Very basic of mathematics: Cosine similarity (see calculus of variation) have important role in semantic similarity between texts. Have a look at the concept https://shyambhu20.blogspot.com/2020/10/calculate-word-similarity-spacy-nlp.html

OK, thank you.

So it looks like SpaCy only uses the cosine distance, which I already know how to calculate, but I was thinking that SpaCy used a CNN before to prepare the information for the calculation like, embed the words and train a neural network to improve the calculation.

I will read your recommendation text in detail, but I want to use a neural network to calculate the semantic similarity, because the cosine distance I have already calculated and I want another algorithm or technique to compare it to my previous results.

Im guess I have to try TesnorFlow that maybe use a Neural Netowork to calculate that...

Thank you for your help.

Hi Josivan Pereira da Silva ,

I just took a quick look at the SpaCy source code repo on GitHub and it seems you and Muhammad Ali are both right: the .similarity() method just applies simple cosine similarity to two text fragments (words, sentences, paragraphs etc.) based on their representations in a given vector space model (VSM). You can however, have SpaCy change the default VSM and replace it with a model you trained yourself or with a pre-trained model you gathered elsewhere (see the paragraph on 'Adding Word Vectors' in this documentation: https://spacy.io/usage/linguistic-features#vectors-similarity). Training such a vector could, as you mention, also involve a CNN - though, the state-of-the-art language models are currently rather the RNN-based transformer models like the famous BERT.

The best place to browse for pre-trained transformer-based language models suitable for your task is currently probably the hugging face library: https://huggingface.co/models

And conveniently SpaCy seems to already have a wrapper for some of those models:

https://spacy.io/universe/project/spacy-transformers

Good luck with your research!

Hi, Annika Hamachers . Thank you so much, I will study this subject by your suggestions.

Thanks!

see this link below

What theory does SpaCy / Python use behind the code to calculate the semantic similarity between texts?