🎓 Education

A survey · about 8 minutes to read · every paper named here has its own page in this library

The Education shelf, a subfield of the social sciences, gathers work where language technology meets teaching and learning: how people learning a language actually write and speak, how software can read, correct, score or support that output, and how new tools such as large language models might change classrooms. Most of the papers here sit at the junction of computational linguistics and second-language acquisition, and many of them also stand on the Language and Linguistics shelf.

The shelf is small and uneven. A 2023 position paper on large language models in education draws almost all of its citations; the rest is a steady stream of learner corpora, annotation schemes and tools for processing non-standard text, presented mostly at the specialist workshops of the field. Together they give a compact picture of what educational NLP looks like when it starts from data produced by learners themselves.

What it covers

The central object is learner language: essays, answers and recorded speech produced by people who are still acquiring a language, whether a second language or, at school age, their first written register. Such text is systematically noisy: misspellings, missing articles and constructions that native-trained tools never saw. The shelf's dominant topic is therefore grammatical error correction and learner language, alongside part-of-speech tagging and Universal Dependencies applied to learner text.

A second strand concerns educational applications: reading aids, short-answer scoring, graded reading material and dialogue tutors. Here the shelf overlaps with language models and pretraining, dialogue systems and text generation.

History and milestones

Learner corpora predate the papers here: the International Corpus of Learner English was assembled from the 1990s, and the Cambridge FCE essays became a standard source for error research. The earliest work on this shelf, from 2010, already shows the shift from training error detectors on clean native text to training them on learner essays with teacher corrections. The mid-2010s brought new resources for less-studied settings (German as a first language at school, Swedish as a second language) and the first syntactic treebank of learner English, which tied learner data to the cross-lingual Universal Dependencies framework.

From around 2020 the methods follow the wider field: recurrent networks, then BERT-style transformers, then large language models as simulated students or story writers. The newest papers return to corpus building with richer annotation.

Main problems and methods

Processing noisy text. Taggers and parsers trained on edited prose stumble on learner input. Work here measures where the damage comes from — unknown words caused by misspellings, capitalisation errors, a different distribution of word classes — and adapts models with character-level information and learner data. A recurring finding is that errors hurt automatic analysis less than feared, and that spelling correction before tagging buys little.

Annotating what learners wrote. A methodological question runs through the treebank and corpus papers: should the analysis describe the sentence as written or as the learner presumably intended? Annotating both an original and a corrected version is the common answer.

Measuring proficiency and development. Several papers turn learner data into scales: assigning vocabulary to CEFR levels, tracing how children's writing grows more complex across school grades, or measuring with contextual embeddings the gap between learner and native usage of a word.

Building tools for learners and teachers. Reading assistants that propose context-appropriate synonyms, analytic scoring of short science answers, curriculum-controlled story generation and tutoring dialogues all appear here. The newest ones rely on fine-tuning compact models rather than prompting large ones.

Landmark papers in this library

Large language models in education. ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education (2023) is a position paper by an interdisciplinary Munich–Tübingen group. It sketches uses for pupils, students, professionals and teachers, weighs risks such as bias, brittleness and unequal access, and argues for AI literacy and system-level strategies. It is by far the most cited work on the shelf.

Tutoring dialogues. MathDial: A Dialogue Tutoring Dataset with Rich Pedagogical Properties Grounded in Math Reasoning Problems (2023) had experienced teachers tutor an LLM that played a confused student on GSM8k problems. The resulting dialogues, labelled with teacher moves, let small fine-tuned models tutor more correctly than a prompted ChatGPT while giving away answers far less often.

A treebank of learner English. Universal Dependencies for Learner English (2016) built the Treebank of Learner English from FCE essays by writers with ten native languages, annotating each sentence in original and corrected form under a "literal reading" principle. Later work reused it, including Learning with learner corpora: Using the TLE for native language identification (2017), which found that a parser trained on the erroneous sentences gave slightly better native-language identification despite parsing worse.

Learning from corrected essays. Using an Error-Annotated Learner Corpus to Develop an ESL/EFL Error Correction System (2010) trained a maximum-entropy preposition corrector on tutor-corrected essays by Korean schoolchildren. Though tutors missed many errors, it beat native-trained models on precision.

First-language school writing. KoKo: an L1 Learner Corpus for German (2014) collected texts from upper-secondary students in three German-speaking regions, with orthographic error annotation and rich questionnaire metadata. Becoming Linguistically Mature: Modeling English and German Children's Writing Development Across School Grades (2020) predicted school grade from sequences of complexity measures computed along a text, beating models that see only text-level averages.

Vocabulary and proficiency. From distributions to labels: A lexical proficiency analysis using learner corpora (2016) derived a single CEFR level per word from Swedish learner essays, correcting frequencies for how many different learners used a word. Comparing Native and Learner Englishes Using a Large Pre-trained Language Model (2022) used BERT embeddings to show learners' word use drawing closer to native norms as proficiency rises, while warning that topic differences between corpora distort the picture.

Reading support. Using Word Semantics To Assist English as a Second Language Learners (2015) turned a lexical substitution system into a browser extension that highlights hard words and offers synonyms in context.

Tagging learner English. Analyzing the Impact of Spelling Errors on POS-Tagging and Chunking in Learner English (2017) showed that affix features recover most of what a spell checker would add. A POS Tagging Model Designed for Learner English (2018) followed with a BiLSTM tagger using character embeddings and pre-training on learner text, and reported that a native-language feature did not help.

Scoring and generation. Transformer-based Hebrew NLP models for Short Answer Scoring in Biology (2023) fine-tuned AlephBERT for rubric-based scoring of high-school biology answers, including across items. Children's English Reading Story Generation via Supervised Fine-Tuning of Compact LLMs with Controllable Difficulty and Safety (2026) fine-tuned 8B models to write phonics-aligned stories for early readers.

People

• Ryo Nagata and Tomoya Mizumoto — the most prolific pair here, with two linked studies on tagging learner English; Nagata is now associated with Konan University, Mizumoto with Waseda University.
• Yevgeni Berzak and Boris Katz — led the Treebank of Learner English; Berzak is now at the Technion, Katz at MIT.
• Debopam Das and Izabela Czerniak — at Åbo Akademi University, behind the ICLE-RC corpus of relative clauses, with Peter Bourgonje.
• Iryna Gurevych and Mrinmaya Sachan — senior authors of MathDial; Gurevych heads a large NLP group at TU Darmstadt, Sachan is linked with ETH Zurich.
• Daryna Dementieva — co-author of the position paper on LLMs in education, now at the Munich Center for Machine Learning, TU Munich.
• Joel Tetreault and Na-Rae Han — authors of the early learner-trained preposition corrector.
• Elena Volodina — University of Gothenburg, central to Swedish second-language resources.
• Elma Kerz and Daniel Wiechmann — modelled writing development through complexity contours.
• Rada Mihalcea and Chris Hokamp — built the context-aware reading assistant for ESL learners.
• Beata Beigman Klebanov — Educational Testing Service, co-author of the Hebrew short-answer scoring study.
• Egon Stemle — Eurac Research, part of the KoKo corpus team.

Institutions and places

Authors span Japan, North America and Europe. Japanese work concentrates on tagging learner English at Konan University and Waseda University; Nordic work on learner corpora at Åbo Akademi University. In the United States, Princeton University, the Educational Testing Service, the Massachusetts Institute of Technology, Georgetown University and the Florida Institute for Human and Machine Cognition appear; in Europe, ETH Zurich, Ludwig-Maximilians-Universität München, the Hessian Center for Artificial Intelligence, Masaryk University and the French Laboratoire de Linguistique Formelle.

Publication is scattered across the computer-assisted language learning workshops, the BEA workshops, LREC and a few main conferences.

Open questions

• Literal versus intended analysis. How should tools represent ungrammatical sentences so that both learners' errors and their intended meaning remain visible?
• Machine normalisation. Speech recognition can quietly correct learner speech; Potential of ASR for the study of L2 learner corpora (2024) found Whisper erasing the very forms acquisition researchers study.
• Evaluation of generated material. Story and tutor generators are often judged with the same automatic metrics they were tuned on, without human experts or held-out lessons.
• Languages beyond English. Resources for Romanian, Swedish, German, French and Hebrew exist here, but English still dominates, a gap the LLM position paper also flags.

Neighbouring topics

The shelf leans heavily on Language and Linguistics and touches Artificial Intelligence. Beyond its own topics, related reading includes psycholinguistics and language acquisition, treebanks, speech and spoken language, reasoning for the maths tutoring work, Hebrew & Aramaic for the scoring study, and evaluation and benchmarks.

Key works in this library

• ChatGPT for Good? On Opportunities and Challenges of Large Language Models for Education — framework for responsible LLM use in schools.
• MathDial: A Dialogue Tutoring Dataset with Rich Pedagogical Properties Grounded in Math Reasoning Problems — teacher–LLM tutoring dialogues.
• Universal Dependencies for Learner English — the first open learner-English treebank.
• Using an Error-Annotated Learner Corpus to Develop an ESL/EFL Error Correction System — preposition correction from tutor feedback.
• KoKo: an L1 Learner Corpus for German — school essays by native German speakers.
• Becoming Linguistically Mature: Modeling English and German Children's Writing Development Across School Grades — complexity contours across grades.
• From distributions to labels: A lexical proficiency analysis using learner corpora — CEFR levels for Swedish words.
• Using Word Semantics To Assist English as a Second Language Learners — in-browser synonym help.
• Analyzing the Impact of Spelling Errors on POS-Tagging and Chunking in Learner English — how much misspellings cost taggers.
• A POS Tagging Model Designed for Learner English — neural tagger adapted to learners.
• Transformer-based Hebrew NLP models for Short Answer Scoring in Biology — rubric scoring in Hebrew.
• Designing the LECOR Learner Corpus for Romanian — learners of Romanian, original and corrected.
• Comparing Native and Learner Englishes Using a Large Pre-trained Language Model — lexical depth via embeddings.
• Children's English Reading Story Generation via Supervised Fine-Tuning of Compact LLMs with Controllable Difficulty and Safety — phonics-controlled stories.
• Potential of ASR for the study of L2 learner corpora — Whisper on learner French.
• ICLE-RC: International Corpus of Learner English for Relative Clauses — multi-level relative-clause annotation.
• Syntax of referents of relative markers: Evidence from a corpus of learner English — antecedent roles by first language.
• Learning with learner corpora: Using the TLE for native language identification — learner treebank for identifying L1.
• Variabilité des syllabes réalisées par des apprenants de l'anglais (Analysing syllable variability in a French learner corpus of English) — syllable-based rhythm of French learners.

The library in figures