A Classification-driven Approach to Document Planning

International Conference RANLP 2009 - Borovets, Bulgaria, pages 324–329 A Classification-driven Approach to Document Planning Rafael L. de Oliveira, Eder M. de Novais, Roberto P. A. de Araujo, Ivandré Paraboni University of São Paulo (USP) School of Arts, Sciences and Humanities (EACH) Av. Arlindo Bettio, 1000 - São Paulo, Brazil { rafaellage, eder.novais, roberto.araujo, ivandre } @usp.br Abstract Document Planning - the task of deciding which content messages should be realised in a target document based on raw data provided by an underlying application, and how these messages should be structured - is arguably one of the most crucial tasks in Natural Language Generation (NLG). In this work we present a machine learning approach to Document Planning that is entirely trainable from annotated corpora, and which paves the way to our long-term goal of developing a text generator system based on a series of classifiers for a simple NLG application in the education domain. Keywords Document Planning; Content Selection. 1. Introduction Natural Language Generation (NLG) systems are used whenever simple, 'canned' text is not sufficient, and greater (i.e., closer to human performance) linguistic variation is required. The traditional NLG architecture is often depicted in simplified form as a 3-stages pipelined process (Document Planning, Sentence Planning and Surface Realisation, cf. [1]), a division that is at least partially motivated by the sheer complexity of the task. Starting from a high-level communicative goal of describing a given domain concept, the system builds up a plan to represent the input data up to the point in which fully-specified text in natural language is produced. The Document Planning module is responsible for deciding what information to communicate (this being the task of Content Determination) and then how this information should be structured for presentation (this being the task of Document Structuring.) For a more comprehensive discussion of the role of Document Planning and its subtasks in the NLG architecture we report to [3]. Document Planning is arguably one of the most crucial components of an NLG system [1]: if a generated document presents the required information in a reasonably coherent structure, then the system may be considered successful even if the text shows surface flaws or limited linguistic variation. On the other hand, if the required information is missing from the text, or if the text is poorly structured, then the overall results are most likely unsatisfactory regardless of how well the individual sentences were realised. When speaking of Data-to-Text generation1, Document Planning is often preceded by a Data 1 For a large-scale application of this kind, see [5]. Interpretation stage [2] that processes raw data application in the first place. In what follows, we discuss the early stages of a simple Data-to-Text NLG application addressing some aspects of both issues, namely, which chunks of information – or messages - should be included in the generated text from the raw data provided by an underlying application, and how such messages should be structured within a standard RST framework [4]. We will argue that at …

Content truncated to 500 words