ACCORD-NLP/CODE-ACCORD-Entities

---
language:
- en
---

# 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](https://accordproject.eu/)). 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](https://github.com/Accord-Project/CODE-ACCORD/blob/main/annotated_data/Annotation_Strategy_V1.0.0.pdf) 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.

```python
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.

```python
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{hettiarachchi-etal-2025-code,
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={Scientific Data},
volume={12},
number={1},
pages={170},
year={2025},
publisher={Nature Publishing Group UK London}
}
```

CODE-ACCORD语料库包含来自英格兰和芬兰建筑法规的标注句子，是Horizon欧洲项目的一部分，旨在支持建筑、翻新或拆除工程的自动合规检查。该语料库为英文，包括英格兰建筑法规和芬兰国家建筑法典的英文翻译。数据集主要关注从文本中提取信息以支持规则生成，特别是实体和关系的标注。实体标注分为四类：对象、属性、质量和值；关系标注分为十类：选择、必要性、部分、非部分、大于、大于等于、等于、小于等于、小于和无。数据集分为训练集和测试集，分别占80%和20%。

原始信息汇总

数据集概述

名称: CODE-ACCORD

描述: CODE-ACCORD是一个包含英格兰和芬兰建筑法规中注释句子的语料库，旨在支持自动合规检查的规则生成。该数据集主要关注从文本中提取的两种关键信息：命名实体和关系。

数据注释

• 实体注释: 数据集定义了四个实体类别：对象、属性、质量和价值。
• 关系注释: 数据集定义了十个关系类别：选择、必要性、部分、非部分、大于、大于等于、等于、小于等于、小于和无。

数据分割

• 分割方式: 数据集分为训练集和测试集，其中训练集占80%，测试集占20%。

数据格式

• 实体数据文件格式:
  • 属性包括：example_id, content, processed_content, label, metadata。
• 关系数据文件格式:
  • 属性包括：example_id, content, metadata, tagged_sentence, relation_type。

数据使用

• 实体数据: 使用以下Python代码加载数据： python 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))

• 关系数据: 使用以下Python代码加载数据： python 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))

引用信息

• 详细信息: 数据注释、统计和分布的更多细节可在以下论文中找到：

  @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} }