Question Generation with Adaptive Copying
Neural Networks
PRESENTER:
Xinyuan Lu
Carleton University
ABSTRACT:
Given the rapid development of communication technology, online business websites
are becoming increasingly popular. People could buy all the goods online rather
than going to physical stores. However, it is often time-consuming for customers to
read long product reviews when they make purchase decisions. Generally speaking,
customers have specific questions in their minds when they seek information in
product specifications. Therefore, if we can identify the most representative questions
that can be answered by extensive product specifications, it will be much easier for
customers to obtain their desired information quickly grasp, which is the motivation
behind this work.
We aim to automatically generate questions from reviews. To
tackle the automatic question generation (QG) task, we proposed a novel adaptive
copying recurrent neural network model, titled Adaptive Copying Neural Network
(ACNN). The proposed model adds a copying mechanism component onto a
bidirectional LSTM architecture to adaptively generate more suitable questions from
the input data. Subsequently, we calculated the generated questions' summarization
score to see if they could be answered by the reviews and to determine whether they
were the most representative questions to summarize those reviews.
In our evaluation experiments, we first evaluated our ACNN model. We confirmed
that our method can outperform the state-of-the-art QG methods in terms of
BLEU and ROUGE evaluation scores; it also showed better performance in human
evaluation. In addition, we combined the summarization scores with our model,
which resulted in a performance boost. Overall, our model is fully data-driven
and can simultaneously handle two significant tasks in natural language processing,
which represents a new direction in this field.