An LLM-powered chatbot for teaching English ergative structures to Chinese-speaking learners

Purpose – This study investigates the potential of large language models (LLMs, e.g., GPT-4) and chatbot technology to support the learning of English ergative structures among Chinese-speaking learners through automated, contrastive grammar analysis. The research focuses on developing computational methods that combine natural language processing (NLP) techniques with an interactive chatbot interface to identify and correct L1 transfer errors, thereby facilitating targeted grammar instruction. The study specifically examines how AI-generated metalinguistic feedback can address persistent challenges in the acquisition of ergative verbs.

Design/methodology/approach – The study employs a fully computational methodology that combines NLP techniques with interactive chatbot simulations. A hybrid system is developed, pairing rule-based error detection with GPT-4's generative capabilities to analyze learner sentences from the Chinese Learner English Corpora. The system diagnoses different types of ergative errors (including overpassivization and transitivity errors) and provides multilevel feedback: (1) error identification, (2) corrected forms, and (3) contrastive English-Chinese examples. Performance is evaluated through three computational metrics: error detection accuracy (vs. human-annotated benchmarks), feedback explicitness scoring, and syntactic pattern generalization capacity.

Findings –Results demonstrate the chatbot system's effectiveness in detecting and classifying ergative errors while producing pedagogically useful explanations. The contrastive feedback mechanism, which juxtaposes English corrections with Chinese equivalents, proves particularly valuable for addressing L1 transfer patterns. Qualitative analysis reveals that the most successful feedback sequences combine structural comparisons with clear metalinguistic explanations. System performance remains consistent across both common and less frequent ergative verb types.

Originality/value/implications – This research contributes to second language learning and educational technology by: (1) proposing a novel framework for human-free grammar instruction, (2) demonstrating LLMs' capacity to model L1-L2 structural contrasts, and (3) establishing best practices for AI-generated grammatical feedback. The findings enable future development of adaptive tutoring systems that address persistent challenges in syntactic acquisition, particularly for typologically distinct language pairs. Practical applications include ESL chatbot design and automated materials generation for ergative structures.