This repository contains code for our paper The Web as a Knowledge-base for Answering Complex Questions. It can be used to train a neural model for answering complex questions, when the answer needs to be derived from multiple web snippets. This model was trained on the dataset ComplexWebQuestions, and the code is in PyTorch.
-
First, clone the repository:
git clone https://alontalmor@bitbucket.org/alontalmor/webkb_dev.git -
Change your directory to where you cloned the files:
cd webkb_dev -
Create a virtual environment with Python 3.6:
virtualenv -p python3 venv python3.6 -m venv venv -
Activate the virtual environment. You will need to activate the venv environment in each terminal in which you want to use WebAsKB.
source venv/bin/activate (or source venv/bin/activate.csh) -
Install the required dependencies:
pip3 install -r requirements_cloud.txt -
Install pytorch 0.3.1 from their website:
torch-0.3.0.post4-cp36-cp36m-macosx_10_7_x86_64.whl http://download.pytorch.org/whl/cu80/torch-0.3.0.post4-cp36-cp36m-linux_x86_64.whl
-
Download external libraries:
wget https://www.dropbox.com/s/k867s25qitdo8bc/Lib.zip unzip Lib.zip -
Download the data:
wget https://www.dropbox.com/s/tn45a3crehht7c1/Data.zip unzip Data.zip -
Optional - install and run Stanford NLP server, to generate noisy supervision:
wget http://nlp.stanford.edu/software/stanford-corenlp-full-2016-10-31.zip cd stanford-corenlp-full-2016-10-31 java -mx4g -cp "*" edu.stanford.nlp.pipeline.StanfordCoreNLPServer -port 9000 -timeout 15000
By default, we expect source data and preprocessed data to be stored in the "data" directory. The expected file locations can be changed by altering config.py. Note -- the dataset downloaded here contains only the question-answer pairs, the full dataset (including web snippets) can be downloaded from ComplexWebQuestions
Now you can do any of the following:
- Generate the noisy supervision data for training
python -m webaskb_run.py gen_noisy_sup. - Run a pointer network to generate split points in the question
python -m webaskb_run.py run_ptrnet. - Train the pointer network
python -m webaskb_run.py train_ptrnet. - Run final predication and calculate p@1 scores
python -m webaskb_run.py splitqa.
Options: ‘—eval_set dev’ or ‘—eval_set test’ to choose between the development and test set.
Please note, Reading Comprehension answer predication data is provided in Data/RC_answer_cache. However the WebAnswer model was not included due to its complexity and reliance on the ability to query a search engine. You may replace the RC component with any other RC model to be used with the web-snippets in ComplexWebQuestions .