Semantic Document Distance Measures and Unsupervised Document Revision Detection

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 947–956, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP Semantic Document Distance Measures and Unsupervised Document Revision Detection Xiaofeng Zhu, Diego Klabjan Northwestern University, USA xiaofengzhu2013@u.northwestern.edu d-klabjan@northwestern.edu Patrick N Bless Intel Corporation, Chandler, AZ, USA patrick.n.bless@intel.com Abstract In this paper, we model the document revi- sion detection problem as a minimum cost branching problem that relies on comput- ing document distances. Furthermore, we propose two new document distance mea- sures, word vector-based Dynamic Time Warping (wDTW) and word vector-based Tree Edit Distance (wTED). Our revision detection system is designed for a large scale corpus and implemented in Apache Spark. We demonstrate that our system can more precisely detect revisions than state-of-the-art methods by utilizing the Wikipedia revision dumps 1 and simulated data sets. 1 Introduction It is a common habit for people to keep several versions of documents, which creates duplicate data. A scholarly article is normally revised sev- eral times before being published. An academic paper may be listed on personal websites, digital conference libraries, Google Scholar, etc. In major corporations, a document typically goes through several revisions involving multiple editors and authors. Users would benefit from visualizing the entire history of a document. It is worthwhile to develop a system that is able to intelligently iden- tify, manage and represent revisions. Given a col- lection of text documents, our study identifies re- vision relationships in a completely unsupervised way. For each document in a corpus we only use its content and the last modified timestamp. We assume that a document can be revised by many users, but that the documents are not merged to- gether. We consider collaborative editing as revis- ing documents one by one. 1https://snap.stanford.edu/data/wiki-meta.html The two research problems that are most rele- vant to document revision detection are plagiarism detection and revision provenance. In a plagiarism detection system, every incoming document is compared with all registered non-plagiarized doc- uments (Si et al., 1997; Oberreuter and Vel´aSquez, 2013; Hagen et al., 2015; Abdi et al., 2015). The system returns true if an original copy is found in the database; otherwise, the system returns false and adds the document to the database. Thus, it is a 1-to-n problem. Revision provenance is a 1- to-1 problem as it keeps track of detailed updates of one document (Buneman et al., 2001; Zhang and Jagadish, 2013). Real-world applications in- clude GitHub, version control in Microsoft Word and Wikipedia version trees (Sabel, 2007). In con- trast, our system solves an n-to-n problem on a large scale. Our potential target data sources, such as the entire web or internal corpora in corpora- tions, contain numerous original documents and their revisions. The aim is to find all revision doc- ument pairs within a reasonable time. Document revision detection, plagiarism detec- tion and revision provenance all rely on compar- ing the content of two documents and assessing a distance/similarity score. The classic document similarity measure, especially for plagiarism de- …

Content truncated to 500 words