Collective Content Selection for Concept-to-Text Generation

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 331–338, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Collective Content Selection for Concept-To-Text Generation Regina Barzilay Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology regina@csail.mit.edu Mirella Lapata School of Informatics University of Edinburgh mlap@inf.ed.ac.uk Abstract A content selection component deter- mines which information should be con- veyed in the output of a natural language generation system. We present an effi- cient method for automatically learning content selection rules from a corpus and its related database. Our modeling frame- work treats content selection as a col- lective classification problem, thus allow- ing us to capture contextual dependen- cies between input items. Experiments in a sports domain demonstrate that this approach achieves a substantial improve- ment over context-agnostic methods. 1 Introduction Content selection is a fundamental task in concept- to-text generation (Reiter and Dale, 2000). A practi- cal generation system typically operates over a large database with multiple entries that could potentially be included in a text. A content selection compo- nent determines what subset of this information to include in the generated document. For example, consider the task of automatically generating game summaries, given a database con- taining statistics on Americal football. Table 1 shows an excerpt from such a database, and its cor- responding game summary written by a journalist. A single football game is typically documented in hundreds of database entries — all actions, player positions, and scores are recorded, along with a wide range of comparative and aggregate statistics. Only a small fraction of this information is featured in a game summary. The content selection component aims to identify this subset.1 In existing generation systems the content se- lection component is manually crafted. Specify- ing content selection rules is, however, notoriously difficult, prohibitively so in large domains. It in- volves the analysis of a large number of texts from a domain-relevant corpus, familiarity with the associ- ated database, and consultation with domain experts. Moreover, the task must be repeated for each domain anew. This paper proposes a data-driven method for learning the content-selection component for a concept-to-text generation system. We assume that the learning algorithm is provided with a parallel corpus of documents and a corresponding database, in which database entries that should appear in doc- uments are marked. One possible approach is to formulate content se- lection as a standard binary classification task: pre- dict whether an item is to be included on the basis of its attributes alone. In fact, this method is com- monly used for content selection in text summariza- tion (e.g., Kupiec et al., 1995). However, by treating each instance in isolation, we cannot guarantee that the selected database entries are related in a mean- ingful way, which is essential for the generation of a coherent text. Rather than selecting each item separately, we propose a method for collective content selection, where all candidates are considered simultaneously for selection. Collective selection thereby allows …

Content truncated to 500 words