Structural Tags, Annealing and Automatic Word Classification

This paper describes an automatic word classification system which uses a locally optimal annealing algorithm and average class mutual information. A new word-class representation, the structural tag is introduced and its advantages for use in statistical language modelling are presented. A summary of some results with the one million word LOB corpus is given; the algorithm is also shown to discover the vowel-consonant distinction and displays an ability to cluster words syntactically in a Latin corpus. Finally, a comparison is made between the current classification system and several leading alternative systems, which shows that the current system performs respectably well.