Explanation mining

Bhavya

Permalink

https://hdl.handle.net/2142/108346 Copy

Description

Title

Explanation mining

Author(s)

Bhavya

Issue Date

2020-05-13

Director of Research (if dissertation) or Advisor (if thesis)

Zhai, ChengXiang

Department of Study

Computer Science

Discipline

Computer Science

Degree Granting Institution

University of Illinois at Urbana-Champaign

Degree Name

M.S.

Degree Level

Thesis

Keyword(s)

• explanation mining
• slide explainer

Abstract

In this thesis, we propose the idea of computational analysis of explanations. Explanations are used to provide an understanding of a concept, procedure or reasoning to others. Although explanations are present online ubiquitously within textbooks, videos, blogposts, discussion forums, and many more, there is no way to mine them automatically. As a result, users in need of an explanation have to rely on search engines and potentially read through multiple documents in an attempt to find a suitable explanation. This process can be highly tedious for them and may not even be successful in some cases. On the other hand, there are many users such as educators, authors, who write explanations and can benefit from assistants that help enhance the quality of their explanations. The goal of computational analysis of explanations is to assist both these kinds of users. In this work, our focus is on Explanation Mining to assist users seeking explanations. For understanding some of the linguistic features of explanations across multiple domains, we first apply standard Learning-to-rank models to rank explanations collected from the Explain Like I'm Five (ELI5) reddit forum. Based on cross-domain experiments, we find that a model trained on a sufficiently large dataset achieves decent performance across all domains which suggests that there are some common markers of explanations. Next, to apply this knowledge to the practical problem of mining explanations of educational concepts, we propose a baseline approach based on the popular Language Modeling approach of information retrieval. We show that incorporating knowledge from a model trained on the ELI5 dataset in the form of a document prior helps increase the performance of a standard retrieval model. This is encouraging because our method requires minimal in-domain supervision, as a result it can be deployed for multiple online courses. Finally, we show a demo system that acts as an assistant to online learners while viewing slides. The system enables users to select any piece of text on the slide and find an explanation for it. We conclude with some interesting directions for future work in this field.

Graduation Semester

2020-05

Type of Resource

Thesis

Permalink

http://hdl.handle.net/2142/108346

Copyright and License Information

Copyright 2020 Bhavya

Owning Collections