|
--- |
|
pipeline_tag: text-classification |
|
inference: false |
|
language: en |
|
tags: |
|
- transformers |
|
--- |
|
|
|
# Prompsit/paraphrase-bert-en |
|
|
|
This model allows to evaluate paraphrases for a given phrase. |
|
We have fine-tuned this model from pretrained "bert-base-uncased". |
|
|
|
|
|
|
|
# How to usage |
|
|
|
The model answer the following question: Is "phrase B" paraphrases of "phrase A". |
|
Please note that we're considering phrases instead of sentences. Therefore, we must take into account that the model doesn't expect to find punctuation marks or long pieces of text. |
|
|
|
Resulting probabilities correspond to classes: |
|
* 0: Not a paraphrase |
|
* 1: It's a paraphrase |
|
|
|
|
|
You can usage the model like this: |
|
|
|
``` |
|
import torch |
|
from transformers import AutoTokenizer, AutoModelForSequenceClassification |
|
|
|
tokenizer = AutoTokenizer.from_pretrained("Prompsit/paraphrase-bert-en") |
|
model = AutoModelForSequenceClassification.from_pretrained("Prompsit/paraphrase-bert-en") |
|
|
|
input = tokenizer('may be addressed','could be included',return_tensors='pt') |
|
logits = model(**input).logits |
|
soft = torch.nn.Softmax(dim=1) |
|
print(soft(logits)) |
|
``` |
|
Output of previous code is: |
|
``` |
|
tensor([[0.1592, 0.8408]], grad_fn=<SoftmaxBackward>) |
|
``` |
|
As the probability of 1 is 0.84, we can conclude from the previous example that "could be included" is paraphrase of "may be included". |
|
|
