SPECTER2

SPECTER2 is the successor to SPECTER and is capable of generating task specific embeddings for scientific tasks when paired with 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.

Dec 2023 Update:

Model usage updated to be compatible with latest versions of transformers and adapters (newly released update to adapter-transformers) libraries.

Aug 2023 Update:

1. The SPECTER2 Base and proximity adapter models have been renamed in Hugging Face based upon usage patterns as follows:

Old Name	New Name
allenai/specter2	allenai/specter2_base
allenai/specter2_proximity	allenai/specter2

2. We have a parallel version (termed aug2023refresh) where the base transformer encoder version is pre-trained on a collection of newer papers (published after 2018). However, for benchmarking purposes, please continue using the current version.

Adapter allenai/specter2_classification for allenai/specter2_base

An adapter for the allenai/specter2_base model that was trained on the allenai/scirepeval dataset.

This adapter was created for usage with the adapters library.

Adapter Usage

First, install adapters:

pip install -U adapters

Note: adapters is built as an add-on transformers that acts as a drop-in replacement with adapter support. More

Now, the adapter can be loaded and activated like this:

from adapters import AutoAdapterModel

model = AutoAdapterModel.from_pretrained("allenai/specter2_base")
adapter_name = model.load_adapter("allenai/specter2_classification", source="hf", set_active=True)

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 (Retrieval)
• 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

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

Uses

Direct Use

Model	Name and HF link	Description
Proximity*	allenai/specter2	Encode papers as queries and candidates eg. Link Prediction, Nearest Neighbor Search
Adhoc Query	allenai/specter2_adhoc_query	Encode short raw text queries for search tasks. (Candidate papers can be encoded with the proximity adapter)
Classification	allenai/specter2_classification	Encode papers to feed into linear classifiers as features
Regression	allenai/specter2_regression	Encode papers to feed into linear regressors as features

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

from transformers import AutoTokenizer
from adapters import AutoAdapterModel

# load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained('allenai/specter2_base')

#load base model
model = AutoAdapterModel.from_pretrained('allenai/specter2_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_classification", source="hf", load_as="classification", set_active=True)

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

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

Training Details

Training Data

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.

Model	SciRepEval In-Train	SciRepEval Out-of-Train	SciRepEval Avg	MDCR(MAP, Recall@5)
BM-25	n/a	n/a	n/a	(33.7, 28.5)
SPECTER	54.7	57.4	68.0	(30.6, 25.5)
SciNCL	55.6	57.8	69.0	(32.6, 27.3)
SciRepEval-Adapters	61.9	59.0	70.9	(35.3, 29.6)
SPECTER2 Base	56.3	73.6	69.1	(38.0, 32.4)
SPECTER2-Adapters	62.3	59.2	71.2	(38.4, 33.0)

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

@inproceedings{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},
  booktitle={Conference on Empirical Methods in Natural Language Processing},
  year={2022},
  url={https://api.semanticscholar.org/CorpusID:254018137}
}