This repo is not currently supported as intent classifier became a part of DeepPavlov open-source library.

Try it here.

Neural Networks for Intent Classifier

In this repo one can find code for training and infering intent classification that is presented as shallow-and-wide Convolutional Neural Network[1].

Currently considered fasttext version in this repo does not works on Windows correctly.

Also this repo contains pre-trained model for intent classification on SNIPS dataset

SNIPS dataset considers the following intents: AddToPlaylist, BookRestaurant, GetWeather, PlayMusic, RateBook, SearchCreativeWork, SearchScreeningEvent.

Test results for other intent recognition services are from https://www.slideshare.net/KonstantinSavenkov/nlu-intent-detection-benchmark-by-intento-august-2017

How to install

First of all, one have to download this repo:

git clone https://github.com/deepmipt/intent_classifier.git

cd intent_classifier

The next step is to install requirements:

pip install -r requirements.txt

How to use pre-trained model (SNIPS)

Now one is able to infer pre-trained model:

./intent_classifier.py ./snips_pretrained/snips_config.json

The script loads pre-trained model, if necessary downloads pre-trained fastText embedding model [2], and then it is ready to predict class and probability of given phrase to belong with this class.

Example:

./intent_classifier.py ./snips_pretrained/snips_config.json
>I want you to add 'I love you, baby' to my playlist
>(0.99986315, 'AddToPlaylist')

How to train on your own data

The repo contains script train.py for training multilabel classifier.
Training data file should be presented in the following data.csv form:

request	class_0	class_1	class_2	class_3	...
text_0	1	0	0	0	...
text_1	0	0	1	0	...
text_2	0	1	0	0	...
text_3	1	0	0	0	...
...	...	...	...	...	...

Then one is ready to run train.py that includes reading data, tokenization, constructing data, building dataset, initializing and training model with given parameters on dataset from data.csv:

./train.py config.json data.csv 

The model will be trained using parameters from config.json file. There is a description of several parameters:

• Directory named model_path should exist. For example, if config.json contains "model_path": "./cnn_model", then configuration parameters for the trained model will be saved to ./cnn_model/cnn_model_opt.json and weights of the model will be saved to ./cnn_model/cnn_model.h5.

• Parameter model_from_saved means whether to load pre-trained model

• Parameter lear_metrics is a string that can include either metrics from keras.metrics or custom metrics from the file metrics.py (for example, fmeasure).

• Parameter confident_threshold is within the range [0,1] and means the boundary whether sample belongs to the class.

• Parameter fasttext_model contains path to pre-trained binary skipgram fastText [2] model for English language. If one prefers to use default model, it will be downloaded when one will train model.

• Parameter text_size means the number of words for padding of each tokenized text request.

• Parameter model_name contains name of the class method from multiclass.ry returning uncompiled Keras model. One can use cnn_model that is shallow-and-wide CNN (config.json contains parameters for this model), dcnn_model that is deep CNN model (be attentive to provide necessary parameters for the model), also it is possible to write own model.

• All other parameters refer to learning and network configuration.

How to infer

Infering can be done in two ways:

./infer.py config.json

or

./intent_classifier.py config.json

The first one runs infer.py file that contains reading parameters from config.json file, initializing tokenizer, initializing and infering model. The second one is doing the same but reads samples from command line.

References

[1] Kim Y. Convolutional neural networks for sentence classification //arXiv preprint arXiv:1408.5882. – 2014.

[2] P. Bojanowski*, E. Grave*, A. Joulin, T. Mikolov, Enriching Word Vectors with Subword Information.