Academic Commons

Presentations (Communicative Events)

An architecture for aggregation in text generation

McKeown, Kathleen; Shaw, James

A traditional natural language generation system architecture consists of a content planner and a surface realizer. The content planner packages information as verb based clausesized propositions, each of which is realized as a single sentence. The surface realizer maps the semantic propositions into actual sentences. In practical applications, many of these propositions share common features, such as the entity being described or discussed. If a generation system simply generates each proposition as a sentence, the output will contain many repetitive and redundant references to common features. A better approach is to detect shared entities among the adjacent propositions and combine them to remove redundancies. To achieve such capability, we added a sentence planner between the content planner and surface realizer. Its main task is aggregation, the combining of semantically related propositions in order to produce concise and fluent expressions. During the aggregation process before an operator is applied to propositions the lexicon is consulted to make sure that the operator is applicable This guarantees that the combined proposition can be realized as a surface string. The system also uses an ontology to generate referring expressions and generalization, both of which result in concise expressions.


More About This Work

Academic Units
Computer Science
Proceedings of the 15th International Joint Conference on Artificial Intelligence, Poster session
Published Here
April 29, 2013