e of Duplicates

In plagiarism detection the goal is usually to identify the similarities between a suspicious document and a set of documents that it might have been copied from. However, in some cases there is no set of documents to compare the suspicious document to and thus a candidate set needs to be identified. Once this candidate set has been identified, more accurate analysis can be performed. Source Retrieval involves using a search engine to identify this candidate set. Queries from the suspicious document are submitted to the search engine and a decision is made about which search results are worth analyzing. In this poster, we present a competition winning approach to source retrieval that achieved the best overall performance in two plagiarism detection competitions (Williams et al., 2014a; Potthast et al., 2014) by using a supervised method to classify the search engine results as potential sources of plagiarism. The source retrieval problem is formally described below. Source Retrieval Problem: Given a suspicious document and a search engine, use the search engine to retrieve candidate documents that may be sources of plagiarism while minimizing the number false positives retrieved. We solve this problem by automatically extracting queries from the suspicious document, submitting the queries to a search engine, using a machine learning approach to classify the search results returned by the search engine and then only retrieve the results classified as being potential sources of plagiarism while ignoring the others. This process is demonstrated in Figure 1