Events
CIS SIG-NLP Ivanka Zhou Li
This is a past event
Title: An Exploration of Ranking Methods for Infographic Retrieval
Abstract: Information graphics from popular media, such as newspapers, magazines, blogs, and social networking sites, are a rich knowledge source that should be made accessible by information retrieval systems. Infographics in popular media generally have a high-level message that they intended to convey instead of merely displaying data for analysis or visualization. Also, the information conveyed by infographics in popular media is often not repeated in the accompanying article’s text.
The GRAPH research group at the University of Delaware has been working for a number of years on understanding infographics. Low-level computer vision techniques are used to transform an image of a bart chart, line graph, or grouped bar chart into a
representation that captures the various components and high level intended message of the graph. My research goal is to build an effective infographic retrieval system based on this representation with deep understandings of the infographics.
In earlier stages of my PhD, I have developed a query understanding system that takes a full sentence natural language query searching for infographics, and identifies words in the query that describe what should be depicted on the independent and dependent axes of relevant infographics and the type of intended message that should be conveyed.
In this talk, I will discuss and compare different ranking models to rank-order infographics, given the above knowledge about the infographics and user queries. A naive yet common way to determine the relevance between an infographic and a query is to treat the entire words in the infographic as a whole and so with the entire query, and then measure the relevance between the two. Instead of naively treating the infographic as a whole and the query as another, the different components of the infographic and the query could be estimated separately, and then combined in a Mixture Model. Different ranking models ranging from a naive model to a Mixture Model will be compared in this talk.
I will also talk about different methods of measuring the textual relevance of each infographic components to each query elements, for example, the relevance of the infographic independent axis to the hypothesized query independent axis, or the relevance of the infographic intended message to the hypothesized intended messages from the user query. I will show the experiment results of applying these different approaches in all of the proposed ranking models. Based on the experiment result, decision on which textual relevance measuring method should be taken and which ranking model to chose is made to achieve the best retrieval result.
Monday, October 28, 2013 at 1:25pm to 2:15pm
Smith Hall, Room 102A
Smith Hall, University of Delaware, Newark, DE 19716, USA
- Event Type
-
College of Engineering, Academics, Lectures & Programs, Students, Lectures and Programs, Community
- Calendar
- Departments
- Group
- ENGR - Computer & Information Sciences
- Contact Email
- Contact Name
-
Vijay Shanker
- Subscribe
Recent Activity
No recent activity