstream logic branches on the label, either constrain yourself to the types the model is confident about (PERSON, DATE, MONEY are dependable) or treat ORG/GPE as one bucket and disambiguate yourself.

Rendering it for a UI

If the end goal is showing highlighted text to a human rather than processing spans in code, /ner/display skips a step and returns rendered markup:

curl "https://use.textapi.com/ner/display?text=Apple+hired+Jane+Doe+in+Paris&format=html"

format=svg (the default) returns an inline fragment you can drop into an existing page; format=html returns a complete standalone document. You can also pass your own spans as JSON — useful when you’ve corrected the model’s output, or when the entities come from your own database rather than from the model at all.

Practical notes

Trim long documents. Recognition on a 40,000-character article is slow, and entities from paragraph 200 are rarely what you’re after. The opening section usually carries the ones that matter.

Send real casing and punctuation. These models learned on ordinary prose; uppercasing everything or stripping punctuation measurably hurts accuracy.

Deduplicate downstream, not upstream. Every mention is returned separately, including repeats. That’s what you want for highlighting; for a summary of “companies in this document,” collapse on the entity text yourself.

Keep reading