I have two groups of texts that are very similar (e.g. reviews written on fridays and reviews written on mondays), and I want to build a LSTM that can classify them into positive and negative reviews.
The question is then how to input the categories (Monday/Sunday) such that the model also takes those into account.
- Make two models (which I would rather not given the text similarities)
- Add a variable to the start of the tokenized texts (0 and 1 for the Friday/Monday). This will likely have very little impacts due to their small weight in the entire context of the LSTM.
- Like (2.) but give these variables a higher weight
- Just hope that the network finds out itself (should one balance the categories friday/monday and the categories positive/negative for each?).
It is quite simple in most other non-sequential models, as one e.g. with a bag-of-words and random forest just could force the dummy variable to be in all trees, or one could normalize the variables, such that the variable would have a weight just as high, as all the text combinedly.
Do you have any suggestions?