A Neural QA Model for DBpedia - GSoC2021

This project started in 2018 and is now at its 4th consecutive year at DBpedia’s Google Summer of Code.

Introduction

Neural SPARQL Machines aim at building an end-to-end system to answer questions posed by user not versed with writing SPARQL queries.

Currently, DBpedia hosts billions of such data points and corresponding relations in the RDF format. Accessing such data is difficult for a lay user, who does not know how to write a SPARQL query. This GSoC project consists of building upon the NSpM​ question answering system, which tries to make this humongous linked data accessible to a larger user base in their natural language (as of now restricted to English) by improving, adding and amending upon the existing codebase.

Documentation

Source Code

The latest codebase is available at this forked repo: https://github.com/dbpedia/neural-qa

Blogs

To better understand the project please look into the following links:

  1. [GSoC 2018] Aman’s Blog: https://amanmehta-maniac.github.io/
  2. [GSoC 2019] Anand’s Blog: https://anandpanchbhai.com/A-Neural-QA-Model-for-DBpedia/
  3. [GSoC 2020] Zheyuan’s Blog: https://baiblanc.github.io/

Reading Material

The first 3 papers introduce and elaborate on Neural SPARQL Machines. Work number 3 was carried out by our GSoC 2019 student and published at KGSWC 2020. The 4th paper is an almost-complete survey of related approaches.

  1. SPARQL as a Foreign Language: https://arxiv.org/abs/1708.07624
  2. Neural Machine Translation for Query Construction and Composition: https://arxiv.org/abs/1806.10478
  3. Exploring Sequence-to-Sequence Models for SPARQL Pattern Composition: https://arxiv.org/abs/2010.10900
  4. Introduction to Neural Network based Approaches for Question Answering over Knowledge Graphs: https://arxiv.org/pdf/1907.09361.pdf

Warm-up tasks

  1. Read through the blogs and the reading list to get a good understanding of the code. This will allow you to get a good idea about the project.
  2. Run the pipelines in the ./gsoc/anand and ./gsoc/zheyuan folders of the base repository using examples of your choice.

Your proposal

Now that you have a good understanding of the current state of the project, we suggest you to build proposals pondering on some of the following points, feel free to bring your own solutions to tackle the problems that the project faces.

  1. How can we automatically build the right question from the property label only?
    • example a) from <s> dbo:birthPlace <o> infer where was <s> born?
    • example b) from <s> dbo:timeZone <o> infer what time zone is <s> in?
  2. How can we automatically build question-query templates that feature one or more of the following?
    • subordinate clauses or genitive: which / that / of / ’s
    • con-/disjunctions: and / or / as well as
    • modifiers: which + mod / what + mod / demonyms
    • comparative: more than / -er than
    • superlative: most … / -est
    • numeric / quantitative: how many / long / tall

Consider experimenting with advanced approaches such as GPT-2 or BERT.

Mentors

@panchbhai1969, @tsoru, TBD

Feel free to contact us for more information. We eagerly look forward to working with you and contributing towards making data accessible to all.

3 Likes

Hey there! I’m Riya Elizabeth John, a sophomore from IIT Roorkee, India.
I have a strong interest in this domain and am part of the Vision and Language Group(DL research group) of my institute.
Thrilled to see the ideas released by DBpedia this year, I’ll get started on the reading material ASAP. Really excited to work with you guys! :smiley:

1 Like

Hi @riyabelle25 and welcome!

Glad to see you interested in this project.

1 Like

Hello, @tsoru @panchbhai1969.
I am Siddhant Jain, a pre-final year student from Pune. The overall concept of the project and looking at previous blogs it seems really intuitive and interesting.
Have been going through the references, feels positive :slight_smile:
Hoping to have a steep learning curve this summer.

1 Like

Welcome, @siddhantjain07, and thank you for your interest in the project.

1 Like