Approximate Maximum-Entropy Integration of Syntactic and Semantic Constraints

Statistical approaches to natural language parsing and interpretation have a number of advantages but thus far have failed to incorporate compositional generalizations found in traditional structural models. A major reason for this is the inability of most statistical language models being used to represent relational constraints, the connectionist variable binding problem being a prominent case. This paper proposes a basis for integrating probabilistic relational constraints using maximum entropy, with standard compositional feature-structure or frame representations. In addition, because full maximum entropy is combinatorically explosive, an approximate maximum entropy (AME) technique is introduced. As a sample problem, the task of integrating syntactic and semantic constraints for nominal compound interpretation is considered.