Unsupervised Text Segmentation Based on Native Language Characteristics

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1457–1469 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1134 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1457–1469 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1134 Unsupervised Text Segmentation Based on Native Language Characteristics Shervin Malmasi1,2 Mark Dras2 Mark Johnson2 Lan Du3 Magdalena Wolska4 1Harvard Medical School, Harvard University smalmasi@bwh.harvard.edu 2Department of Computing, Macquarie University { shervin.malmasi, mark.dras, mark.johnson }@mq.edu.au 3Faculty of IT, Monash University lan.du@monash.edu 4LEAD Graduate School, Universit¨at T¨ubingen magdalena.wolska@uni-tuebingen.de Abstract Most work on segmenting text does so on the basis of topic changes, but it can be of interest to segment by other, stylistically expressed characteristics such as change of authorship or native language. We pro- pose a Bayesian unsupervised text seg- mentation approach to the latter. While baseline models achieve essentially ran- dom segmentation on our task, indicating its difficulty, a Bayesian model that incor- porates appropriately compact language models and alternating asymmetric priors can achieve scores on the standard metrics around halfway to perfect segmentation. 1 Introduction Most work on automatically segmenting text has been on the basis of topic: segment boundaries correspond to topic changes (Hearst, 1997). There are various contexts, however, in which it is of in- terest to identify changes in other characteristics; for example, there has been work on identifying changes in authorship (Koppel et al., 2011) and poetic voice (Brooke et al., 2012). In this paper we investigate text segmentation on the basis of change in the native language of the writer. Two illustrative contexts where this task might be of interest are patchwriting detection and lit- erary analysis. Patchwriting is the heavy use of text from a different source with some modifica- tion and insertion of additional words and sen- tences to form a new text. Pecorari (2003) notes that this is a kind of textual plagiarism, but is a strategy for learning to write in an appropri- ate language and style, rather than for deception. Keck (2006), Gilmore et al. (2010) and Vieyra et al. (2013) found that non-native speakers, not surprisingly in situations of imperfect mastery of a language, are strongly over-represented in this kind of textual plagiarism. In these cases the boundaries between the writer’s original text and (near-)copied native text are often quite appar- ent to the reader, as in this short example from Li and Casanave (2012) (copied text italicised): “Nevertheless, doubtfulness can be cleared rea- sonably by the experiments conducted upon the ‘split-brain patients’, in whom intra-hemispheric communication is no longer possible. To illustrate, one experiment has the patient sit at a table with a non-transparent screen blocking the objects be- hind, who is then asked to reach the objects with different hand respectively.” Because patchwrit- ing can indicate imperfect comprehension of the source (Jamieson and Howard, 2013), identifying it and supporting novice writers to improve it has become a focus of programmes …

Content truncated to 500 words