Unsupervised Decomposition of a Document into Authorial Components

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 1356–1364, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Unsupervised Decomposition of a Document into Authorial Components Moshe Koppel Navot Akiva Idan Dershowitz Nachum Dershowitz Dept. of Computer Science Dept. of Bible School of Computer Science Bar-Ilan University Hebrew University Tel Aviv University Ramat Gan, Israel Jerusalem, Israel Ramat Aviv, Israel {moishk,navot.akiva}@gmail.com dershowitz@gmail.com nachumd@tau.ac.il Abstract We propose a novel unsupervised method for separating out distinct authorial compo- nents of a document. In particular, we show that, given a book artificially “munged” from two thematically similar biblical books, we can separate out the two consti- tuent books almost perfectly. This allows us to automatically recapitulate many con- clusions reached by Bible scholars over centuries of research. One of the key ele- ments of our method is exploitation of dif- ferences in synonym choice by different authors. 1 Introduction We propose a novel unsupervised method for separating out distinct authorial components of a document. There are many instances in which one is faced with a multi-author document and wishes to deli- neate the contributions of each author. Perhaps the most salient example is that of documents of his- torical significance that appear to be composites of multiple earlier texts. The challenge for literary scholars is to tease apart the document’s various components. More contemporary examples include analysis of collaborative online works in which one might wish to identify the contribution of a particular author for commercial or forensic pur- poses. We treat two versions of the problem. In the first, easier, version, the document to be decom- posed is given to us segmented into units, each of which is the work of a single author. The challenge is only to cluster the units according to author. In the second version, we are given an unsegmented document and the challenge includes segmenting the document as well as clustering the resulting units. We assume here that no information about the authors of the document is available and that in particular we are not supplied with any identified samples of any author’s writing. Thus, our me- thods must be entirely unsupervised. There is surprisingly little literature on this problem, despite its importance. Some work in this direction has been done on intrinsic plagiarism de- tection (e.g., Meyer zu Eisen 2006) and document outlier detection (e.g., Guthrie et al. 2008), but this work makes the simplifying assumption that there is a single dominant author, so that outlier units can be identified as those that deviate from the document as a whole. We don’t make this simpli- fying assumption. Some work on a problem that is more similar to ours was done by Graham et al. (2005). However, they assume that examples of pairs of paragraphs labeled as same- author/different-author are available for use as the basis of supervised learning. We make no such assumption. The obvious approach to our unsupervised ver- sion of the problem would be to segment the text …

Content truncated to 500 words