Knowledge Tracing in Sequential Learning of Inflected Vocabulary

Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), pages 238–247, Vancouver, Canada, August 3 - August 4, 2017. c⃝2017 Association for Computational Linguistics Knowledge Tracing in Sequential Learning of Inflected Vocabulary Adithya Renduchintala and Philipp Koehn and Jason Eisner Department of Computer Science Johns Hopkins University {adi.r,phi,eisner}@jhu.edu Abstract We present a feature-rich knowledge tracing method that captures a student’s acquisition and retention of knowledge dur- ing a foreign language phrase learning task. We model the student’s behavior as making predictions under a log-linear model, and adopt a neural gating mechanism to model how the student updates their log-linear parameters in response to feedback. The gating mechanism allows the model to learn complex patterns of retention and acquisi- tion for each feature, while the log-linear parameterization results in an interpretable knowledge state. We collect human data and evaluate several versions of the model. 1 Introduction Knowledge tracing attempts to reconstruct when a student acquired (or forgot) each of several facts. Yet we often hear that “learning is not just memorizing facts.” Facts are not atomic objects to be discretely and independently manipulated. Rather, we suppose, a student who recalls a fact in a given setting is demonstrating a skill—by solving a structured prediction problem that is akin to recon- structive memory (Schacter, 1989; Posner, 1989) or pattern completion (Hopfield, 1982; Smolensky, 1986). The attempt at structured prediction may draw on many cooperating feature weights, some of which may be shared with other facts or skills. In this paper, for the task of foreign-language vo- cabularylearning, wewilladoptaspecificstructured prediction model and learning algorithm. Different knowledge states correspond to model parameter settings (feature weights). Different learning styles correspond to different hyperparameters that govern the learning algorithm.1 As we interact with each student through a simple online tutoring system, we 1In the present paper, we assume that all students share the same hyperparameters (same learning style), although each stu- dentwillhavetheirownparameters, whichchangeastheylearn. would like to track their evolving knowledge state and identify their learning style. That is, we would like to discover parameters and hyperparameters that can explain the evidence so far and predict how the student will react in future. This could help us make good future choices about how to instruct this student, although we leave this reinforcement learning problem to future work. In this paper, we show that we can predict the student’s next answer. In short, we expand the notion of a knowledge tracing model to include representations for a student’s (i) current knowledge, (ii) retention of knowledge, and (iii) acquisition of new knowledge. Our reconstruction of the student’s knowledge state remains interpretable, since it corresponds to the weights of hand-designed features (sub-skills). Interpretability may help a future teaching system provide useful feedback to students and to human teachers, and help it construct educational stimuli that are targeted at improving particular sub-skills, such as features that select correct verb suffixes. Our present paper considers a verb conjugation task, where a foreign language learner learns the verb conjugation paradigm …

Content truncated to 500 words