Unsupervised Concept-to-text Generation with Hypergraphs

2012 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 752–761, Montr´eal, Canada, June 3-8, 2012. c⃝2012 Association for Computational Linguistics Unsupervised Concept-to-text Generation with Hypergraphs Ioannis Konstas and Mirella Lapata Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB i.konstas@sms.ed.ac.uk, mlap@inf.ed.ac.uk Abstract Concept-to-text generation refers to the task of automatically producing textual output from non-linguistic input. We present a joint model that captures content selection (“what to say”) and surface realization (“how to say”) in an unsupervised domain-independent fashion. Rather than breaking up the generation pro- cess into a sequence of local decisions, we de- fine a probabilistic context-free grammar that globally describes the inherent structure of the input (a corpus of database records and text describing some of them). We represent our grammar compactly as a weighted hypergraph and recast generation as the task of finding the best derivation tree for a given input. Experi- mental evaluation on several domains achieves competitive results with state-of-the-art sys- tems that use domain specific constraints, ex- plicit feature engineering or labeled data. 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 includ- ing databases of records, expert system knowledge bases, simulations of physical systems and so on. Figure 1 shows input examples and their correspond- ing text for three domains, air travel, sportscasting and weather forecast generation. A typical concept-to-text generation system im- plements a pipeline architecture consisting of three core stages, namely text planning (determining the content and structure of the target text), sentence planning (determining the structure and lexical con- tent of individual sentences), and surface realiza- tion (rendering the specification chosen by the sen- tence planner into a surface string). Traditionally, these components are hand-engineered in order to generate high quality text, however at the expense of portability and scalability. It is thus no surprise that recent years have witnessed a growing interest in automatic methods for creating trainable genera- tion components. Examples include learning which database records should be present in a text (Duboue and McKeown, 2002; Barzilay and Lapata, 2005) and how these should be verbalized (Liang et al., 2009). Besides concentrating on isolated compo- nents, 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 creating output that con- sists of a few sentences, thus obviating the need for document planning, or by treating sentence planning and surface realization as one component. A com- mon modeling strategy is to break up the genera- tion process into a sequence of local decisions, each learned separately (Reiter et al., 2005; Belz, 2008; Chen and Mooney, 2008; Angeli et al., 2010; Kim and Mooney, 2010). In this paper we describe an end-to-end gen- eration model that performs content …

Content truncated to 500 words