Browse Theses

Machine translation (MT) is the translation of natural language (NL) text by computer, mapping a source language (SL) text onto a target language (TL) while preserving the meaning. Interlingua-based MT systems first map the SL into an intermediate language or interlingua (IL) and then map the IL out to the TL. There are, however, no established specifications that are generally accepted among MT researchers for constructing an interlingua.

This thesis addresses the basic question of what belongs in an interlingua. The work demonstrates that the following “division of labor” is viable in an MT system: an interlingua can be constructed as its own level of representation, distinct from NL syntactic and conceptual levels of representation and linguistically relevant to the task of preserving meaning in MT. This IL captures a level of abstraction between NL syntactic representations and conceptual or knowledge representations (KR).

The thesis focuses on the translation of spatial expressions, i.e., natural language sentences that convey the location, orientation, or motion of physical objects in the real, 3-dimensional world. The focus is both computationally and linguistically motivated: localist research holds that spatial expressions are more basic structurally and semantically than expressions in other fields and, thus, are critical for both language learners and computational systems in establishing the basic linguistic relations in other types of expressions.

The research presented in the thesis spells out the consequences of the division-of-labor approach for an MT system. The contributions of the thesis are: (i) the design of a modular, multi-level MT system, (ii) the formalization of a theory of tiered IL forms and associated processing algorithms, and (iii) the encoding of single-language and cross-language generalizations in an IL for spatial expressions.