--- title: "reparandum: overridden disfluency β€” Elamur Research" description: "reparandum is the Universal Dependencies relation for a speech disfluency. Spoken language is full of restarts." canonical: "https://research.elamur.ai/free/en/concepts/ud-reparandum/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) Β· [this page in HTML](https://research.elamur.ai/free/en/concepts/ud-reparandum/) [home](https://md.research.elamur.ai/free/en/) / [concepts](https://md.research.elamur.ai/free/en/concepts/) / reparandum: overridden disfluency # 🌐 reparandum: overridden disfluency Universal Dependencies Β· 6 key papers Β· [all concepts](https://md.research.elamur.ai/free/en/concepts/) Part of: [πŸ—¨οΈ Natural language processing](https://md.research.elamur.ai/free/en/concepts/natural-language-processing/) Β· [🌐 Universal Dependencies](https://md.research.elamur.ai/free/en/concepts/universal-dependencies/) reparandum is the Universal Dependencies relation for a speech disfluency. It links a word that the speaker abandoned or corrected to the word that replaces it, so a false start stays in the tree without being read as part of the sentence's meaning. Also: overridden disfluency, speech repair, self-correction, reparandum relation ## What it is Spoken language is full of restarts. A speaker begins a word or phrase, breaks off, and says something else instead. In [Universal Dependencies](https://md.research.elamur.ai/free/en/concepts/universal-dependencies/), the abandoned material is called the reparandum. The term comes from the classic analysis of speech repairs (Levelt 1983; Shriberg 1994), which separates three parts: - the reparandum, which is the part that gets replaced; - an optional interregnum, such as "uh" or "I mean"; - the repair, which is the replacing material. UD is a lexicalist dependency scheme. Its relation inventory is small and closed: version 2 lists 37 universal types, and languages may refine them with subtypes but may not add new base categories ([Nivre et al. 2020](https://md.research.elamur.ai/free/en/papers/nivre-2020-universal-dependencies-v2-evergrowing-multilingual/)). The inventory goes back to the universal Stanford dependencies ([Marneffe et al. 2014](https://md.research.elamur.ai/free/en/papers/marneffe-2014-universal-stanford-dependencies-cross-linguistic/); [Nivre et al. 2016](https://md.research.elamur.ai/free/en/papers/nivre-2016-universal-dependencies-v1-multilingual-treebank/)). Within it, reparandum is the label for an overridden disfluency. The relation runs in one direction: - The head is the repair, the word that is kept. - The dependent is the head of the reparandum, the word that was overridden. Material inside the abandoned phrase keeps its ordinary internal relations. Only the phrase's head attaches to the repair. As a result, the tree of the corrected utterance stays intact, and a user who wants the "fluent" sentence can simply prune every reparandum subtree. ### Neighbouring relations The relation must be kept apart from several others: | Relation | What it covers | | --- | --- | | discourse | fillers and interregna such as "uh", "um", "euh", which are not replaced material | | conj | genuine alternatives or coordination, where both parts remain asserted | | parataxis | two clause-like units placed side by side, where neither cancels the other | | goeswith | one word split by a typing error, not a spoken restart | No subtypes of the form reparandum:subtype are part of the universal guidelines. Treebanks generally use the bare label. ## Examples Arrows point from the head to the dependent. 1. "Turn to the left, to the right." reparandum(right β†’ left) "to" and "the" in the first phrase stay attached to "left". 2. "I was, I went home." reparandum(went β†’ was) The first "I" remains the subject of "was". The verb that was cut off depends on the verb that replaced it. 3. French: "le le chat dort" ("the the cat sleeps") reparandum(leβ‚‚ β†’ le₁) This is a simple repetition. The second determiner is the one integrated into the noun phrase. The second example shows a verb replacing a tensed verb. The third shows determiner repetition, one of the most common patterns in transcribed speech. ## Across languages The relation occurs only where a treebank contains transcribed speech, so its use depends heavily on corpus design. The guidelines aim for constructions to be annotated in parallel across typologically different languages ([Marneffe et al. 2021](https://md.research.elamur.ai/free/en/papers/marneffe-2021-universal-dependencies/)), and disfluencies are treated with the same mechanism everywhere. What differs between languages is which word classes get repaired. In the [English](https://md.research.elamur.ai/free/en/topics/english/) treebanks, the dominant pairings are: - verb β†’ verb; - pronoun β†’ pronoun; - verb β†’ auxiliary, where the speaker restarts the verb group. In [French](https://md.research.elamur.ai/free/en/topics/french/), verb β†’ verb also leads, but determiner β†’ determiner repetitions are far more prominent. This reflects how often French speakers restart at the article. Noun β†’ preposition attachments are also frequent in French, typically for a prepositional phrase abandoned before its noun. In both languages, the relation covers a similarly small share of all tokens. That is expected for a label confined to [speech](https://md.research.elamur.ai/free/en/topics/speech/) data. Some treebanks use reparandum and others never do. Comparing annotation across corpora therefore requires the kind of consistency checking that has been applied to UD ([Marneffe et al. 2017](https://md.research.elamur.ai/free/en/papers/marneffe-2017-assessing-annotation-consistency-universal-dependencies/)). ## In treebanks and research For [treebanks](https://md.research.elamur.ai/free/en/topics/treebanks/), reparandum is part of the basic dependency tree, which every UD treebank must provide ([Nivre et al. 2020](https://md.research.elamur.ai/free/en/papers/nivre-2020-universal-dependencies-v2-evergrowing-multilingual/)). The enhanced graph layer adds edges to the basic tree but does not remove them ([Schuster & Manning 2016](https://md.research.elamur.ai/free/en/papers/schuster-2016-enhanced-english-universal-dependencies-improved/)). Disfluencies are therefore still present there, and stripping them is left to the user. Tutorials on UD principles present reparandum among the relations specific to spoken data (Zeman 2021). Recent work on how UD encodes constructions and strategies places such special relations within the scheme's functional design (Nivre 2025). In [dependency parsing](https://md.research.elamur.ai/free/en/concepts/dependency-parsing/), reparandum is a rare, low-frequency label. The multilingual CoNLL shared tasks of 2017 and 2018 used UD to parse raw text in many languages ([Marneffe et al. 2021](https://md.research.elamur.ai/free/en/papers/marneffe-2021-universal-dependencies/)). Parsers trained there mostly saw written data, so they meet speech repairs seldom and tend to confuse them with conj or parataxis. The label is nonetheless valuable for several lines of work: - [corpus linguistics](https://md.research.elamur.ai/free/en/concepts/corpus-linguistics/) of spoken language; - [psycholinguistics](https://md.research.elamur.ai/free/en/concepts/psycholinguistics/) research on self-monitoring; - disfluency detection, the classic task on conversational speech since the Switchboard corpus (Godfrey et al. 1992). In each case, the label lets researchers locate repairs while keeping a syntactically coherent analysis of what the speaker finally said. How often the relation reparandum occurs: its share of all tokens Counted by code over the Universal Dependencies 2.18 treebanks of English and French (every token once, subtypes with their base relation). **The numbers** | Treebanks | Share | | --- | --- | | English treebanks | 0.20% of tokens (2,261 tokens) | | French treebanks | 0.21% of tokens (2,242 tokens) | What reparandum joins: the most frequent head tag β†’ dependent tag The five most frequent in each language's treebanks, by tokens. **The numbers** | Pair | Tokens | | --- | --- | | English: VERB β†’ VERB | 296 | | English: PRON β†’ PRON | 279 | | English: VERB β†’ AUX | 175 | | English: NOUN β†’ NOUN | 118 | | English: DET β†’ DET | 116 | | French: VERB β†’ VERB | 285 | | French: DET β†’ DET | 255 | | French: PRON β†’ PRON | 227 | | French: NOUN β†’ NOUN | 186 | | French: NOUN β†’ ADP | 146 | ## 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. - [Universal Dependencies v1: A Multilingual Treebank Collection](https://md.research.elamur.ai/free/en/papers/nivre-2016-universal-dependencies-v1-multilingual-treebank/) β€” Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter et al. Β· 2016 Β· 1,741 citations - [Universal Stanford dependencies: A cross-linguistic typology](https://md.research.elamur.ai/free/en/papers/marneffe-2014-universal-stanford-dependencies-cross-linguistic/) β€” Marie-Catherine de Marneffe, Timothy Dozat, Natalia Silveira et al. Β· 2014 Β· 627 citations - [Universal Dependencies](https://md.research.elamur.ai/free/en/papers/marneffe-2021-universal-dependencies/) β€” Marie-Catherine de Marneffe, Christopher D. Manning, Joakim Nivre et al. Β· 2021 Β· 480 citations - [Enhanced English Universal Dependencies: An Improved Representation for Natural Language Understanding Tasks](https://md.research.elamur.ai/free/en/papers/schuster-2016-enhanced-english-universal-dependencies-improved/) β€” Sebastian Schuster, Christopher D. Manning Β· 2016 Β· 336 citations - [Universal Dependencies v2: An Evergrowing Multilingual Treebank Collection](https://md.research.elamur.ai/free/en/papers/nivre-2020-universal-dependencies-v2-evergrowing-multilingual/) β€” Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter et al. Β· 2020 Β· 325 citations - [Assessing the Annotation Consistency of the Universal Dependencies Corpora](https://md.research.elamur.ai/free/en/papers/marneffe-2017-assessing-annotation-consistency-universal-dependencies/) β€” Marie-Catherine de Marneffe, Matias Grioni, Jenna Kanerva et al. Β· 2017 Β· 30 citations ## Related concepts - [Universal Dependencies](https://md.research.elamur.ai/free/en/concepts/universal-dependencies/) β€” Universal Dependencies is a framework and an open collection of treebanks that annotates the grammar of many languages with one shared… - [Dependency parsing](https://md.research.elamur.ai/free/en/concepts/dependency-parsing/) β€” Dependency parsing is the task of recovering a sentence's syntactic structure as directed, labelled links between its words. Each word is… ## More papers Topics: [🌐 Universal Dependencies](https://md.research.elamur.ai/free/en/topics/universal-dependencies/) Β· [🌳 Dependency parsing](https://md.research.elamur.ai/free/en/topics/dependency-parsing/) Shelves: [πŸ—£οΈ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) - [← previous concept Question answering](https://md.research.elamur.ai/free/en/concepts/question-answering/) - [next concept β†’ Retrieval-augmented generation](https://md.research.elamur.ai/free/en/concepts/retrieval-augmented-generation/) - [↑ back to concepts](https://md.research.elamur.ai/free/en/concepts/)