I have a datadet with many phrases which I would like to embed them from scratch. I dont want the cosine of the words in order to get a phrase embedding, this is because the phrases may appear in a different enviroment and I want to embed the two words together, or the tree words together in their own envoroment.

Is this possible?

If yes how exactly?

Thank You in advance.


2 Answers 2


There are several ways word-embeddings are trained, however most of them require a ton of data. They usually involve learning vector representations that are useful for some self-supervised objective, which all tend to be pretty data-hungry.

  • word2vec (and variants) learn representations by training a model to use those representations to predict adjacent words
  • Approaches like ELMo and BERT use intermediate representations from a language model, which are pretrained on large text corpora

If you have a large enough dataset you could train new domain-specific embeddings from scratch, but it probably be more effective and much easier to finetune existing embeddings (i.e., initialize on models/representations and train on your domain data).

See: this post for finetuning word2vec, for example.


If you want to train on phrases, then you will have to devise your tokenizer that way. Pseudo code will look like:

 1. Write tokenizer that maps phrases to a single token  
 2. Using tokenizer generate vocabulary from text corpus
   a. this will assign multi word phrase a single token. for example the phrase "Natural language Processing" will be considered as one token.
 3. Once you have devised tokenizer that way, the remaining process will look same.

For example if you are training using gensim, Following line will use your custom tokenizer instead of default gensim function.

yield utils.simple_preprocess(line)

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.