User Modeling in Language Learning with Macaronic Texts

Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics, pages 1859–1869, Berlin, Germany, August 7-12, 2016. c⃝2016 Association for Computational Linguistics User Modeling in Language Learning with Macaronic Texts Adithya Renduchintala and Rebecca Knowles and Philipp Koehn and Jason Eisner Department of Computer Science Johns Hopkins University {adi.r,rknowles,phi,eisner}@jhu.edu Abstract Foreign language learners can acquire new vocabulary by using cognate and con- text clues when reading. To measure such incidental comprehension, we devise an experimental framework that involves reading mixed-language “macaronic” sen- tences. Using data collected via Ama- zon Mechanical Turk, we train a graphi- cal model to simulate a human subject’s comprehension of foreign words, based on cognate clues (edit distance to an English word), context clues (pointwise mutual in- formation), and prior exposure. Our model does a reasonable job at predicting which words a user will be able to understand, which should facilitate the automatic con- struction of comprehensible text for per- sonalized foreign language education. 1 Introduction Second language (L2) learning requires the ac- quisition of vocabulary as well as knowledge of the language’s constructions. One of the ways in which learners become familiar with novel vocab- ulary and constructions is through reading. Ac- cording to Krashen’s Input Hypothesis (Krashen, 1989), learners acquire language through inciden- tal learning, which occurs when learners are ex- posed to comprehensible input. What constitutes “comprehensible input” for a learner varies as their knowledge of the L2 increases. For example, a student in their first month of German lessons would be hard-pressed to read German novels or even front-page news, but they might understand brief descriptions of daily routines. Comprehen- sible input need not be completely familiar to the learner; it could include novel vocabulary items or structures (whose meanings they can glean from context). Such input falls in the “zone of proxi- mal development” (Vygotski˘ı, 2012), just outside of the learner’s comfort zone. The related con- cept of “scaffolding” (Wood et al., 1976) consists of providing assistance to the learner at a level that is just sufficient for them to complete their task, which in our case is understanding a sentence. Automatic selection or construction of comprehensible input—perhaps online and personalized—would be a useful educational technology. However, this requires modeling the student: what can an L2 learner understand in a given context? In this paper, we develop a model and train its parameters on data that we collect. For the remainder of the paper we focus on native English speakers learning German. Our methodology is a novel solution to the problem of controlling for the learner’s German skill level. We use subjects with zero previous knowledge of German, but we translate portions of the sentence into English. Thus, we can presume that they do already know the English words and do not al- ready know the German words (except from see- ing them in earlier trials within our experiment). We are interested in whether they can jointly infer the meanings of the remaining German words in the sentence, so …

Content truncated to 500 words