Domain-oriented two-stage aggregation: Generating baseball play-by-play narratives

This paper presents an end-to-end natural language generation system that performs aggregation in two stages: the first takes advantage of the information implicit in the source knowledge base in order to aggregate event components into complex sentences. The second stage examines the developing context of the text in order to aggregate similar adjacent events into more fluent text. The source knowledge base is the Retrosheet collection of play-by-play baseball scoresheets encoded in machine-readable form. The output is reasonably fluent and natural, human-readable play-by-play narratives of historical baseball games. The system was tested against all regular season major league games played from 1950 to 1969, taking less than a second to produce three to five pages of text for each game. The aggregation achieved resulted in a substantial improvement in native speaker judgments of fluency and readability.