Preprocessing before using `model.encode(sentences)`

First, I’m a huge fan of this project and I use it very actively - thanks guys!

To my question.

1) Cased vs. uncased Assume I running the following code:

from sentence_transformers import SentenceTransformer
model = SentenceTransformer('stsb-bert-base')

#Our sentences we like to encode
sentences = ['This framework generates embeddings for each input sentence',
    'Sentences are passed as a list of string.',
    'The quick brown fox jumps over the lazy dog.']

#Sentences are encoded by calling model.encode()
embeddings = model.encode(sentences)

Since the underlying pre-trained model is uncased but the sentences are cased, is this code doing some “uncasing” under the hood?

2) Text preprocessing More generally, when I preprocess my sentences (lemmatize, remove stop words, remove special characters, …), I tend to get different results compared to when I don’t. As a rule-of-thumb, would you recommend that I preprocess the sentences myself (perhaps for consistency across methods) or should I leave this to your under-the-hood preprocessing because you shouldn’t, say, remove stop words?

Looking forward to your answers!

[nreimers]

1. Yes, the text is lower cased before processing for uncased models

2. Specific preprocessing is not needed. Only when you have a lot of noise in your sentences, like URLs, it can make sense to remove them. Lemmatization and stop word removal is not needed

[muhlbach]

Thanks, this was really helpful!