[Paper Review] Resurrecting My Revolution: Using Social Link Neighborhood in Bringing Context to the Disappearing Web

Hany M. SalahEldeen, Michael L. Nelson|arXiv (Cornell University)|Sep 10, 2013

Social Media and Politics16 references3 citations

TL;DR

This paper proposes using the 'tweet signature'—a set of top five terms from co-occurring tweets about a missing web resource—as a query to search engines to recover semantically similar replacements. By mining the social link neighborhood of a missing resource, the method achieves 70%+ textual similarity to the original in 41% of cases, offering a practical solution for reconstructing lost web content in digital preservation efforts.

ABSTRACT

In previous work we reported that resources linked in tweets disappeared at the rate of 11% in the first year followed by 7.3% each year afterwards. We also found that in the first year 6.7%, and 14.6% in each subsequent year, of the resources were archived in public web archives. In this paper we revisit the same dataset of tweets and find that our prior model still holds and the calculated error for estimating percentages missing was about 4%, but we found the rate of archiving produced a higher error of about 11.5%. We also discovered that resources have disappeared from the archives themselves (7.89%) as well as reappeared on the live web after being declared missing (6.54%). We have also tested the availability of the tweets themselves and found that 10.34% have disappeared from the live web. To mitigate the loss of resources on the live web, we propose the use of a "tweet signature". Using the Topsy API, we extract the top five most frequent terms from the union of all tweets about a resource, and use these five terms as a query to Google. We found that using tweet signatures results in discovering replacement resources with 70+% textual similarity to the missing resource 41% of the time.

Motivation & Objective

• To investigate the persistence and archival status of web resources shared via social media, particularly in the context of historical events like the Egyptian revolution.
• To address the problem of web resource disappearance from the live web and public archives, which threatens digital preservation and access to first-draft historical records.
• To develop and evaluate a method for reconstructing missing web resources using contextual information from social media, specifically the 'social link neighborhood' of co-occurring tweets.
• To assess the effectiveness of tweet signatures as search queries in retrieving semantically similar replacements for missing resources.

Proposed method

• Extract up to 500 recent tweets per resource using the Topsy API to gather contextual information.
• Generate a 'tweet signature' by identifying the top five most frequent terms from the union of all tweets about a given resource.
• Use the tweet signature as a query in Google search to retrieve candidate replacement resources.
• Apply the boilerplate-removal library 'boilerpipe' to extract clean textual content from both the original missing resource and candidate replacements.
• Calculate cosine similarity between the original resource and each candidate to measure textual similarity.
• Rank and evaluate candidate replacements based on similarity scores, with the top result used for comparison.

Experimental results

Research questions

• RQ1 What is the rate of disappearance of web resources shared via social media, and how does this compare to archival preservation rates?
• RQ2 To what extent can the social link neighborhood of a missing resource—defined by co-occurring tweets—provide sufficient context for reconstruction?
• RQ3 How effective is the tweet signature method in retrieving replacement resources with high textual similarity to the original?
• RQ4 What is the impact of timemap revisions and archive shrinkage on the accuracy of archival availability estimates?
• RQ5 How do the availability and re-appearance of resources in archives and on the live web affect the reliability of digital preservation?

Key findings

• The rate of disappearance of web resources from the live web remains approximately 11% in the first year and 7.3% annually thereafter, confirming prior findings.
• The model for estimating archived resources showed a higher error rate of about 11.5%, likely due to timemap shrinkage and changes in archive indexing.
• 7.89% of archived resources were found to have disappeared from archives themselves, and 6.54% reappeared on the live web after being marked missing.
• 10.34% of the tweets themselves were found to be missing from the live web, indicating instability in social media content.
• Using tweet signatures as search queries achieved at least 70% textual similarity to the original missing resource in 41% of cases.
• The mean reciprocal rank (MRR) of the search results using tweet signatures was 0.43, indicating moderate effectiveness in retrieving the original resource.