---
title: "Lemma — definition and key papers — Elamur Research"
description: "A lemma is the citation form of a word: the single form, such as a dictionary headword, chosen to stand for every inflected form of one lexeme."
canonical: "https://research.elamur.ai/free/en/concepts/lemma/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/concepts/lemma/)

[home](https://md.research.elamur.ai/free/en/) / [concepts](https://md.research.elamur.ai/free/en/concepts/) / Lemma

# 📌 Lemma

Concept · 18 key papers · [all concepts](https://md.research.elamur.ai/free/en/concepts/)

Part of: [🗨️ Natural language processing](https://md.research.elamur.ai/free/en/concepts/natural-language-processing/)

A lemma is the citation form of a word: the single form, such as a dictionary headword, chosen to stand for every inflected form of one lexeme. In English, run is the lemma of runs, ran and running.

Also: citation form, dictionary form, canonical form, base form, headword (in dictionaries), lemmata (plural)

## What it is

Look up "ran" in a dictionary and you will not find it as an entry. You find "run", and "ran" is listed under it. That entry word is the lemma. It is a conventional label for a whole family of [word forms](https://md.research.elamur.ai/free/en/concepts/word-form/) that differ only in grammar (tense, number, case, person) and not in basic meaning. Finding the lemma of each word in a text is called [lemmatization](https://md.research.elamur.ai/free/en/concepts/lemmatization/). It is one of the oldest and most common steps in [natural language processing](https://md.research.elamur.ai/free/en/concepts/natural-language-processing/) and [corpus linguistics](https://md.research.elamur.ai/free/en/concepts/corpus-linguistics/).

The lemma is one of several related notions that are easy to confuse:

| Term | What it is | Example |
| --- | --- | --- |
| Word form | A concrete inflected shape as it occurs in text | ran, running |
| Lexeme | The abstract word: a unit of meaning plus its full paradigm | RUN (the verb) |
| Lemma | The one form chosen to name the lexeme | run |
| Stem | The part that affixes attach to, which may not be a word on its own | run-, or Latin amā- in amāre |

A stem is a unit of [morphology](https://md.research.elamur.ai/free/en/concepts/morphology/), and [stemming](https://md.research.elamur.ai/free/en/concepts/stemming/) often produces strings that no dictionary lists. A lemma, by contrast, is usually a real word form. The main exception is root-based dictionaries of Semitic languages, described below.

The word also has a separate meaning in mathematics, where a lemma is an auxiliary proven statement used to prove a larger result. Current work studies whether language models check a lemma's hypotheses before invoking it (Xu et al. 2026). In interactive theorem proving, "hammer" tools pick relevant lemmas from large formal libraries and pass them to automated provers (Kaliszyk & Urban 2012; Blanchette et al. 2016). The rest of this article deals with the linguistic sense.

## The key idea

A lemma names one lexeme, so all the variation in form that the lexeme shows is grouped under a single label. A lexicon, a frequency list or a search index can then count, store and retrieve the lexeme once, rather than once for each form. Three consequences follow.

First, lemmatization is a many-to-one mapping, and it often depends on context. A Latvian form like "ceļu" can belong to a verb or to either of two nouns, and only the sentence decides which (Bergmanis & Goldwater 2018). In languages written without short vowels, such as Arabic, the same written word can match several lexemes. Picking the right one becomes a disambiguation task of its own ([Roth et al. 2008](https://md.research.elamur.ai/free/en/papers/roth-2008-arabic-morphological-tagging-diacritization-lemmatization/)).

Second, the gap between form and lemma varies greatly across languages. Morfette was evaluated on Polish data from the Korpus Słownika Frekwencyjnego, with about 219,000 training tokens. In that data, only 44.44% of tokens had a lemma identical to their surface form ([Chrupala et al. 2008](https://md.research.elamur.ai/free/en/papers/chrupala-2008-learning-morphology-morfette/)).

Third, the lemma is the natural unit for counting words. [Zipf's law](https://md.research.elamur.ai/free/en/concepts/zipfs-law/) is usually stated over word-form frequencies, with types ranked by how often they occur. It applies to lemma counts as well:

```
f(r) ≈ C / r^α
```

Here f(r) is the frequency of the item at rank r in a list sorted by frequency. C is a constant that depends on corpus size, and α is an exponent close to 1. Whether you count word forms or lemmas changes the shape of the long tail. It also changes measures such as the [type–token ratio](https://md.research.elamur.ai/free/en/concepts/type-token-ratio/). The difference is largest in highly inflected languages, where a single frequent lemma can be spread over dozens of rare forms.

## History and founding papers

### Dictionaries and headwords

The lemma began in [lexicography](https://md.research.elamur.ai/free/en/concepts/lexicography/). Each language tradition fixed a citation form: the infinitive for English verbs, the first-person singular present for Latin and Greek verbs, the nominative singular for nouns. For Semitic languages, the tradition is sometimes an abstract root rather than a word.

Dictionary definitions are written for headwords, and this pairing has been reused as a learning signal. Neural encoders have been trained to map a definition onto the embedding of the headword it defines, which also gives a reverse dictionary ([Hill et al. 2016](https://md.research.elamur.ai/free/en/papers/hill-2016-learning-understand-phrases-embedding-dictionary/)).

Lexical semantic studies still take the lexeme as the unit of analysis. Examples include Ziegert's frame-semantic study of Biblical Hebrew space-related verbs from the roots prd and bdl ([Anonymous 2024](https://md.research.elamur.ai/free/en/papers/ziegert-2024-framing-space-determining-meaning-space/)). Another example is a study of near-synonymous nouns for rain ([Pratesi 2024](https://md.research.elamur.ai/free/en/papers/pratesi-2024-syntagmatic-paradigmatic-analysis-lexemes-rain/)).

### Lexemes in computational morphology

Finite-state work treated the lexeme as the hidden object behind surface words. MAGEAD maps Arabic words, both standard and dialectal, to and from a lexeme plus a bundle of features ([Habash & Rambow 2006](https://md.research.elamur.ai/free/en/papers/habash-2006-magead-morphological-analyzer-generator-arabic/)).

Roth and colleagues extended the MADA disambiguation system with lexeme models and defined lexeme choice as a new task for Arabic. They tested it on the held-out half of ATB3-DevTest, about 25,000 words. Their tuned system with lexeme features chose the correct lemma 96.3% of the time, compared with 85.5% for an untuned baseline built on 14 morphological features ([Roth et al. 2008](https://md.research.elamur.ai/free/en/papers/roth-2008-arabic-morphological-tagging-diacritization-lemmatization/)).

### Data-driven lemmatizers

Three systems mark successive steps in statistical lemmatization. The latest of them was benchmarked against the other two (Bergmanis & Goldwater 2018).

| System | Core idea |
| --- | --- |
| Morfette | Separate maximum-entropy models for tags and for shortest edit scripts, combined by a beam-search decoder ([Chrupala et al. 2008](https://md.research.elamur.ai/free/en/papers/chrupala-2008-learning-morphology-morfette/)) |
| Lemming | Log-linear ranking of edit-tree candidates, coupled with a conditional random field (CRF) tagger in a joint model ([Müller et al. 2015](https://md.research.elamur.ai/free/en/papers/muller-2015-joint-lemmatization-morphological-tagging-lemming/)) |
| Lematus | Character-level encoder-decoder that reads surrounding characters as context (Bergmanis & Goldwater 2018) |

Lemming showed that predicting lemmas and morphological tags together helps both. It was compared with Morfette and JCK as baselines. For Czech, lemmatization error fell by about 60%, from 4.05% to 1.58% ([Müller et al. 2015](https://md.research.elamur.ai/free/en/papers/muller-2015-joint-lemmatization-morphological-tagging-lemming/)).

Lematus showed that context matters most for ambiguous forms. It was tested on 20 Universal Dependencies treebanks with full training data. With 20 characters of context on each side, accuracy on ambiguous words rose from 88.8% to 92.4%, compared with the variant that sees no context. Unseen words gained far less, moving from 83.6% to 84.3%. Even without context, the model outperformed all three baselines on unseen words (Bergmanis & Goldwater 2018).

The reverse direction, generating a form from a lemma plus features, became a task in its own right. Hard monotonic attention models did well on small training sets ([Aharoni & Goldberg 2017](https://md.research.elamur.ai/free/en/papers/aharoni-2017-morphological-inflection-generation-hard-monotonic/)). Data hallucination and transfer from related languages raised macro-averaged inflection accuracy by 15 percentage points over the previous state of the art for [low-resource languages](https://md.research.elamur.ai/free/en/topics/low-resource-languages/). That system ranked first by accuracy in SIGMORPHON 2019 ([Anastasopoulos & Neubig 2019](https://md.research.elamur.ai/free/en/papers/anastasopoulos-2019-pushing-limits-low-resource-morphological/)).

## How it is used today

### Lemmas in treebanks

In [Universal Dependencies](https://md.research.elamur.ai/free/en/concepts/universal-dependencies/) (UD), every token carries a LEMMA field next to its FORM. These fields sit in CoNLL-U files, UD's tab-separated file format. Tools are therefore trained and scored on lemmas together with [part-of-speech tagging](https://md.research.elamur.ai/free/en/concepts/part-of-speech-tagging/) and [dependency parsing](https://md.research.elamur.ai/free/en/concepts/dependency-parsing/).

UDPipe learns tokenization, tagging, lemmatization and parsing from nothing more than CoNLL-U files. Some treebanks, including Persian and Latin-PROIEL, required UDPipe to store whole ("absolute") lemmas rather than edit rules ([Straka & Straková 2017](https://md.research.elamur.ai/free/en/papers/straka-2017-tokenizing-pos-tagging-lemmatizing-parsing/)). The CoNLL 2018 shared task added BLEX, a parsing metric that also requires correct lemmas ([Zeman et al. 2018](https://md.research.elamur.ai/free/en/papers/zeman-2018-conll-2018-shared-task-multilingual/)).

Conventions differ between languages and between treebanks. Projects must decide how to cite verbs, whether to lemmatize pronouns or compounds, and how to handle spelling variants. Converting older annotations of Alsatian and Occitan to UD required re-deciding lemma and tag choices (Miletić et al. 2019).

UniMorph stores paradigms as lemma–features–form triples. Release 3.0 held 14.8 million triples across 118 languages. An audit before that release found Latin lemmas among the worst error sources ([McCarthy et al. 2020](https://md.research.elamur.ai/free/en/papers/mccarthy-2020-unimorph-3-0-universal-morphology/)).

### Lemmas as keys to meaning

In [word sense disambiguation](https://md.research.elamur.ai/free/en/concepts/word-sense-disambiguation/) (WSD), the lemma is what connects a word in running text to a dictionary entry. Babelfy relies on the fact that in WSD the lemma matches a dictionary sense exactly, while names in entity linking may appear only partially ([Moro et al. 2014](https://md.research.elamur.ai/free/en/papers/moro-2014-entity-linking-meets-word-sense/)). GlossBERT puts the target's lemma in front of each candidate gloss as a weak signal ([Huang et al. 2019](https://md.research.elamur.ai/free/en/papers/huang-2019-glossbert-bert-word-sense-disambiguation/)).

Distributional models traditionally assign one vector per lemma. That vector averages over a word's uses, which is why context-sensitive models were developed ([Erk & Padó 2008](https://md.research.elamur.ai/free/en/papers/erk-2008-structured-vector-space-model-word/)). AutoExtend breaks a pretrained word vector down into vectors for its lexemes and WordNet synsets (Rothe & Schütze 2015). Selectional-preference models judge a candidate headword by its similarity to headwords already seen in a predicate's argument slot (Erk et al. 2010).

Lexicalized parsing uses a different notion of "head word". There, probabilities are conditioned on the syntactic head of each phrase, not on a dictionary headword ([Collins 2003](https://md.research.elamur.ai/free/en/papers/collins-2003-head-driven-statistical-models-natural/)).

Simple lemma identity can be a surprisingly strong baseline. On the EventCorefBank (ECB) event coreference benchmark, a baseline that linked mentions only when their lemmas matched reached 52.52 CoNLL F1 across topics. CoNLL F1 is the mean of the MUC, B³ and CEAF-e F1 scores. Within topics, the same baseline reached 64.76. This exposed how little lexical variety the benchmark contained ([Cybulska & Vossen 2014](https://md.research.elamur.ai/free/en/papers/cybulska-2014-sledgehammer-crack-nut-lexical-diversity/)).

### Lemma banks and linked data

Lemmas also serve as identifiers in the [knowledge graph](https://md.research.elamur.ai/free/en/concepts/knowledge-graph/) sense. By 2020, Wikidata held lexeme entries (Nielsen 2020), including a Danish collection (Nielsen 2019). For [Ancient Greek](https://md.research.elamur.ai/free/en/topics/ancient-greek/), a lemma bank links corpora and lexical resources through shared lemma identifiers (Swaelens et al. 2026).

### Robustness

Because each lemma has a known set of inflections, those inflections can be used to test models. Swapping words for other inflected forms of the same lemma seriously damaged question-answering and English–French translation systems. Models trained on SQuAD 2.0 often switched to giving no answer. One epoch of fine-tuning on such variants made the systems far more robust. For the translation model, degradation under attack fell from 56.25% to 20.20% ([Tan et al. 2020](https://md.research.elamur.ai/free/en/papers/tan-2020-it-s-morphin-time-combating/)).

## Open questions

- Which form to cite. Choices between infinitive and other verb forms, root and stem, or modern and historical spelling are conventions. Comparing lemmatizers or lemma counts across treebanks therefore partly compares annotation policies (Miletić et al. 2019; [McCarthy et al. 2020](https://md.research.elamur.ai/free/en/papers/mccarthy-2020-unimorph-3-0-universal-morphology/)).
- Where lemma ends and sense begins. Homographs, such as the noun and verb readings in Latvian, make lemmatization a sense decision (Bergmanis & Goldwater 2018). Splitting near-synonyms raises the same issue from the lexicographic side ([Pratesi 2024](https://md.research.elamur.ai/free/en/papers/pratesi-2024-syntagmatic-paradigmatic-analysis-lexemes-rain/)).
- Whether lemmas are still needed. [Subword tokenization](https://md.research.elamur.ai/free/en/concepts/subword-tokenization/) lets neural models work without explicit lemmas. Yet inflectional perturbations still break them ([Tan et al. 2020](https://md.research.elamur.ai/free/en/papers/tan-2020-it-s-morphin-time-combating/)), and treebanks and linked lexicons keep lemmas as their main keys ([Zeman et al. 2018](https://md.research.elamur.ai/free/en/papers/zeman-2018-conll-2018-shared-task-multilingual/); Swaelens et al. 2026).
- Low-resource and historical languages. Unseen forms remain the main source of errors (Bergmanis & Goldwater 2018; [Chrupala et al. 2008](https://md.research.elamur.ai/free/en/papers/chrupala-2008-learning-morphology-morfette/)). Transfer and hallucination help, and hallucination alone comes close enough to be an option for language isolates. How far these gains carry over to poorly documented varieties is unclear ([Anastasopoulos & Neubig 2019](https://md.research.elamur.ai/free/en/papers/anastasopoulos-2019-pushing-limits-low-resource-morphological/)).
- How the counting unit shapes frequency laws. How does lemmatization change Zipfian fits and vocabulary-growth curves? And how does it affect their comparison across languages with different morphology?

## Key papers

The works this page cites, the most cited first; citations as counted by OpenAlex plus the papers of this library that cite the work, read 2026-10-24.

- [Head-Driven Statistical Models for Natural Language Parsing](https://md.research.elamur.ai/free/en/papers/collins-2003-head-driven-statistical-models-natural/) — Michael Collins · 2003 · 2,001 citations
- [Entity Linking meets Word Sense Disambiguation: a Unified Approach](https://md.research.elamur.ai/free/en/papers/moro-2014-entity-linking-meets-word-sense/) — Andrea Moro, Alessandro Raganato, Roberto Navigli · 2014 · 931 citations
- [Tokenizing, POS Tagging, Lemmatizing and Parsing UD 2.0 with UDPipe](https://md.research.elamur.ai/free/en/papers/straka-2017-tokenizing-pos-tagging-lemmatizing-parsing/) — Milan Straka, Jana Straková · 2017 · 811 citations
- [A Structured Vector Space Model for Word Meaning in Context](https://md.research.elamur.ai/free/en/papers/erk-2008-structured-vector-space-model-word/) — Katrin Erk, Sebastian Padó · 2008 · 508 citations
- [GlossBERT: BERT for Word Sense Disambiguation with Gloss Knowledge](https://md.research.elamur.ai/free/en/papers/huang-2019-glossbert-bert-word-sense-disambiguation/) — Luyao Huang, Chi Sun, Xipeng Qiu et al. · 2019 · 365 citations
- [CoNLL 2018 Shared Task: Multilingual Parsing from Raw Text to Universal Dependencies](https://md.research.elamur.ai/free/en/papers/zeman-2018-conll-2018-shared-task-multilingual/) — Daniel Zeman, Jan Hajič, Martin Popel et al. · 2018 · 316 citations
- [Using a sledgehammer to crack a nut? Lexical diversity and event coreference resolution](https://md.research.elamur.ai/free/en/papers/cybulska-2014-sledgehammer-crack-nut-lexical-diversity/) — Agata Cybulska, Piek Vossen · 2014 · 246 citations
- [Learning to Understand Phrases by Embedding the Dictionary](https://md.research.elamur.ai/free/en/papers/hill-2016-learning-understand-phrases-embedding-dictionary/) — Felix Hill, Kyunghyun Cho, Anna Korhonen et al. · 2016 · 234 citations
- [Arabic Morphological Tagging, Diacritization, and Lemmatization Using Lexeme Models and Feature Ranking](https://md.research.elamur.ai/free/en/papers/roth-2008-arabic-morphological-tagging-diacritization-lemmatization/) — Ryan Roth, Owen Rambow, Nizar Habash et al. · 2008 · 217 citations
- [Morphological Inflection Generation with Hard Monotonic Attention](https://md.research.elamur.ai/free/en/papers/aharoni-2017-morphological-inflection-generation-hard-monotonic/) — Roee Aharoni, Yoav Goldberg · 2017 · 209 citations
- [MAGEAD: A Morphological Analyzer and Generator for the Arabic Dialects](https://md.research.elamur.ai/free/en/papers/habash-2006-magead-morphological-analyzer-generator-arabic/) — Nizar Habash, Owen Rambow · 2006 · 209 citations
- [Learning Morphology with Morfette](https://md.research.elamur.ai/free/en/papers/chrupala-2008-learning-morphology-morfette/) — Grzegorz Chrupala, Georgiana Dinu, Josef van Genabith · 2008 · 189 citations
- [Joint Lemmatization and Morphological Tagging with Lemming](https://md.research.elamur.ai/free/en/papers/muller-2015-joint-lemmatization-morphological-tagging-lemming/) — Thomas Müller, Ryan Cotterell, Alexander Fraser et al. · 2015 · 180 citations
- [Pushing the Limits of Low-Resource Morphological Inflection](https://md.research.elamur.ai/free/en/papers/anastasopoulos-2019-pushing-limits-low-resource-morphological/) — Antonios Anastasopoulos, Graham Neubig · 2019 · 137 citations
- [It's Morphin' Time! Combating Linguistic Discrimination with Inflectional Perturbations](https://md.research.elamur.ai/free/en/papers/tan-2020-it-s-morphin-time-combating/) — Samson Tan, Shafiq Joty, Min-Yen Kan et al. · 2020 · 130 citations
- [UniMorph 3.0: Universal Morphology](https://md.research.elamur.ai/free/en/papers/mccarthy-2020-unimorph-3-0-universal-morphology/) — Arya D. McCarthy, Christo Kirov, Matteo Grella et al. · 2020 · 122 citations
- [Framing Space. Determining the Meaning of Space-Related Biblical Hebrew Lexemes](https://md.research.elamur.ai/free/en/papers/ziegert-2024-framing-space-determining-meaning-space/) — Carsten Ziegert · 2024
- [Syntagmatic and Paradigmatic Analysis of Lexemes for ‘Rain’ in Biblical Hebrew](https://md.research.elamur.ai/free/en/papers/pratesi-2024-syntagmatic-paradigmatic-analysis-lexemes-rain/) — Chiara Stornaiuolo Pratesi · 2024

## Related concepts

- [Lemmatization](https://md.research.elamur.ai/free/en/concepts/lemmatization/) — Lemmatization is the task of reducing an inflected or variant word form to its lemma, the canonical dictionary form under which it would be…
- [Word form](https://md.research.elamur.ai/free/en/concepts/word-form/) — A word form is one concrete written or spoken shape of a word as it appears in text, such as walks or walked. Each form realises a lexeme…
- [Type–token ratio](https://md.research.elamur.ai/free/en/concepts/type-token-ratio/) — Type–token ratio (TTR) is a measure of lexical diversity: the number of distinct word forms (types) in a text divided by the total number…
- [Lexicography](https://md.research.elamur.ai/free/en/concepts/lexicography/) — Lexicography is the discipline of compiling, describing and designing dictionaries and related lexical resources. It covers both the…

## More papers

Topics: [🧩 Morphology](https://md.research.elamur.ai/free/en/topics/morphology/) · [📖 Lexical semantics & word embeddings](https://md.research.elamur.ai/free/en/topics/lexical-semantics/)

Shelves: [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/)

Updated 24 October 2026

- [← previous concept Latent semantic indexing](https://md.research.elamur.ai/free/en/concepts/latent-semantic-indexing/)
- [next concept → Lemmatization](https://md.research.elamur.ai/free/en/concepts/lemmatization/)
- [↑ back to concepts](https://md.research.elamur.ai/free/en/concepts/)