SBERT's Interpretability: can we better understand the final results?

user_9151070 0
Viewed 80

I'm familiar with SBERT and its pre-trained models and they are amazing! But at the same time, I want to understand how the results are calculated, and I can't find anything more specific in their website. For example, I have a document and I want to find other documents that are similar to it. I used 2 documents containing 200-250 words each (I changed the model.max_seq_length to 350 so the model can handle bigger texts), and in the end we can see that the cosine-similarity is 0.79. Is that all we can see? Is there a way to extract the main phrases/keywords that made the model return this high value of similarity?

Thanks in advance!

sentence-transformers transformer-model nlp
1 Answers

Have you tried to make either a simple word-count-comparison between the two documents and other random documents? Or a tf-idf, if the two documents are part of a bigger corpus?

Another thing you can do, is to look inside the "stored_embeddings" matrix (see code below and here) in which SBERT encodes your sentences (e.g. for 20.000 documents you'll get a 20.000*384 matrix), after having saved it into a pickle file like:

from sentence_transformers import SentenceTransformer
import pickle

embeddings = model.encode(sentences)

with open('embeddings.pkl', "wb") as fOut:
    pickle.dump({'sentences': sentences, 'embeddings': embeddings}, fOut, protocol=pickle.HIGHEST_PROTOCOL)

with open('embeddings.pkl', "rb") as fIn:
    stored_data = pickle.load(fIn)
    stored_embeddings = stored_data['embeddings']

The stored embeddings variable can be handled as a numpy matrix and can therefore be (for example) indexed to access single elements. By looking at the values of the single 384 dimensions (to do this, you can go column by column but in case of a big matrix I suggest you not to .enumerate(), it'll take forever) and compare the values that the two documents take in one precise dimension. You can see which dimension has the highest values or variance, for example.

I'm not saying it'll be interpretable what you'll find, but at least you can try and see what you find.

user_144909150
Related