QAS - question answering system over linked data.
The system is created as a part of the master's thesis. Abstract:
The thesis deals with question answering over structured data. In knowledge databases, a structured data is usually represented by graphs. However, to satisfy information needs using natural language interfaces the system is required to hide the underlying schema from users. A question answering system with a schema-agnostic graph-based approach was developed as a part of this work. In contrast to traditional question answering systems that rely on deep linguistic analysis and statistical methods, the developed system explores provided graph to yield and reuse semantic connection for a known question-answer pair. Lack of large domain-specific structured data made us perform evaluation with the help of prominent open linked datasets such as Wikidata and DBpedia. Quality of separate answering stages and the approach in general was evaluated using adapted evaluation dataset and standard metrics.
Please, make sure, that you have a python3 (3.6.1) and pip3 and/or easy_install-3.x installed. Also curl utility is needed.
sudo apt-get install python3-setuptools python3.6-dev python3-pip
It is possible to create a virtual environment with python dependencies. Make sure, that you have virtualenv. To check that virtualenv is installed:
which virtualenv
You should see a path to the executable. If it is available, you can run the initialization script:
. ./init.sh
init.sh file creates a virtual environment, install needed dependencies and download WordNet and spaCy corporas.
After, you can run the system:
qa_system --help
To leave the virtual environment:
deactivate
To activate the environment again:
source env/bin/activate
To leave and delete the virtual environment:
. ./clean.sh
All dependencies will be automatically installed in case of the environment setup. This dependencies required only in case of a custom installation.
The QAS uses spaCy for natural language processing. You can install spaCy using pip:
spaCy and NLTK models installation:
python3 -m spacy download en_core_web_sm
echo "import nltk; nltk.download('wordnet')" > download_wordnet.py
python download_wordnet.py
For WebSocket wrapper you also need a Flask webserver:
pip3 install flask
Generating the evaluation dataset:
cd evaluation
python3 generate.py qald/ | tee dataset.test.json
Running the evaluation script:
python3 evaluate.py dataset.test.json 2> processing.log
Generating distribution graphs:
python3 processing.py
To manually run the system under the environment:
qa_system -q "In what city is the Heineken brewery?"
Please, provide a question-answer pair to find a semantic path:
qa_system -q "In what city is the Heineken brewery?" -a "Amsterdam"
This project is licensed under the MIT License - see the LICENSE.md file for details
Author: Mark Birger.
Supervisor: Doc. RNDr. Pavel Smrž, Ph.D.