Neural network for intent classification built from scratch in Golang.
It is a 2-layer neural network (hidden and output), it uses ReLU for hidden layer activation and Softmax in the output layer which return the probabilities for each intent category.
Text inputs are preprocessed removing non letter characters, removing diacritics, converting to lower case and finally encoding as one hot vectors. Spell correction is made using levenshtein distance.
The optimization process is made with the Adam algorithm.
To train the model put the train data in the folder named data. The model params will be saved in output folder.
$ ./train.sh
Results with the example data
To run a web application demo using the model.
$ ./run.sh
The bot responses mapping intent categories are stored in data/responses.json.
