Dialog System NLU

Dialog NLU library that contains Tensorflow and Keras Implementation of the state of the art researches in Dialog System NLU.

It is built on Tensorflow 2 and Huggingface Transformers library for better models coverage and other languages support.

Implemented Papers

NLU Papers

• BERT for Joint Intent Classification and Slot Filling

  Model Compression Papers

• Poor Man’s BERT: Smaller and Faster Transformer Models

  BERT / ALBERT for Joint Intent Classification and Slot Filling

Poor Man’s BERT: Smaller and Faster Transformer Models

Supported data format:

• Data format as in the paper Slot-Gated Modeling for Joint Slot Filling and Intent Prediction (Goo et al):
  • Consists of 3 files:
    • seq.in file contains text samples (utterances)
    • seq.out file contains tags corresponding to samples from seq.in
    • label file contains intent labels corresponding to samples from seq.in

Datasets included in the repo:

• Snips Dataset (Snips voice platform: an embedded spoken language understanding system for private- by-design voice interfaces )(Coucke et al., 2018), which is collected from the Snips personal voice assistant.
  • The training, development and test sets contain 13,084, 700 and 700 utterances, respectively.
  • There are 72 slot labels and 7 intent types for the training set.

Integration with Huggingface Transformers library

Huggingface Transformers has a lot of transformers-based models. The idea behind the integration is to be able to support more architectures as well as more languages.

And more models integration to come

Installation

You may choose to create python environment before installation.

git clone https://github.com/MahmoudWahdan/dialog-nlu.git
cd dialog-nlu
pip install .

Examples:

We provide examples of how to use the library. You can find Jupyter notebboks under notebooks and python scripts examples of how to use the library

Training, Evaluation, and simple API script:

We provide scripts to train, incremental training, and simple flask API.

Quick tour

# imports
from dialognlu import TransformerNLU, AutoNLU
from dialognlu.readers.goo_format_reader import Reader

# reading datasets
train_path = "data/snips/train"
val_path = "data/snips/valid"
train_dataset = Reader.read(train_path)
val_dataset = Reader.read(val_path)

# configurations of the model
config = {
    "pretrained_model_name_or_path": "distilbert-base-uncased",
    "from_pt": False,
}
# create a joint NLU model from configurations
nlu_model = TransformerNLU.from_config(config)

# training the model
nlu_model.train(train_dataset, val_dataset, epochs=3, batch_size=64)

# saving model
save_path = "saved_models/joint_distilbert_model"
nlu_model.save(save_path)

# loading the model and do incremental training

# loading model
nlu_model = AutoNLU.load(save_path)

# Continue training
nlu_model.train(train_dataset, val_dataset, epochs=1, batch_size=64)

# evaluate the model
test_path = "../data/snips/test"
test_dataset = Reader.read(test_path)
token_f1_score, tag_f1_score, report, acc = nlu_model.evaluate(test_dataset)
print('Slot Classification Report:', report)
print('Slot token f1_score = %f' % token_f1_score)
print('Slot tag f1_score = %f' % tag_f1_score)
print('Intent accuracy = %f' % acc)

# do prediction
utterance = "add sabrina salerno to the grime instrumentals playlist"
result = nlu_model.predict(utterance)

Use Layer Pruning with NLU model

It is supported only in transformer-based NLU models

# imports
from dialognlu import TransformerNLU, AutoNLU
from dialognlu.readers.goo_format_reader import Reader

# reading datasets
train_path = "data/snips/train"
val_path = "data/snips/valid"
train_dataset = Reader.read(train_path)
val_dataset = Reader.read(val_path)

# configurations of the model
config = {
    "pretrained_model_name_or_path": "distilbert-base-uncased",
    "from_pt": False,
    "layer_pruning": {
        "strategy": "top",
        "k": 2
    }
}
# create a joint NLU model from configurations
nlu_model = TransformerNLU.from_config(config)

# training the model
nlu_model.train(train_dataset, val_dataset, epochs=3, batch_size=64)