ele-sage/person-company-names-classification

Dataset Card for Person vs. Company Names Classification

This is a large-scale, highly curated dataset containing ~7,5 million examples, designed specifically for training a binary text classification model. The primary task is to distinguish between a person's full name and a company/organization name with high precision.

Dataset Details

Dataset Description

This dataset was built to train a robust "Privacy Gatekeeper" classifier. It is designed to handle real-world database inconsistencies (like swapped first/last names) and difficult edge cases (like companies named after people).

The data has undergone a rigorous cleaning process using NER models, manual curation, and synthetic augmentation to ensure the model learns semantic patterns rather than simple heuristics.

• —Curated by: ele-sage
• —Language(s) (NLP): English, French
• —License: MIT
• —Task: Binary Classification (0: Person, 1: Company)

Uses

Direct Use

This dataset is intended for training models to identify if a string represents a legal entity (Company) or a human being (Person). It is particularly optimized for:

• —Database Cleaning: Standardizing mixed columns of names.

from datasets import load_dataset

dataset = load_dataset("ele-sage/person-company-names-classification")
print(dataset["train"][0])
# {'text': 'Entreprise Shopicar Inc.', 'label': 1}

Dataset Structure

The dataset is provided in a train / test split (approx 95/5). Each record contains:

• —`text`: (string) The cleaned, title-cased name.
• —`label`: (integer)
• —0: Person (e.g., "Jean Tremblay")
• —1: Company (e.g., "Tremblay Construction Inc.")

Dataset Creation

Curation Rationale

Standard NER datasets often lack the specific edge cases found in administrative databases. A simple model might learn that "starts with a digit" means Company, or "looks like a name" means Person. This dataset was engineered to break those lazy patterns by:

1. 1.Removing "easy" signals: Drastically reducing numbered companies.
2. 2.Adding "hard" signals: Generating synthetic companies using real person names (Hard Negatives).

Source Data & Processing

The dataset combines two distinct sources, processed through a multi-stage pipeline.

1. Person Names (Label 0)

• —Source: A large dataset of ~3.5M names (Facebook leak origin).
• —Filtering Pipeline:
• —Character Validation: Removed all entries containing characters outside standard English/French diacritics (removed Cyrillic, Chinese, Emojis, etc.).
• —Semantic Blacklisting (NER/POS): Used a Named Entity Recognition (NER) and Part-of-Speech model to analyze all words. Identified common non-name tokens (verbs, objects, junk).
• —Blacklist Cleanup: Filtered the dataset using the final curated blacklist.
• —Formatting:
• —75% FirstName LastName
• —25% LastName FirstName (To force the model to learn word semantics rather than just position).

2. Company Names (Label 1)

• —Source: Quebec Enterprise Register (public government data).
• —Downsampling Numbered Companies:
• —Raw data contained millions of entries like "9123-4567 QUEBEC INC".
• —Action: These were downsampled to ~30,000 examples.
• —Reasoning: To prevent the model from over-fitting on digits. We want the model to read the text, not just check for numbers.
• —Augmentation (Hard Negatives):
• —Generated ~300,000 synthetic company names by combining real Person Names with business suffixes.
• —Examples: "Jean Tremblay" (Person) vs "Tremblay Plumbing" (Company).
• —Reasoning: To teach the model that a valid person's name can be part of a company name, forcing it to pay attention to the suffix.

3. Final Processing

• —Merging: All sources combined.
• —Cleaning: Applied Title Case and removed excess whitespace.
• —Deduping: Removed exact string duplicates.
• —Splitting: Stratified split into Train/Test.

Annotations

Labels were inferred from the source provenance:

• —Curated Person Data -> 0
• —Curated Company Data (Real + Synthetic) -> 1

Bias, Risks, and Limitations

• —Geographic Bias: The dataset is heavily skewed towards Quebec/Canada naming conventions (French/English mix).
• —Synthetic Data: The "Hard Negative" companies are synthetically generated. While they follow realistic patterns, they may not represent real legal entities.
• —Ambiguity: Certain strings (e.g., "Marshall & Co") are inherently ambiguous without further context. The model is trained to treat these as Companies (Label 1).