Complex Predicates are Multi-Word Expressions

Practitioners of English Natural Language Processing often feel fortunate because their tokens are clearly marked by spaces on either side. However, the spaces can be quite deceptive, since they ignore the boundaries of multi-word expressions, such as noun-noun compounds, verb particle constructions, light verb constructions and constructions from Construction Grammar, e.g., caused-motion constructions and resultatives. Correctly identifying and handling these types of expressions can be quite challenging, even from the viewpoint of manual annotation. This talk will review the pervasive nature of these constructions, touching on Arabic and Hindi as well as English. Using several illustrative examples from newswire and medical informatics, current best practices for annotation and automatic identification will be described, with an emphasis on contributions from predicate argument structures.