ChineseQG

This project includes two ways to generate simple questions based on RDF triples in the knowledge base, template extraction method and template based sequence-to-sequence(seq2seq) model.

Model Overview

Template Extraction

As shown above, we extract templates for a specific predicate ("相关人物"/related people in this case) from training set. Then we randomly select one of those extracted templates to generate new questions of the given triples from testing set.

Template-based Seq2seq

For a given triple, the input for triple encoder is the concatenation of (SUBJECT, SEP, PREDICT). SEP here is '|||'.

For template decoder, instead of using the entire question for input, we replace the SUBJECT(topic words) in the question with '(SUB)' token.

Installation

For training tseq2seq, we strongly recommend using GPU for accelerating the speed.

Tensorflow

The code for tseq2seq is based on Tensorflow. You can find installation instructions here.

Dependencies

The code is based on Python 2.7. The dependencies are listed in the file requirements.txt. You can install these dependencies as follows:

pip install -r requirements.txt

Data

We focus on Chinese simple question generation based on open-domain Chinese knowledge base provided by NLPCC 2017 Shared Task 5. The training file nlpcc-iccpol-2016.kbqa.training-data in our experiments is from the KBQA challenge.

Format

KB-based simple questions are the questions which only involve one-hop inference on the knowledge graph. In other words, we generate simple questions in this work only by single triples.

The (question, triple, answer) triples in nlpcc-iccpol-2016.kbqa.training-data and nlpcc-iccpol-2016.kbqa.testing-data are formatted as follows:

<question id=5>	《线性代数》的isbn码是什么？
<triple id=5>	线性代数 ||| isbn ||| 978-7-111-36843-4
<answer id=5>	978-7-111-36843-4

Usage

Here we provide implementations for template extraction model and template-based seq2seq (tseq2seq) model.

Preprocess

First of all, we do some preprocessing work (using python3 instead of python2 to avoid chinese character encoding errors) on the original training and testing data, including creating vocabulary table, vectorizing input chinese characters and preparing for BLEU and ROUGE evaluations.

python3 preprocess.py

Template Extraction

Template extraction contains two phases: template collection and selective generation. You can implement the two phases in a pipeline by template.py and get automatic evaluation results:

python template-extraction/template.py

Template-based Seq2seq

We implement a general seq2seq framework that contains both vanilla seq2seq and tseqseq model. You can train with vanilla seq2seq model by setting tseq2seq to False.

python tseq2seq/Main.py

You can also adjust the learning rate if the loss is decreasing too slow or seems to over-fit. The dimension of word embedding and hidden units can also easily be changed in Main.py.