Konferensartikel

This paper presents a method for the automatic generation and semantic evaluation of exercise sentences for Esperanto teaching. Our sentence grader exploits both corpus data and lexical resources (verb frames and noun/adjective ontologies) to either generate meaningful sentences from scratch, or to determine the acceptability of a given input sentence. Individual words receive scores for how well they match the semantic conditions projected onto their place in the syntactic tree. In a CALL context, the system works with a lesson-/level-constrained vocabulary and can be integrated into e.g. substitution table or slot filler exercises. While the method as such is language-independent, we also discuss how morphological clues (affixes) can be exploited for semantic purposes. When evaluated on out-of-corpus course materials and short stories, the system achieved a rejection precision, in terms of false positives, of 98-99% at the sentence level, and 93-97% at the word level.

corpus- and parsing-based CALL, sentence generation, sentence grading, semantic sentence evaluation, Esperanto teaching, AWE