demdecuong/stroke_sup_simcse

This is finetune version of SimCSE: Simple Contrastive Learning of Sentence Embeddings

• Train supervised on 100K triplet samples samples related to stroke domain from : stroke books, quora medical, quora's stroke, quora's general and human annotates.
• Positive sentences are generated by paraphrasing and back-translate.
• Negative sentences are randomly selected in general domain.

Extract sentence representation

from transformers import AutoTokenizer, AutoModel  
tokenizer = AutoTokenizer.from_pretrained("demdecuong/stroke_sup_simcse")
model = AutoModel.from_pretrained("demdecuong/stroke_sup_simcse")

text = "What are disease related to red stroke's causes?"
inputs = tokenizer(text, return_tensors='pt')
outputs = model(**inputs)[1]

Build up embedding for database

database = [
    'What is the daily checklist for stroke returning home',
    'What are some tips for stroke adapt new life',
    'What  should I consider when using nursing-home care'
]

embedding = torch.zeros((len(database),768))

for i in range(len(database)):
  inputs = tokenizer(database[i], return_tensors="pt")
  outputs = model(**inputs)[1]
  embedding[i] = outputs

print(embedding.shape)

Result

On our company's PoC project, the testset contains positive/negative pairs of matching question related to stroke from human-generation.

• SimCSE supervised + 100k : Train on 100K triplet samples contains : medical, stroke and general domain
• SimCSE supervised + 42k : Train on 42K triplet samples contains : medical, stroke domain

Model	Top-1 Accuracy
SimCSE supervised (author)	75.83
SimCSE unsupervised (ours)	76.66
SimCSE supervised + 100k (ours)	73.33
SimCSE supervised + 42k (ours)	75.83