This repository was created as a part of a diploma thesis Semantic Sentence Similarity for Intent Recognition Task. While all the included code works and is well documented, as of now, it might be hard for anyone to actually use it, but feel free to try. It is my intention to convert it into a proper Python library in the future.

Template-based intent recognition system built on word embedding models.

Install Python 3.6 and higher

sudo python3 python3-pip python3-dev build-essential

Install Python package manager pipenv

pip3 install --user pipenv

Create and enter the virtual environment

pipenv shell

Install dependencies (including development):

pipenv install --dev

The intent recognition system is dependent on the used embedding model. These models are loaded using the wrappers in embeddings directory.

Currently supported embedding algorithms:

• InferSent: https://github.com/facebookresearch/InferSent
• sent2vec: https://github.com/epfml/sent2vec
• GloVe: https://github.com/maciejkula/glove-python or SpaCy package implementation
• word2vec: gensim package implementation
• FastText: https://github.com/facebookresearch/fastText or gensim package implementation
• StarSpace: https://github.com/facebookresearch/StarSpace

Depending on the embedding algorithm used, it might be needed to install its implementation. You will find the installation instructions for each algorithm on the respective repository.

Module model_compression.py includes functions for compressing embedding models. It is able to compress the models by using different versions of vocabulary pruning and by using vector quantization.

The vector quantization is based on the LBG clustering algorithm, which is implemented in module lbg.py.

The resulting intent recognition system is implemented in intent_query.py. As of now, it loads an embedding model and a set of intent templates from a JSON file (the templates are further quantized). The user can then write sentences in the command line and the algorithm will output the matched intent, together with the respective template and its cosine similarity to the input sentence.

Examples of an embedding model and a JSON template file can be found in the data directory.