rpo19 / AIQuestionAnswering

"Question Answering: Structured vs Unstructured" project for Artificial Intelligence master course of University Milano Bicocca

Question Answering: Structured vs Unstructured

This project is for the master course of Artificial Intelligence from Università Milano Bicocca.

The project aims to explore and compare open domain question answering approaches over both structured (Knowledge Graph) and unstructured data (Free-text), using respectively DBPedia and Wikipedia as information sources. A prototype of both types of approaches has been implemented. This has highlighted the issues and complexities behind the implementation of a QA system in a real-world scenario. To understand the pros and cons of the implemented approaches different kinds of evaluation have been performed.

To know more about the project look at the Presentation.pdf

Directory structure

demo # demo interface
knowledgeGraph # knowledge graph QA
text # free text QA
evaluation # evaluation script to compare the two approaches
data # create this folder and put models there

Run the demo

The demo loads both the Graph based and the Free text QA therefore requires much memory.

With docker (suggested)

Install docker and docker-compose (https://docs.docker.com/get-docker/ https://docs.docker.com/compose/install/), then:

cd docker
docker-compose up -d
# some time is required at first start to download models

Now open http://localhost and you should access QA demo.

Install python requirements

The project has been developed on python3.8

• Create a virtualenv (suggested)

  python -m venv venv
  

  and activate it

  source venv/bin/activate # linux
  # or for windows
  .\venv\Scripts\activate.bat
  

• Install requirements

  pip install -r demo/backend/requirements.txt
  pip install -r knowledgeGraph/kgqalib/requirements.txt
  pip install -r text/ftqalib/requirements.txt
  

• Download required models/data

  • spacy:

    python -m spacy download en_core_web_lg
    

  • nltk stopwords

    python -c 'import nltk;nltk.download("stopwords")'
    

  • Glove embeddings in gensim format; download from https://drive.google.com/file/d/1wHP5tHvBiG5aHuCHxotYQYLeQxRKI-gm/view?usp=sharing and save in data/glove.840B.300d.gensim
  • Pretrained pattern classifier; download from https://drive.google.com/file/d/14OtRDFrT-mAEDSqBX7dQKJGAG39xPd7c/view?usp=sharing and save in data/models/pattern_classifier.h5

• Run the backend with

  python demo/backend/server.py
  

  The first run will download some required models. Then backend will be available at http://localhost:5000.

• Download pre-built frontend from https://drive.google.com/file/d/1UwBqODTaZyRMZW2HrPVRQBqHe-KU1ssL/view?usp=sharing, unzip it and then serve it for instance with

  cd frontend
  python -m http.server 4200
  

  (port 4200 should be already in backend CORS origin as it is the default angularJS development server port)