Model Card for gowitheflow/LASER-cubed-bert-base-unsup

Official model checkpoints of LA(SER)³ (LASER-cubed) from EMNLP 2023 paper "Length is a Curse and a Blessing for Document-level Semantics"

Model Summary

LASER-cubed-bert-base-unsup is an unsupervised model trained on wiki1M dataset. Without needing the training sets to have long texts, it provides surprising generalizability on long document retrieval.

• Developed by: Chenghao Xiao, Yizhi Li, G Thomas Hudson, Chenghua Lin, Noura Al-Moubayed
• Shared by: Chenghao Xiao
• Model type: BERT-base
• Language(s) (NLP): English
• Finetuned from model: BERT-base-uncased

Model Sources

• Github Repo: https://github.com/gowitheflow-1998/LA-SER-cubed
• Paper: https://aclanthology.org/2023.emnlp-main.86/

Usage

Use the model with Sentence Transformers:

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("gowitheflow/LASER-cubed-bert-base-unsup")

text = "LASER-cubed is a dope model - It generalizes to long texts without needing the training sets to have long texts."
representation = model.encode(text)

Evaluation

Evaluate it with the BEIR framework:

from beir.retrieval import models
from beir.datasets.data_loader import GenericDataLoader
from beir.retrieval.evaluation import EvaluateRetrieval
from beir.retrieval.search.dense import DenseRetrievalExactSearch as DRES

# download the datasets with BEIR original repo youself first
data_path = './datasets/arguana' 
corpus, queries, qrels = GenericDataLoader(data_folder=data_path).load(split="test")
model = DRES(models.SentenceBERT("gowitheflow/LASER-cubed-bert-base-unsup"), batch_size=512)
retriever = EvaluateRetrieval(model, score_function="cos_sim")
results = retriever.retrieve(corpus, queries)
ndcg, _map, recall, precision = retriever.evaluate(qrels, results, retriever.k_values)

Downstream Use

Information Retrieval

Out-of-Scope Use

The model is not for further fine-tuning to do other tasks (such as classification), as it's trained to do representation tasks with similarity matching.

Training Details

max seq 256, batch size 128, lr 3e-05, 1 epoch, 10% warmup, 1 A100.

Training Data

wiki 1M

Training Procedure

Please refer to the paper.

Evaluation

Results

BibTeX:

@inproceedings{xiao2023length,
  title={Length is a Curse and a Blessing for Document-level Semantics},
  author={Xiao, Chenghao and Li, Yizhi and Hudson, G and Lin, Chenghua and Al Moubayed, Noura},
  booktitle={Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing},
  pages={1385--1396},
  year={2023}
}