******Update******

This update introduces a new set of SPECTER2 models with the base transformer encoder pre-trained on an extended citation dataset containing more recent papers. For benchmarking purposes please use the existing SPECTER2 models w/o the aug2023refresh suffix.

SPECTER2 (Base)

SPECTER2 is the successor to SPECTER and is capable of generating task specific embeddings for scientific tasks when paired with adapters. This is the base model to be used along with the adapters. Given the combination of title and abstract of a scientific paper or a short texual query, the model can be used to generate effective embeddings to be used in downstream applications.

Note:For general embedding purposes, please use allenai/specter2.

To get the best performance on a downstream task type please load the associated adapter with the base model as in the example below.

Model Details

Model Description

SPECTER2 has been trained on over 6M triplets of scientific paper citations, which are available here. Post that it is trained with additionally attached task format specific adapter modules on all the SciRepEval training tasks.

Task Formats trained on:

• Classification
• Regression
• Proximity
• Adhoc Search

It builds on the work done in SciRepEval: A Multi-Format Benchmark for Scientific Document Representations and we evaluate the trained model on this benchmark as well.

• Developed by: Amanpreet Singh, Mike D'Arcy, Arman Cohan, Doug Downey, Sergey Feldman
• Shared by : Allen AI
• Model type: bert-base-uncased + adapters
• License: Apache 2.0
• Finetuned from model: allenai/scibert.

Model Sources

<!-- Provide the basic links for the model. -->

• Repository: https://github.com/allenai/SPECTER2
• Paper: https://api.semanticscholar.org/CorpusID:254018137
• Demo: Usage

Uses

<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->

Direct Use

*Proximity model should suffice for downstream task types not mentioned above

from transformers import AutoTokenizer, AutoModel

# load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained('allenai/specter2_aug2023refresh_base')

#load base model
model = AutoModel.from_pretrained('allenai/specter2_aug2023refresh_base')

#load the adapter(s) as per the required task, provide an identifier for the adapter in load_as argument and activate it
model.load_adapter("allenai/specter2_aug2023refresh", source="hf", load_as="proximity", set_active=True)
#other possibilities: allenai/specter2_aug2023refresh_<classification|regression|adhoc_query>

papers = [{'title': 'BERT', 'abstract': 'We introduce a new language representation model called BERT'},
          {'title': 'Attention is all you need', 'abstract': ' The dominant sequence transduction models are based on complex recurrent or convolutional neural networks'}]

# concatenate title and abstract
text_batch = [d['title'] + tokenizer.sep_token + (d.get('abstract') or '') for d in papers]
# preprocess the input
inputs = self.tokenizer(text_batch, padding=True, truncation=True,
                                   return_tensors="pt", return_token_type_ids=False, max_length=512)
output = model(**inputs)
# take the first token in the batch as the embedding
embeddings = output.last_hidden_state[:, 0, :]

Downstream Use

<!-- This section is for the model use when fine-tuned for a task, or when plugged into a larger ecosystem/app -->

For evaluation and downstream usage, please refer to https://github.com/allenai/scirepeval/blob/main/evaluation/INFERENCE.md.

Training Details

Training Data

<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->

The base model is trained on citation links between papers and the adapters are trained on 8 large scale tasks across the four formats. All the data is a part of SciRepEval benchmark and is available here.

The citation link are triplets in the form

{"query": {"title": ..., "abstract": ...}, "pos": {"title": ..., "abstract": ...}, "neg": {"title": ..., "abstract": ...}}

consisting of a query paper, a positive citation and a negative which can be from the same/different field of study as the query or citation of a citation.

Training Procedure

Please refer to the SPECTER paper.

Training Hyperparameters

The model is trained in two stages using SciRepEval:

• Base Model: First a base model is trained on the above citation triplets. batch size = 1024, max input length = 512, learning rate = 2e-5, epochs = 2 warmup steps = 10% fp16
• Adapters: Thereafter, task format specific adapters are trained on the SciRepEval training tasks, where 600K triplets are sampled from above and added to the training data as well. batch size = 256, max input length = 512, learning rate = 1e-4, epochs = 6 warmup = 1000 steps fp16

Evaluation

We evaluate the model on SciRepEval, a large scale eval benchmark for scientific embedding tasks which which has [SciDocs] as a subset. We also evaluate and establish a new SoTA on MDCR, a large scale citation recommendation benchmark.

Please cite the following works if you end up using SPECTER2:

SPECTER paper:

@inproceedings{specter2020cohan,
  title={{SPECTER: Document-level Representation Learning using Citation-informed Transformers}},
  author={Arman Cohan and Sergey Feldman and Iz Beltagy and Doug Downey and Daniel S. Weld},
  booktitle={ACL},
  year={2020}
}

SciRepEval paper

@article{Singh2022SciRepEvalAM,
  title={SciRepEval: A Multi-Format Benchmark for Scientific Document Representations},
  author={Amanpreet Singh and Mike D'Arcy and Arman Cohan and Doug Downey and Sergey Feldman},
  journal={ArXiv},
  year={2022},
  volume={abs/2211.13308}
}