YouTaQA is a powerful Q&A system that provides a complete pipeline. Unlike existing QA systems, our system allows users to have accurate answers to their questions based entirely on our search engine, saving the user from having to provide anything other than the question, and possibly making their task easier. To achieve the objective of our system, as shown in the general diagram, we've conceived an architecture composed of three basic modules and an UI where the users enter their questions:
- A search engine (MRI) that serves to provide the 5 most relevant contexts to a given question.
- A module of context classification (MC) based on BERT to choose and identify among the 5 search engine results the best context likely to contain the correct answer to the question.
- Answer Extraction module (MER) based on BERT that extracts the correct answer from the context chosen by classifiier in the previous step.
In order to use our system, you will need to install the following packages using pip:
- tqdm 4.47.0
- pathlib
- sklearn
- nltk.corpus
- numpy 1.19.0
- tokenizers 0.8.1rc1
- jupyter
- xml.dom
- torch 1.5.1
- torchvision 0.6.1
- transformers
- django
- matplotlib 3.2.2
In addition, you need to install nltk stop words by launching the linux/MacOS terminal (this project has been tested on version 3.8.0) and apply the following steps:
python
import nltk
nltk.download(stopwords)
For the search engine, you need to install PyLucene by following this tutorial.
PS: In order to install PyLucene properly, after downloading pylucene, you must put the /IR/indexing/PythonEnglishAnalyzer.java file in the pylucene-8.3.0/java/org/apache/pylucene/analysis folder before starting the installation.
Before starting wikipedia indexing, we split each wikipedia article into its own file, in a well-structured tree structure (For more information, please see the project paper).
To do this, you have to download the dump from wikipedia and decompress it. Then run the IR/preprocessing/parse.py script with the command:
parse.py -i <input_file> -o <output_file>
where input_file is the folder of the previously uncompressed wikipedia dump and output_file is the output folder where to put the article files (you need to allow 200GB space on the disk).
To index documents, you need to run the IR/indexing/indexer.py script with the command :
index.py -i <input_folder> -o <output_folder>
Where input_folder is the folder that contains the result of the preprocessing (the folder that contains the files of the Wikipedia articles) and output_folder is the directory in which the index will be saved.
Our classification model is based on BERT and transformers, it allows a precision of 89%. In order to train the classification model, you just have to download the SQuAD dataset and put the train set and the dev set files in the /DeepLearning/Classifier/Data folder and run the script /DeepLearning/Classifier/training_classifier.py.
To train the answer extraction model, please follow the steps indicated on the transformers project. This model scores an F1_score=92% and an Exact_match=87%.
After finishing the previous steps, it is time to run and launch the web application of our YouTaQA system. To do this, you must first change the paths in /Web application/myproject/setting.py following the paths of your configuration (index path, classification model path, etc). After that, you can start the django server by running the script /Web application/manage.py with the following command :
Python manage.py runserver
Then, open the browser in the port indicated in the terminal.
For more information, you can read the PDF paper of this project or contact one of the developers responsible for this project: