metadata
datasets:
- relbert/semeval2012_relational_similarity
model-index:
- name: relbert/relbert-roberta-large-nce-semeval2012-2
results:
- task:
name: Relation Mapping
type: sorting-task
dataset:
name: Relation Mapping
args: relbert/relation_mapping
type: relation-mapping
metrics:
- name: Accuracy
type: accuracy
value: 0.8303968253968254
- task:
name: Analogy Questions (SAT full)
type: multiple-choice-qa
dataset:
name: SAT full
args: relbert/analogy_questions
type: analogy-questions
metrics:
- name: Accuracy
type: accuracy
value: 0.7192513368983957
- task:
name: Analogy Questions (SAT)
type: multiple-choice-qa
dataset:
name: SAT
args: relbert/analogy_questions
type: analogy-questions
metrics:
- name: Accuracy
type: accuracy
value: 0.7091988130563798
- task:
name: Analogy Questions (BATS)
type: multiple-choice-qa
dataset:
name: BATS
args: relbert/analogy_questions
type: analogy-questions
metrics:
- name: Accuracy
type: accuracy
value: 0.8043357420789328
- task:
name: Analogy Questions (Google)
type: multiple-choice-qa
dataset:
name: Google
args: relbert/analogy_questions
type: analogy-questions
metrics:
- name: Accuracy
type: accuracy
value: 0.948
- task:
name: Analogy Questions (U2)
type: multiple-choice-qa
dataset:
name: U2
args: relbert/analogy_questions
type: analogy-questions
metrics:
- name: Accuracy
type: accuracy
value: 0.6798245614035088
- task:
name: Analogy Questions (U4)
type: multiple-choice-qa
dataset:
name: U4
args: relbert/analogy_questions
type: analogy-questions
metrics:
- name: Accuracy
type: accuracy
value: 0.6643518518518519
- task:
name: Analogy Questions (ConceptNet Analogy)
type: multiple-choice-qa
dataset:
name: ConceptNet Analogy
args: relbert/analogy_questions
type: analogy-questions
metrics:
- name: Accuracy
type: accuracy
value: 0.4865771812080537
- task:
name: Analogy Questions (TREX Analogy)
type: multiple-choice-qa
dataset:
name: TREX Analogy
args: relbert/analogy_questions
type: analogy-questions
metrics:
- name: Accuracy
type: accuracy
value: 0.6338797814207651
- task:
name: Analogy Questions (NELL-ONE Analogy)
type: multiple-choice-qa
dataset:
name: NELL-ONE Analogy
args: relbert/analogy_questions
type: analogy-questions
metrics:
- name: Accuracy
type: accuracy
value: 0.6633333333333333
- task:
name: Lexical Relation Classification (BLESS)
type: classification
dataset:
name: BLESS
args: relbert/lexical_relation_classification
type: relation-classification
metrics:
- name: F1
type: f1
value: 0.9169805635076088
- name: F1 (macro)
type: f1_macro
value: 0.9133613159985977
- task:
name: Lexical Relation Classification (CogALexV)
type: classification
dataset:
name: CogALexV
args: relbert/lexical_relation_classification
type: relation-classification
metrics:
- name: F1
type: f1
value: 0.8643192488262911
- name: F1 (macro)
type: f1_macro
value: 0.709680204738525
- task:
name: Lexical Relation Classification (EVALution)
type: classification
dataset:
name: BLESS
args: relbert/lexical_relation_classification
type: relation-classification
metrics:
- name: F1
type: f1
value: 0.6782231852654388
- name: F1 (macro)
type: f1_macro
value: 0.665196173208286
- task:
name: Lexical Relation Classification (K&H+N)
type: classification
dataset:
name: K&H+N
args: relbert/lexical_relation_classification
type: relation-classification
metrics:
- name: F1
type: f1
value: 0.9568060095986646
- name: F1 (macro)
type: f1_macro
value: 0.8745909398702613
- task:
name: Lexical Relation Classification (ROOT09)
type: classification
dataset:
name: ROOT09
args: relbert/lexical_relation_classification
type: relation-classification
metrics:
- name: F1
type: f1
value: 0.9150736446255092
- name: F1 (macro)
type: f1_macro
value: 0.9142555280970402
relbert/relbert-roberta-large-nce-semeval2012-2
RelBERT based on roberta-large fine-tuned on relbert/semeval2012_relational_similarity (see the relbert for more detail of fine-tuning).
This model achieves the following results on the relation understanding tasks:
- Analogy Question (dataset, full result):
- Accuracy on SAT (full): 0.7192513368983957
- Accuracy on SAT: 0.7091988130563798
- Accuracy on BATS: 0.8043357420789328
- Accuracy on U2: 0.6798245614035088
- Accuracy on U4: 0.6643518518518519
- Accuracy on Google: 0.948
- Accuracy on ConceptNet Analogy: 0.4865771812080537
- Accuracy on T-Rex Analogy: 0.6338797814207651
- Accuracy on NELL-ONE Analogy: 0.6633333333333333
- Lexical Relation Classification (dataset, full result):
- Micro F1 score on BLESS: 0.9169805635076088
- Micro F1 score on CogALexV: 0.8643192488262911
- Micro F1 score on EVALution: 0.6782231852654388
- Micro F1 score on K&H+N: 0.9568060095986646
- Micro F1 score on ROOT09: 0.9150736446255092
- Relation Mapping (dataset, full result):
- Accuracy on Relation Mapping: 0.8303968253968254
Usage
This model can be used through the relbert library. Install the library via pip
pip install relbert
and activate model as below.
from relbert import RelBERT
model = RelBERT("relbert/relbert-roberta-large-nce-semeval2012-2")
vector = model.get_embedding(['Tokyo', 'Japan']) # shape of (n_dim, )
Training hyperparameters
- model: roberta-large
- max_length: 64
- epoch: 10
- batch: 32
- random_seed: 2
- lr: 5e-06
- lr_warmup: 10
- aggregation_mode: average_no_mask
- data: relbert/semeval2012_relational_similarity
- data_name: None
- exclude_relation: None
- split: train
- split_valid: validation
- loss_function: nce
- classification_loss: False
- loss_function_config: {'temperature': 0.05, 'num_negative': 100, 'num_positive': 10}
- augment_negative_by_positive: True
See the full configuration at config file.
Reference
If you use any resource from RelBERT, please consider to cite our paper.
@inproceedings{ushio-etal-2021-distilling,
title = "Distilling Relation Embeddings from Pretrained Language Models",
author = "Ushio, Asahi and
Camacho-Collados, Jose and
Schockaert, Steven",
booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2021",
address = "Online and Punta Cana, Dominican Republic",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.emnlp-main.712",
doi = "10.18653/v1/2021.emnlp-main.712",
pages = "9044--9062",
abstract = "Pre-trained language models have been found to capture a surprisingly rich amount of lexical knowledge, ranging from commonsense properties of everyday concepts to detailed factual knowledge about named entities. Among others, this makes it possible to distill high-quality word vectors from pre-trained language models. However, it is currently unclear to what extent it is possible to distill relation embeddings, i.e. vectors that characterize the relationship between two words. Such relation embeddings are appealing because they can, in principle, encode relational knowledge in a more fine-grained way than is possible with knowledge graphs. To obtain relation embeddings from a pre-trained language model, we encode word pairs using a (manually or automatically generated) prompt, and we fine-tune the language model such that relationally similar word pairs yield similar output vectors. We find that the resulting relation embeddings are highly competitive on analogy (unsupervised) and relation classification (supervised) benchmarks, even without any task-specific fine-tuning. Source code to reproduce our experimental results and the model checkpoints are available in the following repository: https://github.com/asahi417/relbert",
}