Update: Provenance Reconstruction
Work package 5 of Data2Semantics focuses on the reconstruction of provenance information. Provenance is a hot topic in many domains, at it is believed that accurate provenance information can benefit measures of trust and quality. In science, this is certainly true. Provenance information in the form of citations is a centuries old practice. However, this is not enough:
- What data is a scientific conclusion based on?
- How was that data gathered?
- Was the data altered in the process?
- Did the authors cite all sources?
- What part of the cited paper is used in the article?
Detailed provenance of scientific articles, presentations, data and images is often missing. Data2Semantics investigates the possibilities for reconstructing provenance information.
Over the past year, we have implemented a pipeline for provenance reconstruction (see picture), that is based on four steps: preprocessing, hypotheses generation, hypotheses pruning, aggragation and ranking. Each of these steps perform multiple analyses on a document that can be run in parallel.
Our first results, running this pipeline on a Dropbox folder (including version history), are encouraging. We achieve an F-score of 0.7 when we compare generated dependency relations to manually specified ones (our gold standard).
Sara Magliacane has a paper accepted at the ISWC 2012 Doctoral Consortium, where she explains her approach and methodology.