Synonymous question generation: Learning to ask in different ways using variational autoencoders

Recently, there has been significant interest in advancing machine comprehension of text through question answering. Motivated by the idea that machine comprehension should be bidirectional, we explore synonymous question generation from knowledge graphs (KGs) to enable machines to learn how to ask semantically equivalent natural language questions with lexical and syntactical variety from KGs. To the best of our knowledge, this problem has not yet been explored in the literature. We propose explicitly modeling variations in natural language questions associated with KG triples through a conditional variational autoencoder-based model, the Template VAE (T-VAE). Evaluating the generated questions via the Fre'chet InferSent Distance (FID) and the Multiset-Jaccard-k-gram (MS-Jaccard-k) Measure, two joint diversity-quality metrics, demonstrates that the proposed model is able to produce fluent questions that accurately capture variations in questions associated with KG triples. Depending on test conditions, the T-VAE achieves a 15-21% improvement in MS-Jaccard-4 score and a 29-47% improvement in FID score relative to baseline methods.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2020

Author dc:creator

• Wei, Bingzhe

Contributors dc:contributor

• Chang, Kevin C.-C.

• question generation
• variational autoencoders
• machine learning
• natural language processing

Statement dc:rights

• Copyright 2020 Bingzhe Wei

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/107875

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/107875