Design And Evaluation Of An Agreement Error Detection System: Testing The Effect Of Ambiguity, Parser And Corpus Type

We present a system for the detection of agreement errors in Basque, a language with agglutinative morphology and free order of the main sentence constituents. Due to their complexity, agreement errors are one of the most frequent error types found in written texts. As the constituents concerning agreement can appear in any order in the sentence, we have implemented a system that makes use of dependency trees of the sentence, which abstract over specific constituent orders. We have used Sairoi, a tool that obtains the analysis trees that fulfill a set of restrictions described by means of declarative rules. This tool is applied to the output of two dependency analyzers: Maltha (data-driven) and EDGK (rule-based). The system has been evaluated on two corpora: a. group of texts containing errors, and another one composed of correct texts. As a secondary result, we have also estimated a. measure of the impact of syntactic ambiguity on the quality of the results.