sultan commited on
Commit
b10f36d
1 Parent(s): c216078

Uploading the Model

Browse files
Files changed (1) hide show
  1. README.md +58 -0
README.md ADDED
@@ -0,0 +1,58 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ BioM-Transformers: Building Large Biomedical Language Models with
2
+ BERT, ALBERT and ELECTRA
3
+
4
+ Abstract
5
+
6
+
7
+ The impact of design choices on the performance
8
+ of biomedical language models recently
9
+ has been a subject for investigation. In
10
+ this paper, we empirically study biomedical
11
+ domain adaptation with large transformer models
12
+ using different design choices. We evaluate
13
+ the performance of our pretrained models
14
+ against other existing biomedical language
15
+ models in the literature. Our results show that
16
+ we achieve state-of-the-art results on several
17
+ biomedical domain tasks despite using similar
18
+ or less computational cost compared to other
19
+ models in the literature. Our findings highlight
20
+ the significant effect of design choices on
21
+ improving the performance of biomedical language
22
+ models.
23
+
24
+ This model is fine-tuned on the SQuAD2.0 dataset. Fine-tuning the biomedical language model on the SQuAD dataset helps improve the score on the BioASQ challenge. If you plan to work with BioASQ or biomedical QA tasks, it's better to use this model over BioM-ELECTRA-Large. This model (TensorFlow version ) took the lead in the BioASQ9b-Factoid challenge under the name of (UDEL-LAB2). To see the full details of BioASQ9B results, please check this link http://participants-area.bioasq.org/results/9b/phaseB/ ( you need to register). Huggingface library doesn't implement a lire-wise decay feature, which affects the performance of the SQuAD dataset. The reported result of BioM-ELECTRA-SQuAD in our paper is 88.3 (F1) since we use ELECTRA open-source code with TF checkpoint, which uses lire-wise decay.
25
+
26
+
27
+ Evaluation results on SQuAD2.0 Dev Dataset
28
+ ```
29
+ exact = 84.33420365535248
30
+ f1 = 87.49354241889522
31
+ total = 11873
32
+ HasAns_exact = 80.43184885290148
33
+ HasAns_f1 = 86.75958656200127
34
+ HasAns_total = 5928
35
+ NoAns_exact = 88.22539949537426
36
+ NoAns_f1 = 88.22539949537426
37
+ NoAns_total = 5945
38
+ best_exact = 84.33420365535248
39
+ best_exact_thresh = 0.0
40
+ best_f1 = 87.49354241889522
41
+ best_f1_thresh = 0.0
42
+ epoch = 2.0
43
+
44
+ ```
45
+ @inproceedings{alrowili-shanker-2021-biom,
46
+ title = "{B}io{M}-Transformers: Building Large Biomedical Language Models with {BERT}, {ALBERT} and {ELECTRA}",
47
+ author = "Alrowili, Sultan and
48
+ Shanker, Vijay",
49
+ booktitle = "Proceedings of the 20th Workshop on Biomedical Language Processing",
50
+ month = jun,
51
+ year = "2021",
52
+ address = "Online",
53
+ publisher = "Association for Computational Linguistics",
54
+ url = "https://www.aclweb.org/anthology/2021.bionlp-1.24",
55
+ pages = "221--227",
56
+ abstract = "The impact of design choices on the performance of biomedical language models recently has been a subject for investigation. In this paper, we empirically study biomedical domain adaptation with large transformer models using different design choices. We evaluate the performance of our pretrained models against other existing biomedical language models in the literature. Our results show that we achieve state-of-the-art results on several biomedical domain tasks despite using similar or less computational cost compared to other models in the literature. Our findings highlight the significant effect of design choices on improving the performance of biomedical language models.",
57
+ }
58
+ ```