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
andSSCI-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.
- 2022-03-24 : SSCI-BERT and SSCI-SciBERT has been put forward for the first time.
- 2022-06-09 : The paper for SsciBERT has been submitted to arxiv(https://arxiv.org/abs/2206.04510).
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")
- Datasets for Pretraining
Abstract data for SSCI papers published in WOS from 1986 to 2021, with duplicate entries and missing abstract entries removed. Each line of the corpus is an abstract of a paper, consisting of 2964743 abstracts from 3250 journals.
- Datasets for JCR Social Science Disciplines Classification
SSCI journal papers published between 2006 and 2020 are then subject-matched to the list of JCR disciplines based on the journal's ISSN number, comprising a total of 46 disciplines. Extract 500 pieces of data from each discipline as a dataset for the JCR Social Science Discipline Classification task. A total of 23,000 header data and 22,000 summary data were obtained.
- Datasets for Identifying Abstract Structures
SSCI journal abstracts published between 2008 and 2020 obtained a total of 1378276 construction notes according to the functional structure of the sentence-by-sentence annotation abstracts in the five categories of background, purpose, methodology, results and conclusions (BPMRC).
- Datasets for Software Entities Recognition in Scientometrics
Using full-text data published in Scientometrics from 2010 to 2020, software entities in the dataset were manually annotated, and a total of 13,269 software entities were identified.
dataset examples can be find in 'datasets/examples' if you want all datasets for researches, please email us and we will provide for free.()
- The version of the model we provide is
PyTorch
.
- Download directly through Huggingface's official website.
- KM4STfulltext/SSCI-BERT-e2
- KM4STfulltext/SSCI-SciBERT-e2
- KM4STfulltext/SSCI-BERT-e4
- KM4STfulltext/SSCI-SciBERT-e4
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 |
- 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.
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 |
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 |
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 |
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 |
- If our content is helpful for your research work, please quote our research in your article.
- If you want to quote our research, https://link.springer.com/article/10.1007/s11192-022-04602-4
- 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.
- 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))