The Use of Granularity in Rhetorical Relation Prediction

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 44–48, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics The Use of Granularity in Rhetorical Relation Prediction Blake Stephen Howald and Martha Abramson Ultralingua, Inc. 1313 SE Fifth Street, Suite 108 Minneapolis, MN 55414 {howald, abramson}@ultralingua.com Abstract We present the results of several machine learning tasks designed to predict rhetori- cal relations that hold between clauses in discourse. We demonstrate that organizing rhetorical relations into different granularity categories (based on relative degree of detail) increases average prediction accuracy from 58% to 70%. Accuracy further increases to 80% with the inclusion of clause types. These results, which are competitive with existing systems, hold across several modes of written discourse and suggest that features of informa- tion structure are an important consideration in the machine learnability of discourse. 1 Introduction The rhetorical relations that hold between clauses in discourse index temporal and event information and contribute to a discourse’s pragmatic coherence (Hobbs, 1985). For example, in (1) the NARRATION relation holds between (1a) and (1b) as (1b) tempo- rally follows (1a) at event time. (1) a. Pascale closed the toy chest. b. She walked to the gate. c. The gate was locked securely. d. So she couldn’t get into the kitchen. The ELABORATION relation, describing the sur- rounding state of affairs, holds between (1b) and (1c). (1c) is temporally inclusive (subordinated) with (1b) and there is no temporal progression at event time. The RESULT relation holds between (1b- c) and (1d). (1d) follows (1b) and its subordinated ELABORATION relation (1c) at event time. Additional pragmatic information is encoded in these relations in terms of granularity. Granularity refers to the relative increases or decreases in the level of described detail. For example, moving from (1b) to (1c), we learn more information about the gate via the ELABORATION relation. Also, moving from (1b-c) to (1d) there is a consolidation of infor- mation associated with the RESULT relation. Through several supervised machine learning tasks, we investigate the degree to which granularity (as well as additional elements of discourse struc- ture (e.g. tense, aspect, event)) serves as a viable organization and predictor of rhetorical relations in a range of written discourses. This paper is orga- nized as follows. Section 2 reviews prior research on rhetorical relations, discourse structure, granular- ity and prediction. Section 3 discusses the analyzed data, the selection and annotation of features, and the construction of several machine learning tasks. Section 4 provides the results which are then dis- cussed in Section 5. 2 Background Rhetorical relation prediction has received consid- erable attention and has been shown to be useful for text summarization (Marcu, 1998). Prediction tasks rely on a number of features (discourse con- nectives, part of speech, etc.) (Marcu and Echihabi, 2002; Lapata and Lascarides, 2004). A wide range of accuracies are also reported - 33.96% (Marcu and Echihabi, 2002) to 70.70% (Lapata and Lascarides, 2004) for all rhetorical relations and, for individ- ual relations, CONTRAST (43.64%) and CONTINU- ATION …

Content truncated to 500 words