LangLingual: LLM-Powered Language Learning

• LangLingual is an LLM-driven, exercise-oriented platform that delivers personalized, real-time, context-aware feedback for effective English learning.
• It integrates advanced orchestration frameworks, granular proficiency modeling, and adaptive feedback pipelines to keep learners in their optimal challenge zone.
• Empirical evaluations highlight enhanced learner motivation, significant reductions in error repetition, and scalable AI-assisted language instruction.

LangLingual is an LLM-driven, exercise-oriented English language learning platform that delivers personalized, real-time, context-aware feedback through a conversational agent. Designed to address critical limitations of both traditional classroom and generic LLM-based learning tools, LangLingual integrates advanced orchestration frameworks, granular proficiency modeling, and adaptive feedback pipelines. Empirical results demonstrate strong usability, enhanced learner motivation, and tangible reductions in error repetition, positioning LangLingual as a research prototype for scalable, AI-augmented language instruction (Gupta et al., 27 Oct 2025).

1. Motivation and Design Principles

LangLingual was conceived to overcome persistent challenges in language education: the inability of classroom and MOOC environments to support highly individualized feedback and practice; the divergence of learner needs and proficiency within a cohort; and the lack of longitudinal progress tracking in general-purpose LLM interfaces. To address these deficits, the following design goals were established:

• Personalization: All exercises and feedback are dynamically tailored to the individual learner’s estimated proficiency and error patterns.
• Real-time, Socratic Feedback: Immediate correction of grammatical and lexical errors using hint-based guidance rather than direct answers, enabling active learner reflection and self-correction.
• Longitudinal Proficiency Modeling: Each user’s competence is tracked on a continuous 1–14 scale, with frequent updates and exercise difficulty adaptation.
• Context-awareness: Both exercises and feedback are conditioned on recent conversational context, allowing nuanced targeting of recurrent mistakes and topical relevance.

These objectives jointly support the principle of keeping learners in their “zone of proximal development”—by modulating challenge in response to evolving proficiency without overwhelming or under-challenging the user (Gupta et al., 27 Oct 2025).

2. System Architecture and Components

The deployed LangLingual system consists of loosely coupled modules orchestrated via the LangChain framework (v0.3), with supplementary storage and analytical backends:

• Front-End Interface: Streamlit application (Streamlit Cloud-deployed) supporting both text and audio input. Speech utterances are transcribed using OpenAI Whisper.
• Backend Orchestration: LangChain manages conversation state, prompt sequencing, memory, and retrieval-augmented generation (RAG) pipelines.
• LLM Backend: GPT-3.5-turbo or GPT-4 serve all generative tasks (language turns, exercise generation, feedback). The system is model-agnostic at this layer.
• Vector Store: ChromaDB provides embedding-based retrieval for relevant learning materials.
• Persistence Layer: Supabase/PostgreSQL authenticates users, persists chat logs, session metadata, and detected improvement areas; data isolation is enforced via row-level security.
• Proficiency Assessment Module: Fuses word-bank statistics and LLM-prediction to yield robust proficiency estimates.
• Exercise Generator: Monitors for “exercise keywords” in LLM outputs to instantiate and track active learning tasks.
• Feedback Module: Aggregates user utterances, invokes LLM-powered categorization of characteristic errors, and supplies targeted hinting.

The architecture is modular, separating UI, orchestration, content storage, and analytic logic for maintainability and extensibility (Gupta et al., 27 Oct 2025).

3. Algorithmic Pipelines and Proficiency Modeling

LangLingual operationalizes its instructional objectives via specialized pipelines:

3.1 Context-Aware Grammar Exercise Generation

1. The LLM produces a candidate reply in the flow of conversation.
2. An Exercise Detector scans responses for trigger keywords (“fill in the blank,” etc.).
3. If an exercise opportunity is identified, the system logs:
   • Exercise type
   • Prompt text
   • Conversation context
4. The learner’s response is recorded for downstream feedback.

3.2 Socratic, Real-Time Feedback

• Every learner utterance is sent to an LLM prompt that requests:
  • Identification of grammatical/lexical errors
  • Hints for correction, employing a Socratic (inductive) presentation style
• The LLM returns errors and hints in structured format (JSON/natural language)
• Feedback is immediately surfaced for self-correction or deeper explanation requests

3.3 Proficiency Scoring

A hybrid model combines:

• Word-bank analysis: Mean (or median) proficiency of lemma-matched vocabulary against a 50 000-term bank
• LLM prediction: Direct LLM-based 1–14 proficiency estimate

The combined score is:

$$\mathrm{Level}_\mathrm{combined} = w_\mathrm{wb}\,\mathrm{Level}_\mathrm{wb} + w_\mathrm{LLM}\,\mathrm{Level}_\mathrm{LLM}$$

with $w_\mathrm{wb}=0.4$, $w_\mathrm{LLM}=0.6$

This ensures stable performance even with out-of-vocabulary or atypical learner input.

3.4 Adaptation and Data Collection

• The proficiency scalar is adjusted after each conversational turn and exercise trial
• Exercise selection adapts in real time to maintain the learner within ±1 level of their current proficiency
• All chat data, error analyses, and exercise performance are logged for personalized analytic and progression purposes

No Bayesian or Elo-style update was implemented; only a weighted moving average is used for progression (Gupta et al., 27 Oct 2025).

4. Evaluation Methodology and Results

Evaluation comprised both human user sessions and scripted persona trials:

• Participants: Seven intermediate-to-advanced users (India, Japan, Indonesia, Australia, ages 18–44)
• Procedures: Each subject completed a full session; additional persona-based tests simulated specialized contexts (academic, job interview, professional presentation)
• Quantitative metrics: 5-point Likert scale ratings for usability, motivation, and learning effectiveness; feature-preference counts
• Qualitative data: Freeform survey comments and researcher observations on instructional clarity and context-sensitivity

Key findings:

• Real-time, individually tailored feedback was highly valued
• Most participants self-reported reductions in repeated errors (notably articles and tenses)
• Motivation for continued practice increased
• Persona runs confirmed that the system flexibly generated appropriate exercises across diverse communicative contexts
• No inferential statistical analysis was conducted; all evidence is descriptive (Gupta et al., 27 Oct 2025)

5. Limitations and Prospective Enhancements

LangLingual’s initial study identified several constraints:

• Small Sample Size and Limited Sessions: Only seven users, each with a single session—limits external validity
• No Objective Skill Assessment: Absence of pre/post standardized proficiency testing
• Proficiency Model Simplicity: Current model lacks pedagogically meaningful thresholds; progression relies on a continuous scalar average
• Occasional Context Drift: The LLM’s conversational context may degrade with increasing dialogue history, resulting in topic irrelevance

Proposed future improvements include gamification (badges, streaks), proactive engagement mechanisms (daily tasks), and refinement of the proficiency scale for alignment with CEFR descriptors (Gupta et al., 27 Oct 2025).

6. Implications for Language Education and Research Directions

LangLingual demonstrates that LLM-enabled systems can approximate core pedagogical functions: real-time grammar feedback, context-adaptive task generation, and fine-grained proficiency tracking in a scalable, remotely accessible framework. Integration of pedagogically-informed progression thresholds, comprehensive learning analytics, and standardized outcome benchmarks (CEFR, EFCAMDAT) are suggested for future iterations.

Research questions remain regarding the optimal weighting of rule-based versus model-based proficiency estimation, tuning the balance between hinting and explicit correction, and scaling up for rigorous comparative A/B testing with established language learning platforms (Gupta et al., 27 Oct 2025).

LangLingual thus acts as a blueprint for R&D at the intersection of LLMs and language pedagogy, offering evidence that AI-driven conversational systems can deliver not only scalable instruction but also personalized, context-sensitive formative assessment.