Reference Specification in Multilingual Document Production

Abstract

To produce documents in multiple languages automatically requires a language-independent representation of the documents' meaning. For a person to build this language-independent representation by communicating in natural language with a computer system, the problem of reference must be addressed. This problem, inherent in natural language, presents itself not only in the specification of the language-independent representation, but also in the generation of documents with the meaning contained in this representation. This thesis presents methods to make both the specification of entities in the user interface and the generation of expressions to refer to these entities in documents more natural and provides empirical evidence demonstrating the efficacy of these methods. More specifically, this thesis describes the development of three types of reference mechanisms: a statistical model that uses domain and lexical knowledge to organize new options in the interface; techniques for controlling coreference specification that take advantage of discourse structure and genre features; and automatically learned models for generating expressions to refer to new and already mentioned entities in a particular domain. The evaluation of these reference mechanisms establishes that specifying new entities using an interface formed by computational linguistic processing reduces the amount of time required to refer to entities in the interface; exploiting discourse structure and genre features is more helpful than traditional knowledge-editing interfaces for referring to entities in the interface that are already contained in the knowledge representation; and using learned linguistic information to generate referring expressions in documents leads to expressions that more closely match the decisions of people.