README.md

---
license: cc-by-sa-4.0
dataset_info:
- config_name: bc5cdr
  features:
  - name: entity
    dtype: string
  - name: label
    dtype: string
configs:
- config_name: bird
  data_files:
  - split: test
    path: data/bird/bird.tsv
- config_name: turney
  data_files:
  - split: test
    path: data/turney/turney.tsv
- config_name: conll
  data_files:
  - split: test
    path: data/conll/conll.tsv
- config_name: bc5cdr
  data_files:
  - split: test
    path: data/bc5cdr/bc5cdr.tsv
- config_name: autofj
  data_files:
  - split: test
    path: data/autofj/autofj.tsv
- config_name: ppdb
  data_files:
  - split: test
    path: data/ppdb/ppdb.tsv
- config_name: ppdb_filtered
  data_files:
  - split: test
    path: data/ppdb/ppdb_filtered.tsv
- config_name: yago
  data_files:
  - split: test
    path: data/yago/yago_test_samples.tsv
- config_name: umls
  data_files:
  - split: umls
    path: data/umls/umls_test_samples.tsv
- config_name: kb
  data_files:
  - split: umls
    path: data/kb/umls_kb.tsv
  - split: yago
    path: data/kb/yago_kb.tsv
language:
- en
tags:
- pearl benchmark
- phrase embeddings
- entity retrieval
- entity clustering
- fuzzy join
- entity matching
- string matching
- string similarity
size_categories:
- 1K<n<10K
---
# PEARL-Benchmark: A benchmark for evaluating phrase representations
[Learning High-Quality and General-Purpose Phrase Representations](https://arxiv.org/pdf/2401.10407.pdf). <br>
[Lihu Chen](https://chenlihu.com), [Gaël Varoquaux](https://gael-varoquaux.info/), [Fabian M. Suchanek](https://suchanek.name/). 
Accepted by EACL Findings 2024 <br>

Our PEARL Benchmark contains 9 phrase-level datasets of five types of tasks, which cover both the field of data science and natural language processing.

## Description
* **Paraphrase Classification**: PPDB and PPDBfiltered ([Wang et al., 2021](https://aclanthology.org/2021.emnlp-main.846/))
* **Phrase Similarity**: Turney ([Turney, 2012](https://arxiv.org/pdf/1309.4035.pdf)) and BIRD ([Asaadi et al., 2019](https://aclanthology.org/N19-1050/))
* **Entity Retrieval**: We constructed two datasets based on Yago ([Pellissier Tanon et al., 2020](https://hal-lara.archives-ouvertes.fr/DIG/hal-03108570v1)) and UMLS ([Bodenreider, 2004](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC308795/))
* **Entity Clustering**: CoNLL 03 ([Tjong Kim Sang, 2002](https://aclanthology.org/W02-2024/)) and BC5CDR ([Li et al., 2016](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4860626/))
* **Fuzzy Join**: AutoFJ benchmark ([Li et al., 2021](https://arxiv.org/abs/2103.04489)), which contains 50 diverse fuzzy-join datasets 
| - | PPDB | PPDB filtered |Turney|BIRD|YAGO|UMLS|CoNLL|BC5CDR|AutoFJ|
|-|-|-|-|-|-|-|-|-|-|
|Task|Paraphrase Classification|Paraphrase Classification|Phrase Similarity|Phrase Similarity|Entity Retrieval|Entity Retrieval|Entity Clustering|Entity Clustering|Fuzzy Join|
|Samples|23.4k|15.5k|2.2k|3.4k|10k|10k|5.0k|9.7k|50 subsets|
|Averaged Length|2.5|2.0|1.2|1.7|3.3|4.1|1.5|1.4|3.8|
|Metric|Acc|Acc|Acc|Pearson|Top-1 Acc|Top-1 Acc|NMI|NMI|Acc|

## Usage
```python
from datasets import load_dataset
turney_dataset = load_dataset("Lihuchen/pearl_benchmark", "turney", split="test")
```

## Evaluation
We offer a python script to evaluate your model: [eval.py](https://huggingface.co/datasets/Lihuchen/pearl_benchmark/blob/main/eval.py)
```python
python eval.py -batch_size 32
```

## Citation
```bibtex
@article{chen2024learning,
  title={Learning High-Quality and General-Purpose Phrase Representations},
  author={Chen, Lihu and Varoquaux, Ga{\"e}l and Suchanek, Fabian M},
  journal={arXiv preprint arXiv:2401.10407},
  year={2024}
}
```