README.md

---
dataset_info:
  features:
  - name: id
    dtype: string
  - name: sentence
    dtype: string
  - name: relation
    dtype: string
  - name: tokens
    sequence: string
  - name: tags
    sequence: string
  splits:
  - name: train
    num_bytes: 288609496
    num_examples: 300067
  - name: test
    num_bytes: 36305820
    num_examples: 37508
  - name: validation
    num_bytes: 36152287
    num_examples: 37509
  download_size: 107574628
  dataset_size: 361067603
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
  - split: test
    path: data/test-*
  - split: validation
    path: data/validation-*
license: mit
task_categories:
- token-classification
- text2text-generation
- summarization
language:
- en
tags:
- engineering design
- knowledge extraction
---

Dataset Copyright - L. Siddharth, Singapore University of Technology and Design, Singapore.

The dataset includes 375,084 example sentences (187200 positive, 187884 negative), each including a pair of entities and the engineering design relation between these.
The dataset was manually constructed using sentences in 4,205 patents granted by USPTO, stratified according to 130 classes.

The dataset is used to train token classification and Seq2Seq transformer models to populate explicit engineering design facts from artefact descriptions.

More details in the following paper.
Siddharth, L., Luo, J., 2024. Retrieval-Augmented Generation using Engineering Design Knowledge. arXiv (cs.CL) https://arxiv.org/abs/2307.06985.

In each example, 
The input is a pair of entities is marked in a sentence using {HEAD ~ ...} and {TAIL ~ ...} notations. 
The output is the relation between the pair of entities as identified using actual tokens in the sentence. If there is no relation, the output in None.

The dataset could be used to train Seq2Seq models, i.e., marked sentence --> relation.
The dataset could also be used to train token classification models, i.e., tokenized marked sentence --> token tags.

For more information, please write to [email protected]