A Statistical NLG Framework for Aggregated Planning and Realization

Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics, pages 1406–1415, Sofia, Bulgaria, August 4-9 2013. c⃝2013 Association for Computational Linguistics A Statistical NLG Framework for Aggregated Planning and Realization Ravi Kondadadi∗, Blake Howald and Frank Schilder Thomson Reuters, Research & Development 610 Opperman Drive, Eagan, MN 55123 firstname.lastname@thomsonreuters.com Abstract We present a hybrid natural language gen- eration (NLG) system that consolidates macro and micro planning and surface re- alization tasks into one statistical learn- ing process. Our novel approach is based on deriving a template bank automatically from a corpus of texts from a target do- main. First, we identify domain specific entity tags and Discourse Representation Structures on a per sentence basis. Each sentence is then organized into semanti- cally similar groups (representing a do- main specific concept) by k-means cluster- ing. After this semi-automatic processing (human review of cluster assignments), a number of corpus–level statistics are com- piled and used as features by a ranking SVM to develop model weights from a training corpus. At generation time, a set of input data, the collection of semanti- cally organized templates, and the model weights are used to select optimal tem- plates. Our system is evaluated with au- tomatic, non–expert crowdsourced and ex- pert evaluation metrics. We also introduce a novel automatic metric – syntactic vari- ability – that represents linguistic variation as a measure of unique template sequences across a collection of automatically gener- ated documents. The metrics for generated weather and biography texts fall within ac- ceptable ranges. In sum, we argue that our statistical approach to NLG reduces the need for complicated knowledge-based ar- chitectures and readily adapts to different domains with reduced development time. ∗*Ravi Kondadadi is now affiliated with Nuance Commu- nications, Inc. 1 Introduction NLG is the process of generating natural-sounding text from non-linguistic inputs. A typical NLG system contains three main components: (1) Doc- ument (Macro) Planning - deciding what content should be realized in the output and how it should be structured; (2) Sentence (Micro) planning - generating a detailed sentence specification and selecting appropriate referring expressions; and (3) Surface Realization - generating the final text after applying morphological modifications based on syntactic rules (see e.g., Bateman and Zock (2003), Reiter and Dale (2000) and McKeown (1985)). However, document planning is arguably one of the most crucial components of an NLG system and is responsible for making the texts ex- press the desired communicative goal in a coher- ent structure. If the document planning stage fails, the communicative goal of the generated text will not be met even if the other two stages are perfect. While most traditional systems simplify develop- ment by using a pipelined approach where (1-3) are executed in a sequence, this can result in er- rors at one stage propagating to successive stages (see e.g., Robin and McKeown (1996)). We pro- pose a hybrid framework that combines (1-3) by converting data to text in one single process. Most NLG systems fall into two broad …

Content truncated to 500 words