Under-resourced and Long-Tail CCG Supertagging

Address the unresolved problem of CCG supertagging for under-resourced languages and for the long tail of CCG categories, including by developing methods that improve prediction of rare categories.

Background

The paper identifies CCG supertagging as particularly difficult for under-resourced languages and for rare categories because CCG contains a large number of sparse, long-tail supertags. In the experiments, TagInsert improves over a left-to-right Transformer and performs competitively with BERT in several low-resource settings, while its performance on rare categories varies substantially across datasets and languages.

The authors suggest combining TagInsert’s dynamic generation order with constructive supertagging, which builds supertags rather than treating them as opaque labels. They motivate this combination because constructive approaches may better handle rare categories, while TagInsert can exploit flexible generation order and decoder-side context.