Evin Schuchardt
Luke Roberts
Reference Information
Title: User-oriented document summarization through vision-based eye-tracking
Authors: Songhua Xu, Hao Jiang, Francis C.M. Lau
Presentation Venue: IUI 2009: Proceedings of the 14th international conference on Intelligent user interfaces; February 8-11, 2009; Sanibel Island, Florida, USA
Summary
This paper discusses a new algorithm for document summarization using vision-based eye-tracking and word semantic analysis.
They use an eye-tracking setup with a web camera and an eye-tracking algorithm described in a different paper. They take gaze samples on the words, noting which words got the most attention. They then compare the semantic similarity between words, and if the determined similarities are similar enough, they assume that the user will give both words the same amount of attention. The algorithm they use to determine semantic similarity is described in a different paper. They then select a number of words that have the highest semantic similarity. They use the same idea when predicting how much attention a user will give to a sentence.
They then make an algorithm unique to the user that will summarize the document based on what the user seems to find important about the article. They restrict the algorithm to only output a percentage of the sentences given the most attention. The percentage is based on the size of the document.
The researchers performed a study to see how accurate their algorithm was by comparing it to two popular text summarization algorithms already in use (MEAD and Microsoft Word AutoSummarize). They also compare the summaries to the summaries produced by the users on their own. They had users read different articles from two sets: science articles and entertainment/leisure articles.
The researchers found that their algorithm could create more personalized summaries based on the user’s interests, especially for the entertainment/leisure articles, than the other two algorithms could.
Discussion
With all the reading we’ve been doing for this class this semester, I’ve got admit I was really excited to read this paper and see how they went about making this algorithm. Aside from the several references they gave to other papers, they were thorough in explaining how their algorithm worked.
However, it seems like they rushed their results section. They didn’t given much information on how the users responded to their system. They just said it was better than the others and presented a few equations on how they calculated the recall and precision of their algorithm. However, I still did find the paper and the techniques described to be interesting.
The researchers note that in future studies they would like to make it so that their algorithm can work even on articles that the user has not read based on results from previous articles that the user has read. They also mention improving the overall algorithm and allowing for feedback from the user to improve summarization.
 |
| Image from paper showing which words were given the most attention based on their algorithm |
Even though I have never used an automated summary generator, I think this idea sounds very original. I am really intrigued by the eye-tracking algorithm and what other types of applications it has been used for.
ReplyDeleteThough they did not explain the user comments on the implementation it was interesting that they noted that it performed better than the other two systems they compared it to. It doesn't seem that they have much credit to do that, but aside from that they seem confident in their results.
ReplyDeleteThis is a very cool concept, but I am skeptical about their assumptions. My first instinct on the reason a user spends a longer time on a certain spot in an article is because it is hard to read or comprehend, not necessarily because they are more interested in it. The fact they somewhat proved their system to perform well would suggest I am wrong though.
ReplyDelete