Computational Models of Problems with Writing of English as a Second Language Learners

Learning a new language is a challenging endeavor. As a student attempts to master the grammar usage and mechanics of the new language, they make many mistakes. Detailed feedback and corrections from language tutors are invaluable to student learning, but it is time consuming to provide such feedback. In this thesis, I investigate the feasibility of building computer programs to help to reduce the efforts of English as a Second Language (ESL) tutors. Specifically, I consider three problems: (1) whether a program can identify areas that may need the tutor’s attention, such as places where the learners have used redundant words; (2) whether a program can auto-complete a tutor’s corrections by inferring the location and reason for the correction; (3) for detecting misusages of prepositions, a common ESL error type, whether a program can automatically construct a set of potential corrections by finding words that are more likely to be confused with each other (known as a confusion set).\ud \ud The viability of these programs depends on whether aspects of the English language and common ESL mistakes can be described by computational models. For each task, building computational models faces unique challenges: (1) In highlighting redundant areas, it is difficult to precisely define “redundancy” in a computer’s language. (2) In auto-completing tutors’ annotations, it is difficult for computers to correctly interpret how many writing problems were addressed during revision. (3) In confusion set construction, it is difficult to infer which words are more likely confused with the given word. To address these challenges, this thesis presents different model alternatives for each task. Empirical experiments demonstrate the degrees of success to which computational models can help with detecting and correcting ESL writing problems\u