ABSTRACT
The development of symbolic machine translation (MT) systems is difficult and expensive, but the development of non-symbolic MT systems is typically not much easier or less expensive. An ideal MT system would be able to identify the structure of the source and target languages without the assistance of human engineers, but at the same time be easily understood, corrected and updated to reflect new information and situations. For the second, a symbolic approach is almost a necessity, as sub-symbolic or completely automatic inference systems tend to produce nearly opaque sets of numbers as their outputs. However, the process of describing human languages in symbolic terms is difficult and knowledge-intensive. A prototype system, Metla, is described which derives linguistically plausible and understandable transfer functions from paired corpora without the necessity of human intervention or pre-analysis.
