ele-sage/person-names-ner

Dataset Card for Person Full Name NER Parsing

This dataset contains 3,383,944 curated and augmented person names, designed specifically for training Token Classification (NER) models. The primary task is to parse a full name string into its FirstName and LastName components, correctly handling multi-word names and different ordering formats.

Dataset Details

Dataset Description

This dataset is built to train robust models that can understand and segment human names. To improve real-world performance, the data has been augmented to include examples in three distinct formats:

• FirstName LastName (50% of the data)

• LastName, FirstName (25% of the data)

• LastName FirstName (25% of the data, the most ambiguous case)

• Curated by: ele-sage

• Language(s) (NLP): English, French

• License: MIT

Dataset Sources

• Repository: https://huggingface.co/datasets/ele-sage/person-full-name-ner-parsing

Uses

Direct Use

This dataset is intended for training token-classification models for Named Entity Recognition on person names. The primary goal is to create a "name splitter" or "name parser" that can handle various formats. It can be loaded directly using the 🤗 Datasets library:

from datasets import load_dataset

# Load the dataset directly from the Hub
dataset = load_dataset("ele-sage/person-full-name-ner-parsing")

print(dataset["train"][0])
# Expected output:
# {'first_name': 'Andrew', 'last_name': 'Burt', 'gender': 'M', 'country_code': 'CA', 'tokens': ['Burt', 'Andrew'], 'ner_tags': [2, 0]}```

Out-of-Scope Use

• General NER: This dataset is highly specialized and should not be used to train models for general-purpose NER (e.g., finding locations, organizations, dates).
• Company Name Parsing: The dataset does not contain company or organization names. The cleaning process was explicitly designed to remove them.

Dataset Structure

The dataset is provided in a train / test split (90/10). Each record is a JSON object with the following fields:

• first_name: (string) The original first name from the source data.
• last_name: (string) The original last name from the source data.
• gender: (string) The gender associated with the name in the source data (M, F, or null).
• country_code: (string) The country code from the source data (always, CA).
• tokens: (list of strings) The full name, pre-tokenized into a list of words, potentially reordered and with a comma.
• ner_tags: (list of integers) The corresponding IOB-style tags for each token in the tokens list. The mapping is as follows:
  • 0: B-FNAME (Beginning of a First Name)
  • 1: I-FNAME (Inside a First Name)
  • 2: B-LNAME (Beginning of a Last Name)
  • 3: I-LNAME (Inside a Last Name)
  • 4: O (Outside, for non-entity tokens like commas)

Dataset Creation

Curation Rationale

The primary motivation was to create a large-scale, high-quality dataset for the specific task of parsing human names. Standard NER datasets are too general, and simple rule-based parsers fail on ambiguous or multi-word names. This dataset was created to train a robust AI model that could overcome these limitations by learning from a diverse set of augmented examples.

Source Data

The dataset originates from a single source: a large CSV file of over 3.4 million Canadian names, originally from a Facebook data leak. Source Link.

Data Collection and Processing

The creation of this dataset involved a multi-stage curation pipeline to ensure high quality and robustness:

1. AI-Powered Cleaning: The raw 3.4M row CSV was first processed using the ele-sage/distilbert-base-uncased-name-classifier model. Any entry that the model did not classify as a "Person" was discarded. This was a critical step to remove a significant amount of non-name text and noise (e.g., company names, phrases, spam).

2. Rule-Based Filtering: The AI-cleaned data was further filtered to:

   • Remove entries where the first and last names were identical.
   • Remove entries containing characters outside a pre-defined set of English, French, and common punctuation characters.

3. Data Augmentation & Tagging: The final, cleaned dataset of 3,383,944 names was programmatically tagged and augmented to create three different name formats:

   • 50% was formatted as FirstName LastName.
   • 25% was formatted as LastName, FirstName (with the comma token tagged as O).
   • 25% was formatted as the ambiguous LastName FirstName (without a comma).

4. Shuffling: The final combined dataset was then shuffled.

Who are the source data producers?

The original data was created by users of the social media platform Facebook. The data was later made public through a data leak.

Annotations

Annotation Process

The annotation for this dataset was a two-stage process:

1. Primary Human Annotation: The ground-truth labels were originally provided by Facebook users themselve when they entered their names into separate, structured first_name and last_name fields in a registration form.

2. Secondary Programmatic Transformation: A script was then used to process this structured data. This script converted the first_name and last_name fields into a sequential format suitable for token classification. It generated the tokens list and the corresponding ner_tags (using the IOB scheme) for each of the three augmented formats (FirstName LastName, LastName, FirstName, and LastName FirstName).

Who are the annotators?

The primary, ground-truth annotation was performed by the original data producers (the users of the social media platform).

The secondary annotation format (the ner_tags) was generated programmatically by the data curation scripts developed by the dataset author (ele-sage).

Bias, Risks, and Limitations

• Geographic & Cultural Bias: The dataset is heavily biased towards North American (specifically Canadian) and European name structures. It will not perform well for training models on names from other cultural contexts (e.g., East Asian, Icelandic) where naming conventions differ.
• Source Data Noise: As the data originates from a data leak of user-provided information, it contains inherent noise, typos, and non-name entries. While a rigorous cleaning process was applied, some noise may persist.
• Limited Scope: The ner_tags are strictly limited to first and last names. The dataset provides no information on titles (Dr., Mr.), suffixes (Jr., III), or middle names.

Dataset Card Authors

ele-sage

Dataset Card Contact

ele-sage on Hugging Face