Files in this item



application/pdfHODOSH-DISSERTATION-2015.pdf (34MB)
(no description provided)PDF


Title:Natural language image description: data, models, and evaluation
Author(s):Hodosh, Micah A
Director of Research:Hockenmaier, Julia
Doctoral Committee Chair(s):Hockenmaier, Julia
Doctoral Committee Member(s):Dolan, Bill; Forsyth, David; Roth, Dan
Department / Program:Computer Science
Discipline:Computer Science
Degree Granting Institution:University of Illinois at Urbana-Champaign
Subject(s):Computer Vision
Natural Language Processing
Image Description
Neural Networks
computer vision (CV)
natural language processing (NLP)
Machine Learning
Machine Learning Applications
Image Captioning
Abstract:Automatically describing an image with a concise natural language description is an ambitious and emerging task bringing together the Natural Language and Computer Vision communities. With any emerging task, the necessary groundwork developing appropriate datasets, strong baseline models, and evaluation frameworks is key. In this thesis, we introduce the rst large datasets speci cally designed with image description in mind, focusing on concrete descriptions that can be gleaned from the image alone. Furthermore, we develop strong baseline models that show the need to model language beyond a simple bag-of-words approach to increase performance. Most importantly, we introduce a ranking based framework for comparing image description models. We show that this framework is more reliable and accurate than the conventional wisdom of evaluating on novel model generated text. As this task has gained popularity recently, we further analyze the drawbacks of current evaluation methods, and put forth concrete extensions to our ranking framework that will guide progress towards modeling the association of natural language and the images the language describes.
Issue Date:2015-11-25
Rights Information:Copyright 2015 Micah A Hodosh
Date Available in IDEALS:2016-03-02
Date Deposited:2015-12

This item appears in the following Collection(s)

Item Statistics