Data annotation and labeling services
Native speakers label your text, documents and transcripts to your guidelines: named entities, sentiment, intent, categories and document fields. Every batch comes with agreement scores.
My order #48213 ORDER arrived in Leeds LOC on Tuesday DATE, but the charger PRODUCT is missing.
Entity spans, intent and sentiment on one message. The label set stays the same in every language.
What data annotation means
Data annotation is the work of adding labels to raw data, such as tagging the names in a sentence or marking the intent of a support message, so a machine learning model can learn from it or be tested on it.
This page covers text, documents and transcripts. Images and video have their own page, and so do speech and audio.
Annotation types
Labeled to your taxonomy and guidelines. If you don't have guidelines yet, we write the first draft with you during the pilot.
Named entity recognition
People, organizations, places, dates, amounts and product names, plus any custom entity types your model needs. Spans are marked to the exact character.
Sentiment and emotion
Positive, negative and neutral, or a five-point scale. Labeled by sentence or by aspect, so “the battery is great but the screen scratches” gets two scores.
Intent and classification
Support tickets, chats and search queries sorted into your intent list or category tree, with more than one label per item where your schema allows it.
Relations and coreference
Which entity connects to which, such as a person and the company they work for, and which “she” or “it” refers back to which name.
Document annotation
Fields and values in invoices, forms, receipts and contracts, plus table cells and reading order. OCR text is corrected by hand to give you ground truth.
Transcript annotation
Call and chat transcripts labeled by speaker turn, with topic, intent and outcome on each segment.
How quality is measured
You get these numbers with every batch, so you can see the quality without re-checking the labels yourself.
- Gold-set accuracy
- Items with agreed answers are mixed into every batch. Each annotator's score against them is tracked and reported.
- Agreement between annotators
- A share of items is labeled by two people. We report agreement as Cohen's kappa or Krippendorff's alpha, whichever you prefer.
- Senior review of disagreements
- When annotators disagree, a senior reviewer decides, and the decision is written into the guidelines.
- Batch reports
- Each delivery lists item counts, agreement and gold-set scores, and the questions raised during the batch.
Two edge cases, and the rules they added
“Great, the charger PRODUCT is missing again.”
Labeled negative. Rule added: score what the writer means, not the words. Sarcastic praise counts as negative.
“Order kal DATE aa gaya, but charger PRODUCT missing hai.”
Rule added: tag code-mixed messages with both languages, and label entities in either one. “Kal” means yesterday here, from the past tense.
How an annotation project runs
The pilot is where most problems surface, so we don't skip it.
Brief
Send your guidelines, label set and a sample of the data, with the languages and volume you need.
Pilot batch
We label a pilot batch, with part of it labeled twice to measure agreement.
Calibration
You review the pilot results with us. The guidelines are updated and fixed as version one.
Your sign-off QAProduction
Batches run on the agreed schedule, with gold items and double labeling throughout.
Delivery
Labels land in your platform or arrive as files in your format, with a report for each batch.
Formats, tools and languages
- Data you can send
- TXT, CSV, JSON, PDF, scanned images, and chat or call transcripts
- Labels we deliver
- JSONL with character offsets, CoNLL with BIO tags, CSV, or labels saved straight into your platform
- Platforms
- Your own labeling tool, or Label Studio or doccano set up for your project
- Languages
- English, European languages, Arabic, Hebrew, Farsi, Urdu, Hindi, Chinese, Japanese, Korean and smaller languages, confirmed per project
Formats and platforms to confirm
How your data is handled
The same rules apply to every AI project. See how we handle data.
- Signed NDAs. Every annotator and reviewer works under one, and we'll sign yours.
- Limited access. Only the people assigned to your project can open your data.
- Deleted at the end. We delete your data when the project ends, unless you ask us to keep it.
- Yours to keep. You own the data and the labels. We don't reuse them for other clients.
Related services
Questions about data annotation
What is data annotation?
Data annotation is adding labels to raw data so a machine learning model can learn from it or be tested on it. For text, that means marking things like names, dates, sentiment or the intent behind a message. For documents, it means marking fields such as invoice numbers and totals.
Which annotation types do you handle?
Named entities, sentiment and emotion, intent and classification, relations and coreference, document fields and transcript segments. Image, video and audio work is covered on their own pages.
Can you work in our labeling tool?
Yes. We can work inside your platform under your access rules. If you don't have one, we set up a tool for the project and deliver the labels as files in your format.
How do you measure agreement between annotators?
Part of each batch is labeled by two annotators, and we calculate agreement as Cohen's kappa or Krippendorff's alpha. Gold items with agreed answers are also mixed into every batch to score each annotator. You get both numbers with each delivery.
Can you annotate languages other than English, including dialects?
Yes. Each language is labeled by native speakers. For dialects, code-mixed text such as Hindi and English written together, and low-resource languages, we confirm we have enough qualified annotators before the pilot starts.
Do you also label images, video and audio?
Yes. See image and video annotation and speech and audio data.
Start with a pilot batch
Send your guidelines and a sample of your data. We'll label a pilot batch, report the agreement scores and the questions it raised, and quote for the full set.
- Phone
- +44 20 3290 0116
- Phone hours
- Weekdays 9:30 to 16:00, weekends 9:30 to 12:00, UK time