Rasa: Intent classification using keras network

[matteoredaelli]

could be useful adding Tensorflow ML? see https://chatbotsmagazine.com/contextual-chat-bots-with-tensorflow-4391749d0077

regards
Matteo

[tmbo]

duplicates #423

[matteoredaelli]

..but adding tensorflow ML and contexts/topics in conversatios are two different features..

[tmbo]

I am not sure what you mean by "adding tensorflow ML" then. Care to elaborate a bit? Thought it was just about adding it as a dependency for the context implementation.

[PHLF]

I think @matteoredaelli mean having a component classifying intents like sklearn but backed by a Tensorflow model.

[tmbo]

Alright, I reopened it and renamed it. I think this would be a good contribution.

[rajesh-bhat]

Wouldn't we need lots of training data if we try to go with deep learning models in tensorflow or keras ?

[PHLF]

Not necessarily as Tensorflow allows one to build more than DeepLearning models.

[rajesh-bhat]

You mean to say word vectors with neural networks would do better?

[PHLF]

I mean all machine learning models for classification you can build with Tensorflow like shallow neural networks, Random Forests, Support Vector Machines, logistic regression don't necessarily require huge training data sets. As for defining which is better, it depends on your usecase.

[rajesh-bhat]

Thank you @PHLF

[ajayrfhp]

I raised something related to this in gitter. Someone was kind enough to point out that performance gain if any would be marginal unless you have real huge datasets. Huge datasets for intent classification are limited. The largest I have found is 20,000 sample big Frames dataset, but dataset is skewed towards a few classes. For the 20 k frames dataset, RASA's SVM was on par with neural nets. The performance gain on unseen datasets, the number of people who have well-balanced datasets for intent classification remains to be seen.

[itsmudassir]

@ajayrfhp can you please provide this 20 k frames dataset. Thanyou

[akane999]

I already have a code for intent recognition with bag of words and a DNN classifier (keras) . Tell me if you still need help with this issue.

[itsmudassir]

Thanks for the reply. yes i would love to learn from your code . If you can
please show me your repository and provide a read-me about how to run the
code and the intents dataset that you have used .
Thanks!

[nahidalam]

@akane999 I would love to look into that intent recognition code too.