User Intent Recognition

Deep Neural Networks can be attentive too..

Detecting User's intent is very crucial in driving the conversations in a dialog system in the right direction. In simple words intent recognition engine classifies user utterance to one of the target class/intent.

Components of Intent Recognition Engine

1. Multilingual Transformer based Language Model

   This component converts input sentence/words to contextual embeddings, by contextual we mean the same word in different contexts would end up getting different embeddings, this contextuality also helps in dealing with homographs. The words in an utterance are tokenized using a sub-word tokenizer (more about this in the next section), these token's id's traverse through the domain fine tuned Language models to output 768-dimensional word vectors. various pooling techniques can be employed to convert these high dimensional word vectors into sentence embeddings.

2. Convolutional Neural Network (CNN) based classifier

   The embeddings extracted from a Domain Adapted Language Model is used to train a CNN classifier

intent-recognition engine