SSCI-BERT: A pretrained language model for social scientific text

Introduction

The research for social science texts needs the support natural language processing tools.

The pre-trained language model has greatly improved the accuracy of text mining in general texts. At present, there is an urgent need for a pre-trained language model specifically for the automatic processing of scientific texts in social science.

We used the abstract of social science research as the training set. Based on the deep language model framework of BERT, we constructed SSCI-BERT and SSCI-SciBERT pre-training language models by transformers/run_mlm.py.

We designed four downstream tasks of Text Classification on different social scientific article corpus to verify the performance of the model.

SSCI-BERT and SSCI-SciBERT are trained on the abstract of articles published in SSCI journals from 1986 to 2021. The training set involved in the experiment included a total of 503910614 words.
Based on the idea of Domain-Adaptive Pretraining, SSCI-BERT and SSCI-SciBERT combine a large amount of abstracts of scientific articles based on the BERT structure, and continue to train the BERT and SSCI-SciBERT models respectively to obtain pre-training models for the automatic processing of Social science research texts.

News

2022-03-24 : SSCIBERT and SSCI-SciBERT has been put forward for the first time.

How to use

Huggingface Transformers

The from_pretrained method based on Huggingface Transformers can directly obtain SSCI-BERT and SSCI-SciBERT models online.

SSCI-BERT

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("KM4STfulltext/SSCI-BERT-e2")

model = AutoModel.from_pretrained("KM4STfulltext/SSCI-BERT-e2")

SSCI-SciBERT

from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("KM4STfulltext/SSCI-SciBERT-e2")

model = AutoModel.from_pretrained("KM4STfulltext/SSCI-SciBERT-e2")

Download Models

The version of the model we provide is PyTorch.

From Huggingface

Download directly through Huggingface's official website.
KM4STfulltext/SSCI-BERT-e2
KM4STfulltext/SSCI-SciBERT-e2
KM4STfulltext/SSCI-BERT-e4
KM4STfulltext/SSCI-SciBERT-e4

From Google Drive

We have put the model on Google Drive for users.

Model	DATASET(year)	Base Model
SSCI-BERT-e2	1986-2021	Bert-base-cased
SSCI-SciBERT-e2 (recommended)	1986-2021	Scibert-scivocab-cased
SSCI-BERT-e4	1986-2021	Bert-base-cased
SSCI-SciBERT-e4	1986-2021	Scibert-scivocab-cased

Evaluation & Results

We use SSCI-BERT and SSCI-SciBERT to perform Text Classificationon different social science research corpus. The experimental results are as follows. Relevant data sets are available for download in the Verification task datasets folder of this project.

JCR Title Classify Dataset

Model	accuracy	macro avg	weighted avg
Bert-base-cased	28.43	22.06	21.86
Scibert-scivocab-cased	38.48	33.89	33.92
SSCI-BERT-e2	40.43	35.37	35.33
SSCI-SciBERT-e2	41.35	37.27	37.25
SSCI-BERT-e4	40.65	35.49	35.40
SSCI-SciBERT-e4	41.13	36.96	36.94
Support	2300	2300	2300

JCR Abstract Classify Dataset

Model	accuracy	macro avg	weighted avg
Bert-base-cased	48.59	42.8	42.82
Scibert-scivocab-cased	55.59	51.4	51.81
SSCI-BERT-e2	58.05	53.31	53.73
SSCI-SciBERT-e2	59.95	56.51	57.12
SSCI-BERT-e4	59.00	54.97	55.59
SSCI-SciBERT-e4	60.00	56.38	56.90
Support	2200	2200	2200

JCR Mixed Titles and Abstracts Dataset

Model	accuracy	macro avg	weighted avg
Bert-base-cased	58.24	57.27	57.25
Scibert-scivocab-cased	59.58	58.65	58.68
SSCI-BERT-e2	60.89	60.24	60.30
SSCI-SciBERT-e2	60.96	60.54	60.51
SSCI-BERT-e4	61.00	60.48	60.43
SSCI-SciBERT-e4	61.24	60.71	60.75
Support	4500	4500	4500

SSCI Abstract Structural Function Recognition (Classify Dataset)

	Bert-base-cased	SSCI-BERT-e2	SSCI-BERT-e4	support
B	63.77	64.29	64.63	224
P	53.66	57.14	57.99	95
M	87.63	88.43	89.06	323
R	86.81	88.28	88.47	419
C	78.32	79.82	78.95	316
accuracy	79.59	80.9	80.97	1377
macro avg	74.04	75.59	75.82	1377
weighted avg	79.02	80.32	80.44	1377
	Scibert-scivocab-cased	SSCI-SciBERT-e2	SSCI-SciBERT-e4	support
B	69.98	70.95	70.95	224
P	58.89	60.12	58.96	95
M	89.37	90.12	88.11	323
R	87.66	88.07	87.44	419
C	80.7	82.61	82.94	316
accuracy	81.63	82.72	82.06	1377
macro avg	77.32	78.37	77.68	1377
weighted avg	81.6	82.58	81.92	1377

Cited

If our content is helpful for your research work, please quote our research in your article.
https://link.springer.com/article/10.1007/s11192-022-04602-4

Disclaimer

The experimental results presented in the report only show the performance under a specific data set and hyperparameter combination, and cannot represent the essence of each model. The experimental results may change due to random number seeds and computing equipment.
Users can use the model arbitrarily within the scope of the license, but we are not responsible for the direct or indirect losses caused by using the content of the project.

Acknowledgment

SSCI-BERT was trained based on [BERT-Base-Cased](google-research/bert: TensorFlow code and pre-trained models for BERT (github.com)).
SSCI-SciBERT was trained based on [scibert-scivocab-cased](allenai/scibert: A BERT model for scientific text. (github.com))

KM4STfulltext
/

SSCI-SciBERT-e2