Named Entity Recognition

Named Entity Recognition (NER) is a natural language processing (NLP) technique that aims to locate and classify named entities in text into predefined categories. Named entities refer to real-world objects such as persons, locations, organizations, dates, and more. NER plays a crucial role in extracting meaningful information from unstructured text data, enabling automated analysis, information retrieval, and data mining.

Example: Identifying PII / PHI with NER

In this example, we will utilize NER to identify and classify different types of Personally Identifiable Information (PII) and Protected Health Information (PHI) within a given text. The following table presents a subset of the example dataset:

Categories of PII

We would like to identify the following types of PII in the given text:

• Name: Full Name - The complete name of an individual.
• Address: Home Address - The residential address of an individual.
• Email: Email Address - Personal email addresses.
• SSN: Social Security Number - A unique number assigned to individuals in the United States for identification purposes.
• Passport: Passport Number - A unique number assigned to a passport document.
• License: Driver's License Number - A unique number assigned to a driver's license.
• CreditCard: Credit Card Numbers - Numbers associated with personal credit cards.
• Birthdate: Date of Birth - The birth date of an individual.
• Phone: Telephone Number - Personal landline or mobile phone numbers.
• MedicalRecord: Medical Record Numbers - Unique identifiers for personal medical records.
• Biometric: Biometric Identifiers - Fingerprints, facial recognition patterns, DNA, etc.
• Vehicle: Vehicle Identifiers - License plate numbers, Vehicle Identification Number, etc.
• InternetActivity: Internet Activity - IP addresses, cookie IDs, device identifiers, etc.
• Employment: Employment Information - Employee ID number, work email, work phone number, etc.
• Education: Educational Records - Student ID number, transcripts, etc.

Using Anote to extract entities within text

Upload Data: Start by uploading the dataset into Anote. We upload the CSV in Unstructured format, choose the NLP task of Named Entity Recognition, and choose the per line decomposition.

Customize Categories: In the annotation interface, we want to set up the entity categories as listed above. We customize the entities in Anote's configuration settings by pressing the + Add Taxonomy button to add the entities.

We can add labeling function to identify key words where entities may appear in the text.

Annotate a Subset of Text Data: Anote provides an intuitive interface where we can view the text data that the model is least certain about, and annotate the edge cases with one of the defined categories above. On Anote, we label few rows of entities by highlight all the entities in the given row of text, and then clicking the "confirm" button once completed. We can also remove entities that the model may predict incorrectly.

Export as CSV: Once finished annotating a few entities, we can download the resulting entities in CSV format. The entity category should be in parenthesis, to the right of where the entity was identified, as seen below.

Notice below that the NER model is able to pick up entities such as credit card information that weren't even annotated yet, this is through a fusion with the zero shot NER model.

Below is what a proper output should look like.

Hierarchical PII Taxonomy

Due to changing business requirements, we want to refine our entity identification requirements on this example dataset. The rationale behind this decision lies in the pursuit of more granular entities that can be precisely identified within text data.

Introduction of Hierarchies: To better organize and manage the diverse types of PII, we introduced hierarchical structures within the entity identification process. This approach allowed us to establish clearer relationships between PII categories, such as IDs and Contact Info.

Addition of New Categories: Recognizing the evolving nature of data, we decided to incorporate additional categories, such as usernames and student IDs.

Table 2: Subcategories - ID and Contact Info

IDs

• License
• Passport
• SSN
• CreditCard

Contact

• Phone
• Username
• Student ID
• Email

Categories without Subcategories

• Name
• Address
• Birthdate

Adapting to Changing NER Requirements

To make adjust to the changing business requirements, we navigate to the customize tag. We can add nested sub-entities underneath existing entities. We can also define new entities. After we make the modifications, we can expand our categories to see all of the categories.

We can also collapse the categories to only see the parent categories.

Once confirmed, we can begin annotating, incorporating the previous knowledge to inform our updated entity predictions. Once again, we can highlight over the key words in the text that are indicative of the entity, predict the entity, and click confirm when finished.

Real Time Updates

One important concept regarding Anote is the dynamic nature of our product. As business requirements change, as we add hierarchies and add new categories, our model is able to auto-adjust, incorporating previous knowledge to update the entity predictions synchronously in real time. Here is the resulting download table, from the CSV output.