Combining Referring Expression Generation and Surface Realization: A Corpus-Based Investigation of Architectures

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1547–1557, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics Combining Referring Expression Generation and Surface Realization: A Corpus-Based Investigation of Architectures Sina Zarrieß Jonas Kuhn Institut f¨ur maschinelle Sprachverarbeitung University of Stuttgart, Germany sina.zarriess,jonas.kuhn@ims.uni-stuttgart.de Abstract We suggest a generation task that inte- grates discourse-level referring expression generation and sentence-level surface re- alization. We present a data set of Ger- man articles annotated with deep syntax and referents, including some types of im- plicit referents. Our experiments compare several architectures varying the order of a set of trainable modules. The results sug- gest that a revision-based pipeline, with in- termediate linearization, significantly out- performs standard pipelines or a parallel architecture. 1 Introduction Generating well-formed linguistic utterances from an abstract non-linguistic input involves making a multitude of conceptual, discourse-level as well as sentence-level, lexical and syntactic decisions. Work on rule-based natural language generation (NLG) has explored a number of ways to com- bine these decisions in an architecture, ranging from integrated systems where all decisions hap- pen jointly (Appelt, 1982) to strictly sequential pipelines (Reiter and Dale, 1997). While inte- grated or interactive systems typically face issues with efficiency and scalability, they can directly account for interactions between discourse-level planning and linguistic realization. For instance, Rubinoff (1992) mentions Example (1) where the sentence planning component needs to have ac- cess to the lexical knowledge that “order” and not “home” can be realized as a verb in English. (1) a. *John homed him with an order. b. John ordered him home. In recent data-driven generation research, the focus has somewhat shifted from full data-to-text systems to approaches that isolate well-defined subproblems from the NLG pipeline. In particular, the tasks of surface realization and referring ex- pression generation (REG) have received increas- ing attention using a number of available anno- tated data sets (Belz and Kow, 2010; Belz et al., 2011). While these single-task approaches have given rise to many insights about algorithms and corpus-based modelling for specific phenomena, they can hardly deal with aspects of the architec- ture and interaction between generation levels. This paper suggests a middle ground between full data-to-text and single-task generation, com- bining two well-studied NLG problems. We in- tegrate a discourse-level approach to REG with sentence-level surface realization in a data-driven framework. We address this integrated task with a set of components that can be trained on flexible inputs which allows us to systematically explore different ways of arranging the components in a generation architecture. Our main goal is to inves- tigate how different architectural set-ups account for interactions between generation decisions at the level of referring expressions (REs), syntax and word order. Our basic set-up is inspired from the Generating Referring Expressions in Context (GREC) tasks, where candidate REs have to be assigned to in- stances of a referent in a Wikipedia article (Belz and Kow, 2010). We have created a dataset of Ger- man texts with annotations that extend this stan- …

Content truncated to 500 words