ifc_ner_dataset / README.md
mazri24's picture
Update README.md
aacaeef verified
metadata
language:
  - en
license: mit
task_categories:
  - token-classification
task_ids:
  - named-entity-recognition
pretty_name: IFC NER Dataset
size_categories:
  - 100K<n<1M
tags:
  - bim
  - ifc
  - aec
  - openbim
  - named-entity-recognition
  - ner
  - synthetic-data
  - ai

IFC NER Dataset

A synthetic Natural Language Processing (NLP) dataset for Named Entity Recognition (NER) in the Architecture, Engineering, and Construction (AEC) domain, generated from Industry Foundation Classes (IFC) models.

This dataset enables training models to extract structured BIM information from natural language sentences.

📌 Overview

The dataset consists of automatically generated sentences derived from IFC files.

Each sentence is:

  • Tokenized
  • Labeled using the BIO tagging scheme

The dataset supports extraction of:

  • IFC entities (e.g., IfcWall, IfcDoor)
  • Relations (e.g., has, contains, part_of)
  • Properties (e.g., Height, FireRating)
  • Materials
  • Property values

📂 Dataset Structure

The dataset follows the Hugging Face DatasetDict format:

  • train
  • validation
  • test

Each example looks like:

{
  "tokens": ["The", "wall", "has", "height", "3.0", "meters"],
  "bio": ["O", "B-IFC_ENTITY", "B-RELATION", "B-PROPERTY", "O", "O"]
}

Where:

  • tokens: Tokenized sentence
  • bio: BIO-formatted NER tags aligned with tokens

BIO Tag meaning:

  • B-XXX → Beginning of entity type
  • I-XXX → Inside entity type
  • O → Outside any entity

🏗 Data Generation

Sentences are programmatically generated from real IFC models.

The generation includes:

  • Entity mention templates
  • Property description templates
  • Relationship-based sentences
  • Passive voice variants
  • Question-style sentences
  • Multi-property sentences

This provides structured consistency with real BIM data while introducing controlled linguistic variation.

🎯 Intended Use

This dataset is suitable for:

  • Token classification (NER)
  • Domain adaptation for AEC NLP models
  • Text → structured BIM workflows
  • Text → IDS generation pipelines
  • BIM-aware information extraction systems

📊 Scale

  • Generated from IFC models
  • 300k+ labeled sentences
  • BIO token-level annotation

⚠️ Limitations

  • Sentences are synthetically generated.
  • Linguistic diversity depends on template variety.
  • Does not cover the full IFC schema.
  • Distribution reflects the source IFC models used.