Level Management System
The Level Management System is the core engine of LingoFit. Our goal is to provide content that matches a user's proficiency so precisely that it feels natural. Early iterations revealed that naïve approaches—such as asking an LLM to "rate this user 1-100" or using a single CEFR score—were insufficient. They lacked transparency, were non-deterministic, and failed to capture the multi-dimensional nature of language difficulty (e.g., a user might know complex words but struggle with fast speech).
To solve this, we adopted an Implicit User Modeling approach grounded in Human-Computer Interaction (HCI) theory (Fischer, 2001).
- Explicit Modeling: Asking users to take quizzes (high friction).
- Implicit Modeling: Inferring difficulty from natural behaviors (low friction).
LingoFit primarily relies on implicit signals—pauses, rewinds, and vocabulary lookups—to estimate difficulty without interrupting the user's flow.
We quantify English listening ability across three independent axes, derived from the DLI-FLC Guide to Measuring Listening [1].
- Definition: The rarity and diversity of words used.
- Ground Truth: We utilize a "Closest Profile Match" algorithm based on research analyzing 17,000 sentences [2].
- Logic: When generating a script for a specific level (e.g., A2), the AI must strictly adhere to that level's statistical word distribution.
| CEFR Level | A1 Word % | A2 Word % | B1 Word % | B2 Word % |
|---|---|---|---|---|
| A1 | 66.3% | 15.2% | 4.8% | 1.3% |
| A2 | 54.6% | 18.2% | 10.1% | 3.2% |
| B1 | 41.7% | 20.1% | 15.5% | 5.9% |
- Definition: The complexity of sentence structures (e.g., clause density).
- Ground Truth: We measure this via Average Sentence Length (ASL).
- Logic: The system targets specific "Success Ranges" for ASL.
| Level | ASL Target (Words per Sentence) |
|---|---|
| A1 | 6.7 ~ 8.7 |
| A2 | 9.9 ~ 11.9 |
| B1 | 13.7 ~ 16.7 |
- Definition: The cognitive load required to process spoken sounds.
- Ground Truth: We reference research on natural speech rates and TED Talk averages [3].
- Logic: We map proficiency to specific WPM (Words Per Minute) ranges.
| Level | WPM Range | Multiplier |
|---|---|---|
| A1 | 110-130 | 0.7x |
| B1 | 130-150 | 1.0x |
| C2 | 180+ | 1.2x |
The LMS updates the user's profile after every session using a feedback loop that processes both implicit and explicit signals.
The update mechanism functions as follows:
-
Input Vector: We aggregate 6 signals into a normalized vector:
- Implicit: Pause Count, Rewind Count, Vocab Lookups, Vocab Saves.
- Explicit: User Understanding Rating, User Speed Rating.
-
Weight Matrix: A fixed 6x3 matrix (
DEFAULT_WEIGHT_MATRIX) maps these signals to the three proficiency axes.- Example: A "Vocab Save" heavily weights the Lexical update.
- Example: A "Rewind" affects both Auditory and Syntactic scores.
-
Clipping: Updates are capped at
[–8, +8]per session to ensure smooth, stable progression rather than volatile jumps. - Finalize: The calculated deltas are applied to the user's profile in the database.