Cyclic feeding interactions between finite-state mal-rules

Intelligent Language Tutoring Systems typically attempt to automatically diagnose learner errors in order to provide individualized feedback. One common approach is the use of mal-rules to extend norma- tive grammars by licensing specific types of learner errors. In finite-state morphologies, mal-rules can be implemented as two-level rules or replace rules. However, unlike the phonological rules of natural languages, mal-rules do not necessarily behave as a coherent system, especially with respect to feed- ing interactions. Using examples from learner errors attested in the RULEC corpus of Russian learner texts, we illustrate the problem of cyclic feeding interactions that can occur between mal-rules. We then describe a formal algorithm for identifying an optimal ordering for mal-rules to be applied to a transducer.