Date of Award

Summer 2016

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

Committee Director

Michael L. Nelson (Director)

Committee Member

Michele C. Weigle

Committee Member

Hussein Abdel-Wahab

Committee Member

M'Hammed Abdous

Abstract

Much of our cultural discourse occurs primarily on the Web. Thus, Web preservation is a fundamental precondition for multiple disciplines. Archiving Web pages into themed collections is a method for ensuring these resources are available for posterity. Services such as Archive-It exists to allow institutions to develop, curate, and preserve collections of Web resources. Understanding the contents and boundaries of these archived collections is a challenge for most people, resulting in the paradox of the larger the collection, the harder it is to understand. Meanwhile, as the sheer volume of data grows on the Web, "storytelling" is becoming a popular technique in social media for selecting Web resources to support a particular narrative or "story".

In this dissertation, we address the problem of understanding the archived collections through proposing the Dark and Stormy Archive (DSA) framework, in which we integrate "storytelling" social media and Web archives. In the DSA framework, we identify, evaluate, and select candidate Web pages from archived collections that summarize the holdings of these collections, arrange them in chronological order, and then visualize these pages using tools that users already are familiar with, such as Storify.

To inform our work of generating stories from archived collections, we start by building a baseline for the structural characteristics of popular (i.e., receiving the most views) human-generated stories through investigating stories from Storify. Furthermore, we checked the entire population of Archive-It collections for better understanding the characteristics of the collections we intend to summarize. We then filter off-topic pages from the collections the using different methods to detect when an archived page in a collection has gone off-topic. We created a gold standard dataset from three Archive-It collections to evaluate the proposed methods at different thresholds. From the gold standard dataset, we identified five behaviors for the TimeMaps (a list of archived copies of a page) based on the page’s aboutness. Based on a dynamic slicing algorithm, we divide the collection and cluster the pages in each slice. We then select the best representative page from each cluster based on different quality metrics (e.g., the replay quality, and the quality of the generated snippet from the page). At the end, we put the selected pages in chronological order and visualize them using Storify.

For evaluating the DSA framework, we obtained a ground truth dataset of hand-crafted stories from Archive-It collections generated by expert archivists. We used Amazon’s Mechanical Turk to evaluate the automatically generated stories against the stories that were created by domain experts. The results show that the automatically generated stories by the DSA are indistinguishable from those created by human subject domain experts, while at the same time both kinds of stories (automatic and human) are easily distinguished from randomly generated stories

ISBN

9781369143713

Share

COinS