Inducing Document Plans for Concept-to-Text Generation

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 1503–1514, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Inducing Document Plans for Concept-to-text Generation Ioannis Konstas and Mirella Lapata Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB ikonstas@inf.ed.ac.uk, mlap@inf.ed.ac.uk Abstract In a language generation system, a content planner selects which elements must be in- cluded in the output text and the ordering be- tween them. Recent empirical approaches per- form content selection without any ordering and have thus no means to ensure that the out- put is coherent. In this paper we focus on the problem of generating text from a database and present a trainable end-to-end generation system that includes both content selection and ordering. Content plans are represented intuitively by a set of grammar rules that op- erate on the document level and are acquired automatically from training data. We de- velop two approaches: the first one is inspired from Rhetorical Structure Theory and repre- sents the document as a tree of discourse re- lations between database records; the second one requires little linguistic sophistication and uses tree structures to represent global patterns of database record sequences within a doc- ument. Experimental evaluation on two do- mains yields considerable improvements over the state of the art for both approaches. 1 Introduction Concept-to-text generation broadly refers to the task of automatically producing textual output from non- linguistic input (Reiter and Dale, 2000). Depend- ing on the application and the domain at hand, the input may assume various representations including databases, expert system knowledge bases, simula- tions of physical systems, or formal meaning rep- resentations. Generation systems typically follow a pipeline architecture consisting of three compo- nents: content planning (selecting and ordering the parts of the input to be mentioned in the output text), sentence planning (determining the structure and lexical content of individual sentences), and surface realization (verbalizing the chosen content in natu- ral language). Traditionally, these components are hand-engineered in order to ensure output of high quality. More recently there has been growing interest in the application of learning methods because of their promise to make generation more robust and adaptable. Examples include learning which con- tent should be present in a document (Duboue and McKeown, 2002; Barzilay and Lapata, 2005), how it should be aligned to utterances (Liang et al., 2009), and how to select a sentence plan among many al- ternatives (Stent et al., 2004). Beyond isolated com- ponents, a few approaches have emerged that tackle concept-to-text generation end-to-end. Due to the complexity of the task, most models simplify the generation process, e.g., by treating sentence plan- ning and surface realization as one component (An- geli et al., 2010), by implementing content selection without any document planning (Konstas and Lap- ata, 2012; Angeli et al., 2010; Kim and Mooney, 2010), or by eliminating content planning entirely (Belz, 2008; Wong and Mooney, 2007). In this paper we present a trainable …

Content truncated to 500 words