Sentiment analysis alternatives

Hello fellows,

I’ve been recently working on a project in NLP. It’s main goal is to do the sentiment analysis on non-english dataset. I’ve been trying to apply model from one of the lessons (on IMDB dataset) but it does not generalize well (I’ve got merely 53% accuracy on dataset with two labels).

Do you guys know any alternative, promising, state-of-the-art approaches that is being used for sentiment analysis? Would appreciate any propositions.

Do you use the model trained with IMDB dataset?
If true, then the approach would not work well because the co-occurrence/sequence patterns that RNN models learn should be different from language to language.

If you train the model from scratch in order to fit your model to your dataset, I’m not sure though, changing hyper-parameters such as number of layers/units of RNN is one alternative.

Yeah, I trained my model from scratch. Maybe this model is just too large to generalize well on my dataset. I’ll try to train smaller one and see if this improves results.

Great, I’ve heard that the imbalance between classes would harm your training.

Is your dataset balanced?

best