Leveraging large language models to supplement corpus-based inductive learning of Chinese as a second language

Corpus tools have proven effective for supporting inductive language learning by enabling learners to observe multiple examples, form hypotheses, and verify the hypotheses based on additional examples. However, when applied to Chinese as a Second Language (CSL), these tools encounter limitations that disrupt the observe-hypothesize-verify process. Sketch Engine, for example, misanalyzes Chinese word boundaries, topicalized objects, and ba -constructions, and provides inaccurate observational data that undermines the effectiveness of inductive learning. This paper proposes integrating Large Language Models (LLMs) with corpus tools to address the limitations. Using Sketch Engine and Claude Opus 4 as exemplars, I demonstrate how LLMs serve three pedagogical functions: (1) error detection to identify misanalyzed features in corpus outputs, (2) guided pattern discovery to help learners recognize linguistic regularities across examples, and (3) hypothesis verification to confirm/refine learners’ observations. Through analysis of specific Chinese features, I show how LLM integration maintains the discovery processes while ensuring accurate linguistic input for the learners. The proposed corpus-LLM integration represents an advancement in leveraging AI for language pedagogy. The paper concludes with future research directions for optimizing this integration in CSL acquisition, and emphasizes the need to balance technological innovation with pedagogical principles.