Unsupervised learning of rhetorical structure with un-topic models

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 2–13, Dublin, Ireland, August 23-29 2014. Unsupervised learning of rhetorical structure with un-topic models Diarmuid ´O S´eaghdha Computer Laboratory University of Cambridge Cambridge, UK do242@cam.ac.uk Simone Teufel Computer Laboratory University of Cambridge Cambridge, UK sht25@cam.ac.uk Abstract In this paper we investigate whether unsupervised models can be used to induce conventional aspects of rhetorical language in scientific writing. We rely on the intuition that the rhetorical language used in a document is general in nature and independent of the document’s topic. We describe a Bayesian latent-variable model that implements this intuition. In two empirical evaluations based on the task of argumentative zoning (AZ), we demonstrate that our generality hypothesis is crucial for distinguishing between rhetorical and topical language and that features provided by our unsupervised model trained on a large corpus can improve the performance of a supervised AZ classifier. 1 Introduction Scientific writing has many conventions. Some exist at the level of sentence construction, such as a preference for the passive voice or for deverbal nominalisations. Others relate to the high-level organisation of a paper: a typical paper at an NLP conference may be divided into sections covering the introduction, related work, methods, experimental results and conclusion. There are also intermediate levels of convention that use lexical and phrasal items to signal the role played by each part of the text in the argument the authors wish to construct. The theory of argumentative zoning (AZ) describes how a scientific article can be analysed in terms of text blocks (or zones) that share a rhetorical function (Teufel, 2010). For example: part of the article may consist of background information, another part may describe the aim of the research, other parts may report the authors’ own work or compare that work to alternative approaches in the literature. Supervised computational systems can be trained to mark up the AZ structure of a text automatically (see Section 2); the output of such systems has been shown to aid summarisation and human browsing of the scientific literature (Teufel and Moens, 2002; Guo et al., 2011a; Contractor et al., 2012). However, supervised systems require manually annotated training data that must be created anew for each discipline (and language) before they can be deployed, while large quantities of unannotated text are often available. For this reason, there is considerable value in developing unsupervised systems that induce aspects of rhetorical structure from unannotated text. In this paper we advance a hypothesis about the generality of rhetorical language. We propose that the words and linguistic constructs used to express rhetorical function in a scientific paper are independent of the paper’s topic. Naturally there will be some variation across research areas and there may be large differences across disciplines, but within a discipline we do not expect that the specific subject of a paper plays a significant role in how the authors construct their argument. For example, the following template could be used to generate an …

Content truncated to 500 words