تأثير زبان اوّل بر دقّت دسته بندها در برچسب دهي خودكار خطاهاي املايي در پيكره ي زبان آموز

Given the existing research gap in developing automatic spelling correction systems tailored to the needs of Persian language learners, this study investigates the impact of metadata information namely nationality an‎d first language (L1) on the accuracy of classifiers in the automatic tagging of spelling errors. In this research, 637 text documents, including spelling an‎d composition texts from learners of various nationalities, were collected, with a focus on four nationalities: Urdu, Arabic, Armenian, an‎d Russian speakers. This data was extracted from the Language Center of the University of Isfahan, Bent-Al-Hoda School, Al-Mustafa International University (Qom), Armenian High School, an‎d the Saadi Foundation Corpus. Spelling errors were coded using the INCEpTION software based on a stan‎dard framework. The error tagging was performed by the researcher. This framework encompasses three hierarchical levels: Error Domain, Error Category, an‎d Error Process. It includes four main categories (Primary Signs, Secondary Signs, Punctuation Marks, an‎d Form Category) an‎d four error processes (Reduction, Addition, Transposition, an‎d Substitution). Following the study by Taghavi Sani (1402/2023), this research analyzed the data from two perspectives: first, a statistical analysis of error frequencies based on nationality an‎d L1; an‎d second, an assessment of classifier prediction accuracy under two distinct conditions: “with consideration” an‎d “without consideration” of linguistic an‎d national features. The findings indicate that incorporating L1 an‎d nationality information significantly enhances the accuracy of classifiers in the automatic tagging of spelling errors within the learner corpus, particularly in linguistically heterogeneous corpora, leading to better performance .eva‎luation results demonstrated that the overall accuracy of the models improved in most cases when the nationality variable was included. In the Ran‎dom Forest model using the Index_Encode method, the overall accuracy was 58% without considering nationality, which increased to 62% upon its inclusion. For the same method, the Support Vector Machi ne model achieved an accuracy of 53% in the “without nationality” condition an‎d 55% in the “with nationality” condition. In the OneHot_Encode method, the RF model achieved an accuracy of 63% without nationality, with no significant change observed upon adding the nationality variable (remaining around 61%). Conversely, the SVM model showed better performance in this method, with accuracy increasing from 66% (without nationality) to 68% (with nationality). Overall, the results suggest that the nationality variable plays a positive role in improving the performance of both machine learning models across both encoding methods, leading to increased accuracy in most cases. However, the extent of this impact was dependent on the specific model an‎d encoding technique used. Future research is recommended to collect larger an‎d more diverse corpora to more precisely investigate the effect of L1 on spelling errors an‎d enhance the generalizability of the findings.

5

159047