Semi-Supervised Never-Ending Learning in Rhetorical Relation Identification

Proceedings of Recent Advances in Natural Language Processing, pages 436–442, Hissar, Bulgaria, Sep 7–9 2015. Semi-Supervised Never-Ending Learning in Rhetorical Relation Identification Erick G. Maziero1,2, Graeme Hirst1 1Department of Computer Science University of Toronto Toronto, ON, M5S 3G4, Canada erick,gh@cs.toronto.edu Thiago A. S. Pardo2 2Department of Computer Science University of S˜ao Paulo S˜ao Carlos, SP, 13566-590, Brazil taspardo@icmc.usp.br Abstract Some languages do not have enough la- beled data to obtain good discourse pars- ing, specially in the relation identification step, and the additional use of unlabeled data is a plausible solution. A workflow is presented that uses a semi-supervised learning approach. Instead of only a pre- defined additional set of unlabeled data, texts obtained from the web are continu- ously added. This obtains near human per- fomance (0.79) in intra sentential rhetori- cal relation identification. An experiment for English also shows improvement using a similar workflow. 1 Introduction A text is composed of coherent propositions (phrases and sentences, for example) ordered and connected according to the intentions of the au- thor of the text. This composition may be recog- nized and structured according to many theories and this type of information is valuable to many natural language processing applications. A pro- cess to recognize, automatically, the coherent or discursive (or also rhetorical) structure of a text is named discourse parsing (DP). The most prominent theory in Computational Linguistics to structure the discourse of a text is the Rhetorical Structure Theory (RST) proposed by Mann and Thompson (1987). In this theory, the text is segmented into elementary discourse units (EDUs), which each contain a proposition (basic idea) of the text. The theory proposes a set of rhetorical relations that may hold between Figure 1: An example of sentence-level struc- ture according to RST. From Soricut and Marcu (2003). the EDUs, explicating the intentions of the author. For example, consider the sentence in Figure 1. It is segmented into three EDUs, numbered from 1 to 3. EDUs 2 and 3 are related by the relation Enablement, forming a new span of text, which is related to 1 by the relation Attribution. In each re- lation, EDUs can be Nucleus (more essential) or Satellite to the writer’s purpose. Many approaches have been used in DP, the majority of them using machine learning algo- rithms, such as probabilistic models (Soricut and Marcu, 2003), SVMs (Reitter, 2003; duVerle and Prendinger, 2009; Hernault et al., 2010; Feng and Hirst, 2012) and dynamic conditional random field (Joty et al., 2012). To obtain acceptable results, these approaches need plenty of labeled data. But even more than other levels of linguistic informa- tion, such as morphology or syntax, the annotation of discourse is an expensive task. Given this fact, what can we do when there is not enough data to perform effective learning of DP, as in languages with little annotated data? This paper describes a methodology to over- come the problem of insufficient labeled data in the task of identifying rhetorical relations between 436 Figure 2: Lexicalized syntactic …

Content truncated to 500 words