dimartinot / text-semantic-similarity

This is the repository for the code that ran the experiments presented in the following article: Introduction to Deep Similarity Learning for Sequences

Text Semantic Similarity

This is the repository for the code that ran the experiments presented in the following article: Introduction to Deep Similarity Learning for Sequences

File Exploration

The most important files are:

• EDA.ipynb Exploratory Data Analysis notebook: used to clean and analyse the dataset. Generates the pickled version of the dataset with pre-computed sentence embeddings
• Training.ipynb Main training pipeline: loads pickled dataset generated using the EDA.ipynb notebook
• contrastiveModel.py : Models are kept in a single file for the moment as they share loads of similarities.

Installation

I advise the use of Anaconda distribution to run the code of this project. An anaconda environment file has been generated and can be used to create a new working environment using the following command:

conda env create -f environment.yml

Dataset generation

To generate the dataset, retrieved the source in kaggle and then import it and run the commands shown in notebook/EDA.ipynb to save a pickled dataset file (approx. 3GB of size).

Execution

To execute the main code of this project, you can either run:

cd notebook
jupyter notebook

and then run the Training.ipynb model.

Or you could just run:

python main.py

Results

The training result of my initial TextSimilarityDeepSiameseLSTM class with a LogReg classifier are the following:

Train Acc: 0.7993654994990785 - Val Acc: 0.7652195423623995 - Test Acc: 0.7669758812615955