Skip to main content
Skills for processing unstructured documents can only be built in Advanced Designer; the cloud-based Skill Designer does not support these scenarios. They use four core NLP activities to identify entities, segment text, and extract fields from freeform content like contracts, letters, and emails:
Each of these activities supports a limited set of languages. See the activity’s reference page for the language list.

​
Pick a scenario

Each scenario below shares a common pattern; only the activities you add to the processing flow differ.

​
Common workflow

1

Create a Document skill

Open Advanced Designer and click Create Document Skill on the start page.
2

Upload documents

On the Documents tab, upload the documents you’ll use to set up the skill.
3

Define fields and label

On the Fields tab, create and configure the fields the skill will extract. Label documents in the Reference section.
4

Add and configure NLP activities

On the Activities tab, add the activities for your scenario (described below). Open each activity in the Activity Editor to configure and train it.
5

Test and publish

Click Test Skill Using Selected Documents to evaluate results. When the results are good enough, publish the skill.

​
Pre-trained named entities (whole document)

Use this scenario when the entities you need can appear anywhere in the document — for example, company names and addresses in a letter. Add a Named Entities (NER) activity and map each named entity to a field. If you also need to break an address into components (street, city, state, country, postal code), add an Address Parsing activity and map the components to fields.
Document processing flow with a Named Entities (NER) activity

​
Pre-trained named entities (specific paragraphs)

Use this scenario when the entity always sits in the same paragraph — for example, a purchase amount in the price clause of a sales agreement. First isolate the paragraph with a Segmentation activity, and then run a Named Entities (NER) or Address Parsing activity on the segmented field. You can also isolate the paragraph with a Fast Learning or NLP Extraction Rules activity instead of Segmentation, and then run NER or Address Parsing on the result.
Pre-trained activities are easy to configure and need no training, but a neural network trained on your documents may extract more accurately. If you have a large document set, also try the custom named entities scenario below and pick whichever performs better.
Document processing flow with Segmentation feeding Named Entities (NER) and Address Parsing

​
Custom named entities (Deep Learning for NLP)

Use this scenario when pre-trained activities can’t disambiguate the entities you need — for example, extracting only one organization’s name from a paragraph that lists both parties to an agreement, or extracting an entity type that NER doesn’t cover (such as an email address). Pair a Segmentation activity with a Deep Learning for NLP activity: Segmentation isolates the paragraph and Deep Learning extracts the targeted fields.
Training a Deep Learning for NLP activity requires at least 50 documents (150 recommended). For best results, also try the pre-trained Named Entities (NER) activity and pick whichever extracts more accurately on your documents.
Document processing flow with Segmentation feeding a Deep Learning for NLP activity