Using Facebook's Infersent model we create sentence embeddings of the existing data. When a new text is queried, we calculate the cosine distance between the the query text and the existing embeddings. The highest value is considered as a match and returned as the answer.

InferSent is a sentence embeddings method that provides semantic representations for English sentences. It is trained on natural language inference data and generalizes well to many different tasks.

Read the original paper - arxiv.

Make sure you have Python 3. Install the Python requirements:

pip install -r requirements.txt

run setup.sh to get the Infersent model and also all the word vectors. This project uses GloVe:

./setup.sh

For training, add any new data in data/ directory. Check data/README.md for format instructions. Then run the training to save the embeddings in embeddings/ dir:

python train.py

To check, import find_best_match from inference:

from inference import find_best_match

print(find_best_match(text="are you open source"))

This repository comes with a Tornado API server, with REST API and a websocket end point. To run the server:

python server.py

Check CONTRIBUTING.md for more details.

• Build a dataset to evaluate accuracy
• Evaluate GLoVe vs FastText
• Work on improving accuracy
• Handle common typos

The mighty MIT license. Check LICENSE for more details.