Reconstructing Human-Generated Provenance Through Similarity-Based Clustering

In this paper, we revisit our method for reconstructing the primary sources of documents, which make up an important part of their provenance. Our method is based on the assumption that if two documents are semantically similar, there is a high chance that they also share a common source. We previously evaluated this assumption on an excerpt from a news archive, achieving 68.2% precision and 73% recall when reconstructing the primary sources of all articles. However, since we could not release this dataset to the public, it made our results hard to compare to others. In this work, we extend the flexibility of our method by adding a new parameter, and re-evaluate it on the human-generated dataset created for the 2014 Provenance Reconstruction Challenge. The extended method achieves up to 86% precision and 59% recall, and is now directly comparable to any approach that uses the same dataset.