Two methods for predicting the order of prenominal adjectives in English

Robert Malouf (Rijksuniversiteit Groningen)

The order of prenominal adjectival modifiers in English is governed by
complex and poorly understood constraints which are difficult or
impossible for a linguist to encode in a computational grammar.  For
parsing grammars, this is not a problem: the grammar may simply admit
all sequences of adjectives.  However, a natural language generation
system must assign a valid ordering.  

In this paper I explore and compare two methods for solving this problem. One approach is to apply a post-filter to an overgenerating symbolic grammar that uses an n-gram language model to select the most likely output. A more successful alternative method extends the work of Shaw and Hatzivassiloglou (1999) using the Memory-Based Learning techniques of, e.g., Daelemans, van den Bosch, and Weijters (1997). This approach fares better, correctly predicting the order of nearly 90% of the sequences in the test set. Given the variability inherent in the data, this is as well as any method would be able to do.