0
$\begingroup$

I used sentence bert to embed sentences from this tutorial https://www.sbert.net/docs/pretrained_models.html

from sentence_transformers import SentenceTransformer, util
model = SentenceTransformer('all-mpnet-base-v2')

This is the event triples t I forgot to concat into sentences,

[('U.S. stock index futures', 'points to', 'start'),
 ('U.S. stock index futures', 'points to', 'higher start')]

model.encode(t) returns a 2d array of shape (2,768), with two idential 768-dimension vectors, and its value is different from both model.encode('U.S. stock index futures') and model.encode('U.S. stock index futures points to start'). What could possibly have it returned?

It is the same situation for other models on huggingface such as https://huggingface.co/sentence-transformers/stsb-distilbert-base

$\endgroup$

1 Answer 1

2
$\begingroup$

There are two valid inputs to MPNet's tokenizer:

Union[TextInputSequence, Tuple[InputSequence, InputSequence]]

When you give a list of tuples as an input, from each tuple only the first two sentences, i.e. "U.S. stock index futures" and "points to" are used to encode, similar to BERT's next sentence prediction pre-training task.

This is tokenized and converted to input_ids with special tokens as follows:

['<s>', 'u', '.', 's', '.', 'stock', 'index', 'futures', '</s>', '</s>', 'points', 'to', '</s>']

<s> indicates start of sentence and </s> indicates end of sentence. I'm not sure about the </s> before 'points', but this is the output from the model's tokenizer. As such NLU models can work with one or two sentences together, these special tokens are added to identify how many sentences are in the model and they help to separate the sentences.

Thus, you get two same vectors from

model.encode(t)

If you want the same vector without using a tuple, the encode function should contain the following input string:

"U.S. stock index futures </s> </s> points to"

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.