ACCORD-NLP/CODE-ACCORD-Relations

CODE-ACCORD: A Corpus of Building Regulatory Data for Rule Generation towards Automatic Compliance Checking

The CODE-ACCORD corpus contains annotated sentences from the building regulations of England and Finland and has been developed as part of the Horizon European project for Automated Compliance Checks for Construction, Renovation or Demolition Works (ACCORD). The corpus is in English, and it consists of both the English Building Regulations and the English translation of the Finnish National Building Code.

Data Annotation

CODE-ACCORD is mainly focused on extracting information from text to support rule generation. There are two key types of information found in the text: named entities and relations, which are essential for comprehending the ideas conveyed in natural language. Hence, this dataset primarily focused on annotating entities and relations.

Four categories were considered for entity annotation: (1) object, (2) property, (3) quality and (4) value. The relations annotations span in ten categories: (1) selection, (2) necessity, (3) part-of, (4) not-part-of, (5) greater, (6) greater-equal, (7) equal, (8) less-equal, (9) less and (10) none. Please refer to our Annotation Stragety for more details about the categories and sample annotations.

Data Splits

Both entity and relation-annotated data consist of two data splits named train and test. The train split forms 80% of the full dataset, while the remaining 20% belongs to the test split.

Entities

The format of an entity data file is as follows:

Attribute	Description
example_id	Unique ID assigned for each sentence
content	Original textual content of the sentence
processed_content	Tokenised (using NLTK's word_tokenize package) textual content of the sentence
label	Entity labelled sequence in IOB format
metadata	Additional information of sentence (i.e. original approved document from which the sentence is extracted)

Using Data

The train and test splits of entity-annotated data can be loaded into Pandas DataFrames using the following Python code.

from datasets import Dataset
from datasets import load_dataset

train = Dataset.to_pandas(load_dataset('ACCORD-NLP/CODE-ACCORD-Entities', split='train'))
test = Dataset.to_pandas(load_dataset('ACCORD-NLP/CODE-ACCORD-Entities', split='test'))

Relations

The format of a relation data file is as follows:

Attribute	Description
example_id	Unique ID assigned for each sentence
content	Original textual content of the sentence
metadata	Additional information of sentence (i.e. original approved document from which the sentence is extracted)
tagged_sentence	Sentence with tagged entity pair
relation_type	Category of the relation in between the tagged entity pair

Using Data

The train and test splits of relation-annotated data can be loaded into Pandas DataFrames using the following Python code.

from datasets import Dataset
from datasets import load_dataset

train = Dataset.to_pandas(load_dataset('ACCORD-NLP/CODE-ACCORD-Relations', split='train'))
test = Dataset.to_pandas(load_dataset('ACCORD-NLP/CODE-ACCORD-Relations', split='test'))

Citation

More details about data annotation, statistics, and distribution are available in the following paper.

@article{hettiarachchi2024code,
  title={{CODE-ACCORD}: A Corpus of Building Regulatory Data for Rule Generation towards Automatic Compliance Checking},
  author={Hettiarachchi, Hansi and Dridi, Amna and Gaber, Mohamed Medhat and Parsafard, Pouyan and Bocaneala, Nicoleta and Breitenfelder, Katja and Costa, Gon{\c{c}}al and Hedblom, Maria and Juganaru-Mathieu, Mihaela and Mecharnia, Thamer and Park, Sumee and Tan, He and Tawil, Abdel-Rahman H. and Vakaj, Edlira},
  journal={arXiv preprint arXiv:2403.02231},
  year={2024},
  url={https://arxiv.org/abs/2403.02231}
}