The surprising effectiveness of explicit semantic analysis in dataless classification

Gupta, Shashank

Permalink

https://hdl.handle.net/2142/105826 Copy

Description

Title

The surprising effectiveness of explicit semantic analysis in dataless classification

Author(s)

Gupta, Shashank

Issue Date

2019-07-16

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

Roth, Dan

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)

• ESA
• Dataless Classification
• Embeddings
• Unsupervised Learning
• EntityESA
• Entity2Vec
• Topic2Vec
• Word2Concept

Abstract

Organizing textual content into broad labels is one of the most basic tasks that some people carry out on a regular basis. This simple task helps people navigate through large document collections by exposing the labels of the documents, which can then be used for selecting the documents of interest. Currently, the most popular techniques for providing this basic functionality are supervised in nature, wherein someone has to annotate a collection of documents with the labels of interest. However, it might not always be possible to create a sizeable labeled dataset for every scenario or domain of interest. Thus, techniques like “Dataless Classification” have been proposed in the past that are able to bootstrap the creation of a classifier by only requiring semantic descriptions of the labels. However, despite the encouraging performance of Dataless Classification on Text Classification tasks, there is still a room for large improvement. In this thesis, we identify the limitations of ESA-driven Dataless Classification and systematically design techniques for addressing each limitation. In the process, we end up developing 4 new embeddings – EntityESA, Entity2Vec, Topic2Vec and Word2Concept. However, despite our best efforts, we found it difficult to outperform the original Dataless Classification system. For some of the techniques we provide an explanation for this observed behavior, however we also attribute some of these observations to the datasets that are being used for evaluation purposes. We then propose a way to create a new dataset that can used for future Dataless evaluations. The new embedding methods proposed in this work are generic enough that they can be of independent interest as well.

Graduation Semester

2019-08

Type of Resource

text

Permalink

http://hdl.handle.net/2142/105826

Copyright and License Information

Copyright 2019 Shashank Gupta

Owning Collections