A data-driven CEFR vocabulary profiler that solves the Polysemy Flatline Problem by disambiguating word senses (A1 to C2) in context.
Watch on YT
Standard CEFR text profilers (such as VocabKitchen or conventional wordlist matchers) assign each word a static difficulty level based solely on lemma frequency. This creates the Polysemy Flatline: polysemous words are systematically flattened to their lowest, beginner level regardless of how they are used.
Consider how the word party shifts difficulty across contexts:
| Sentence Context | Traditional Profiler | SenseCEFR (Context-Aware) | Actual Pedagogical Meaning |
|---|---|---|---|
| "We celebrated at a wonderful birthday party with my best friends." | A1 | A1 | Social celebration |
| "Both political parties debated the election campaign vigorously." | A1 (Incorrect) | B1 | Political organization |
| "The binding contract was formally signed by each legal party." | A1 (Incorrect) | C1 | Legal entity / contracting body |
By misclassifying advanced academic or legal senses as beginner vocabulary, traditional tools severely underestimate the true linguistic difficulty of textbooks, graded readers, examinations, and student essays.
SenseCEFR analyzes entire sentences to identify multi-tier polysemous tokens and resolves their exact CEFR level (A1 to C2) through a two-tier hybrid architecture:
Input Text
│
[Tokenizer & Sentence Splitter]
│
[Rule-Based Lemmatization]
(O(1) inflection map + suffix heuristics)
│
[Polysemy Discrepancy Gate]
Does lemma span multiple CEFR levels?
/ \
NO YES (15-20% of tokens)
│ │
[Direct Lookup] [Word Sense Disambiguation]
(0ms, instant) • Local Context Matcher (Built-in)
• Google Gemini AI Contextual Reasoning
│
[CEFR Level Tier Assignment]
(A1, A2, B1, B2, C1, C2)
│
[Click-to-Inspect Modal]
• Guideword & Sense Definition
• Candidate sense levels & Confidence score
• Manual level selection with live recalculation
- Evaluates tokens across the CEFR scale: A1, A2, B1, B2, C1, and C2.
- Off-list: Terms outside the pedagogical corpus are clearly distinguished.
- Unresolved Senses: Ambiguous or unmapped senses are flagged with
⚑for immediate pedagogical review.
Tokens whose difficulty level was adjusted through word sense disambiguation are marked with ✦ so educators and linguists can easily identify where context altered the vocabulary tier.
Click any profiled token to open the Word Inspector:
- View the active dictionary definition, guideword, and assigned CEFR tier.
- Inspect context confidence scores (e.g.
98% Context Match). - Review rejected candidate senses with the option to click ✓ Select this Sense to manually override the level. All profile metrics and distributions update in real time.
Open Settings (gear icon) to access the built-in lexicon customization center:
- Single Word Overrides: Assign custom CEFR levels to emerging or domain-specific terminology (e.g.
blockchain→C1). - Multi-Sense Polysemy: Define multiple senses for a single word with distinct CEFR tiers, parts of speech, guidewords, and definitions. Both the local matcher and Gemini AI will automatically disambiguate between them.
- CSV Import & Backup: Bulk import vocabulary lists directly from Excel/Google Sheets, export timestamped backups, or reset to factory defaults with 1 click.
Optionally connect your Google Gemini API key (e.g., gemini-flash-lite-latest) for:
- Zero-shot deep reasoning on complex figurative, idiomatic, and metaphorical usage.
- Plain-English pedagogical rationales explaining why a sense was selected.
Click Save Report to generate a comprehensive CSV report containing token counts, type-token ratios, CEFR distributions, and a complete audit log of all disambiguated polysemous words.
SenseCEFR harmonizes multiple academic corpora:
- 8,889 total headwords (CEFR-J 1.5, Octanove C1/C2, WordNet-CEFR, and Academic Word List overlay).
- 1,206 polysemous multi-CEFR words with 3,936 distinct sense definitions and contextual collocation vectors.
- 105 phrasal verb constructions (e.g., bank on, figure out, look after, carry out).
-
32,013 inflection-to-lemma pairings for instant
$O(1)$ morphological resolution.
Access the profiler directly in your web browser: https://nlp4ed.com/SenseCEFR/
Download the source code as a ZIP (or clone this repository) and open index.html in any web browser:
git clone https://github.com/NLP4Ed/SenseCEFR.git
cd SenseCEFRIf you use SenseCEFR in academic research, textbook leveling, or curriculum design, please cite it using CITATION.cff or the following BibTeX entry:
@software{sensecefr2026,
author = {Lchgary, Yassine},
title = {SenseCEFR: Context-Aware Vocabulary Profiler with Word Sense Disambiguation},
year = {2026},
version = {0.5.0},
doi = {10.5281/zenodo.23318092},
url = {https://nlp4ed.com/SenseCEFR/}
}SenseCEFR is built upon several foundational open-source datasets, academic lexical databases, and natural language processing research.
-
CEFR-Annotated WordNet (Kikuchi et al.)
- Role: Primary data source for sense-level CEFR proficiency annotations mapped across WordNet synsets.
- License: CC BY 4.0
- Resource: Hugging Face Dataset
-
Princeton WordNet 3.0 & Morphy (Miller et al.)
- Role: English lexical database providing synset definitions and glosses, plus Morphy morphological exception tables (
*.exc) for client-side irregular lemmatization. - License: WordNet 3.0 License (Permissive, Copyright 2006 Princeton University)
- Resource: Princeton WordNet
- Role: English lexical database providing synset definitions and glosses, plus Morphy morphological exception tables (
-
CEFR-J Lexical Database (v1.5) (Tokyo University of Foreign Studies)
- Role: Core graded vocabulary framework providing baseline A1–B2 word-family proficiency targets.
- License: Open Educational / Academic Citation
- Resource: CEFR-J Project · Open Language Profiles CEFR-J
-
Octanove C1/C2 Vocabulary Profiles (Octanove Labs)
- Role: Advanced lexical profiles extending SenseCEFR coverage to C1 and C2 proficiency levels.
- License: CC BY-SA 4.0
- Resource: Open Language Profiles (CEFR-J & Octanove)
-
Oxford 3000™ / 5000™ Wordlists (Oxford University Press)
- Role: Graded frequency wordlists, pedagogical definitions, and guidewords used to benchmark learner dictionary senses against WordNet synsets.
- Terms: Oxford 3000™ and Oxford 5000™ are trademarks of Oxford University Press (referenced for academic comparison).
- Resource: Oxford Learner's Dictionaries Wordlists
-
CEFR Illustrative Descriptors & Companion Volume (Council of Europe)
- Role: International reference standard defining the 6-level proficiency scale (A1–C2), illustrative descriptors, and communicative proficiency criteria.
- Terms: Official International Standard (Council of Europe, referenced for educational alignment)
- Resource: Council of Europe CEFR Portal
-
Mihalcea WSD Corpora & Resources (Rada Mihalcea)
- Role: Word Sense Disambiguation evaluation benchmarks and methodology used to calibrate the contextual window overlap scorer.
- Terms: Academic & Research Use
- Resource: University of Michigan WSD Downloads
-
CEFR-Annotated WordNet:
Kikuchi, M., Ono, M., Soga, T., Tanabe, T., & Ozono, T. (2024). CEFR-Annotated WordNet: LLM-Based Proficiency-Guided Semantic Database for Language Learning. arXiv preprint arXiv:2510.18466. https://doi.org/10.48550/arXiv.2510.18466 -
Princeton WordNet:
Miller, G. A. (1995). WordNet: A lexical database for English. Communications of the ACM, 38(11), 39–41. https://doi.org/10.1145/219717.219748 -
CEFR-J Framework:
Negishi, M., Takada, T., & Tono, Y. (2013). A progress report on the development of the CEFR-J. In N. Saville & F. Sheehan (Eds.), Critical issues in language testing: Studies in language testing (Vol. 36, pp. 135–163). Cambridge University Press. http://www.cefr-j.org/ -
CEFR Companion Volume:
Council of Europe. (2001, 2020). Common European Framework of Reference for Languages: Learning, teaching, assessment – Companion volume. Council of Europe Publishing. https://www.coe.int/en/web/common-european-framework-reference-languages -
Simplified Lesk Algorithm (WSD):
Lesk, M. (1986). Automatic sense disambiguation using machine-readable dictionaries: How to tell a pine cone from an ice cream cone. In Proceedings of the 5th Annual International Conference on Systems Documentation (SIGDOC '86) (pp. 24–26). Association for Computing Machinery. https://doi.org/10.1145/318723.318728 -
Wikipedia-Based Sense Disambiguation:
Mihalcea, R. (2007). Using Wikipedia for automatic word sense disambiguation. In Human Language Technologies 2007: NAACL-HLT Companion Volume (pp. 197–200). Association for Computing Machinery. https://web.eecs.umich.edu/~mihalcea/downloads.html
This project is licensed under the MIT License.