Context-Aware Geospatial Question Answering Component - GSoC2021

Data accessibility is the key to a better understanding of the world. While most data artifacts have a geospatial context [1] it is crucial to support retrieval of such data with accessible interfaces. In our context we will use a chatbot UI on top of a Question Answering (QA) system. The major part of currently existing Knowledge Graph Question Answering (KGQA) systems are limited by the Knowledge Graphs (KG) that they are working on knowledge bases (c.f., [2]), i.e., they work mostly on logical facts. However, one of the new challenges in such a task is the inclusion of the current users’ context. For example, if a person asks: “What is this building in front of me?” or “What works of art are exhibited in this museum?”, then a system needs to take into account the current geospatial context. Hence, the system has to filter the possible search results regarding the geospatial context. Accordingly, to answer such questions, we must use not only the data from the KG (e.g., DBpedia). Instead we need to narrow down the search space based on the geolocation and geospatial orientation (line of sight) of the user and fetch additional information from external data sources (e.g., OpenStreetMap).

In this project the existing DBpedia chatbot [2] (developed within a GSoC project in 2017) will be extended by a geospatial component. We plan to reuse Question Answering (QA) components of the Qanary Framework [3, 4]. The Qanary QA Framework was already part of one of the past DBpedia GSoC project in 2017 and grew up a lot since that time.

The project proposes to implement a component (or set of components) that are capable of fetching the data based on the geospatial location of the user and build a (virtual) KG to answer the users’ questions. For example, given the question “What is this building in front of me?”, the QA component has to perform the following steps:

  1. Fetch all the buildings nearby and its properties based on the location;
  2. Build and insert a KG into temporal storage;
  3. Select the most relevant object (building) and the corresponding predicate (property of the building) that is user asking for;
  4. Run the SPARQL query on the KG and retrieve an answer.

Note, that this set of steps is just a draft idea of what you can implement, feel free to bring your own unique solutions to this problem.

Goals

  • Establish a working Context-Aware Geospatial Question Answering system;
  • Benchmark the system on the existing datasets.

Impact

The outcomes and contributions of the project might be reused in the research community as well as in industry. The Context-Aware QA field is not yet well studied which might be a good opportunity to become a pioneer in such a research field. As many of QA systems are now included not only inside mobile devices but also into the cars, the project results might become relevant for related industry companies.

Warm-up tasks

  • Have a look at the papers in the references list;
  • Run simple SPARQL Queries on DBpedia to get familiar with the data and technology;
  • Find 5 entities with geospatial context in DBpedia
  • Define 5 natural-language questions expressing the information need for these geospatial Entities and define the current geospatial context of the user (position and direction);
  • Define the corresponding SPARQL queries for these questions;
  • Use the first 10 questions of the CASQAD [5] dataset and analyze them, s.t., you marked all information required to answer the question correctly;
  • Implement a simple Qanary component using Python or Java (see the guides at [7]).

Mentors

Keywords

Question Answering, KBQA, Geospatial Search, Knowledge Graphs, Chatbot

Initialization Process

To our best knowledge there is no standard process to make statements about the completion status of the warm-up tasks. However, we only can make a final statement when Google is opening the application process. There might be something asked there regarding the warm-up tasks.

However, to get in touch with us, you might join the DBpedia Slack channel at https://dbpedia-slack.herokuapp.com/ and create a private conversation with us.

References

[1] D. Punjani, K. Singh, A. Both et. al. “Template-Based Question Answering over Linked Geospatial Data”

[2] D. Diefenbach, V. Lopez, K. Singh et. al. “Core techniques of question answering systems over knowledge bases: a survey”

[3] A. Both, D. Diefenbach, K. Singh et. al. “Qanary – A Methodology for Vocabulary-Driven Open Question Answering Systems”

[4] K. Singh, A. Sethupat, A. Both et. al. “Why Reinvent the Wheel-Let’s Build Question Answering Systems Together”

[5] J. Rose, J. Lehmann. “CASQAD – A New Dataset for Context-Aware Spatial Question Answering”

[6] http://chat.dbpedia.org/

[7] https://github.com/WDAqua/Qanary/wiki

1 Like

Hello @perevalov @anbo, I am really interested in this project. I’m a PhD student at University of Amsterdam working on Complex Question Answering. Hope to work with you soon!

1 Like

Hello @kolk, nice to connect with you. Just let us know if you have questions w.r.t. the warm-up tasks or anything else.

I’m very much interested in contributing to this project. Thanks in advance.

Hi @anbo @perevalov I am an undergraduate student from India and I am highly interested in this project and have an experience of 2 years of working with Q&A based projects. Kindly tell how may I proceed with this project.

Hi @shantanudube,

thanks for your interest in the project and reaching out to us.

A typical start would be to solve the warm-up tasks (c.f., Section “Warm-up tasks”). Additionally, you might contact us for a private discussion via Slack (c.f., Section “Initialization Process”).

Hope it helps.

2 Likes

Thankyou @anbo it is definitely going to help!

Dear Students. Please keep in mind to create your submission as early as possible. We, as mentors, can review your project proposal before the deadline and provide feedback, so you can improve it.