Hi, I wanted to understand why RASA decided to use starspace (tensorflow embedding in RASA) rather than using LSTM for classification?
What do you mean?
starspace is the way to construct a loss function, while LSTM is an architecture of a neural network. They don’t contradict each other. We didn’t use LSTM in nlu classification because we didn’t see performance gain
Let’s take these below lines:
- I am happy not sad
- I am sad not happy
Would starspace be able to classify these two as different intents?
LSTM being sequence to sequence model, i would expect that it should be able to classify above two sentences as different intents.
PS: starspace rather than being I loss function, I thought it is way of creating embedding just like GloVe or word2vec or fasText. StarSpace has the advantage of handling featured labels, and bringing variety of things in single vector space so that those can be compared. Is my understanding not correct?
the fact that starspace idea is used to construct a loss function doesn’t contradict the fact that it can be used to create embeddings