On Lexical Aggregation and Ordering

Aggregation is the process of removing redundant information during language generation while preserving the information to be conveyed. Aggregation is an important component of text or sentence planning. Without aggregation, automated language generation systems would not be able to p roduce f luent text f rom real -wor ld databases and knowledge bases, since information is rarely stored in computers in forms directly supporting fluent expression. Various types of aggregation (syntactic, lexical, referential) have been identified in [Hovy88, Cook84, Reinhart91, Horacek92, Dalianis&Hovy93,Wilkinson95,Dalianis95a, 95b,96a]. This paper investigates lexical aggregation, the process by which a set of items is replaced with a single new lexeme that encompasses the same meaning. We call the elements that will be aggregated the aggregands and the element (the lexeme) which is the result of the aggregation the aggregator. Lexical aggregation can be divided into two major types, bounded and unbounded. With Bounded Lexical (BL) aggregation the aggregator lexeme covers a closed set of concepts and the redundancy is obvious, the aggregated information is recoverable, and the aggregation process must be carried out. In contrast, Unbounded Lexical (UL) aggregation is carried out over an open set of aggregands and consequently the aggregated information is not recoverable and has to be licensed by other factors, such as the hearer's goals. The example in Figure 1 contains both types of aggregation, where f ight and week are the unbounded and bounded aggregators respectively.