Maximal unit segmentation in spoken data

[sylvainkahane]

In spoken productions the boundary of sentences is not explicitly marked and linguists must decide where to cut (and where to eventually add punctuations in their transcriptions).
As part of a working group dedicated to spoken data of the Unidive COST action, we have tried to unify our criteria and we have started a fork of the UD guidelines where we propose specific guidelines for spoken data. If you are interested by the topic, please have a look to
https://grew.fr/spoken-language-guidelines/workgroups/spoken-data/maximal_unit_segmentation.html
When we will be sure we have reached a sufficient consensus, these guidelines will be merged in the UD website.

[dan-zeman]

This will be very useful, thanks!

I noticed one divergence from the UD guidelines (parataxis, ccomp) in your Reported Speech section. In the current UD guidelines, the speech verb is considered inserted (and thus attached as parataxis) only if it is surrounded by parts of the direct speech. When it occurs before or after the direct speech, it is the head and the direct speech is attached to it as ccomp. In contrast, you seem to treat the occurrence after the direct speech as an "insertion", too. Besides deviating from the guidelines, I suppose it would look problematic in verb-final languages such as Turkic or Dravidian.

[sylvainkahane]

The criteria to decide whether the speech verb is the head or it is inserted are language-specific. Indeed in verb-final languages, a verb final can be the head. I add an example from Beja.
But considering that the speech verb is the head as soon as it occurs before or after the direct speech is not reasonable. In French, insertions are easy to spot because the subject must be inverted. They can occurs inside the sentence but most of time they are at the end: "ne fais pas ça, a-t-elle dit" 'don't do that, she said'.
Even in English for "don't do that, she said", it is quite clear that it is an insertion, according to the prosody and English WO rules.
For insertion, we also keep parataxis, but we suggest to type it with parataxis:insert, in order to distinguish insertion and parentheses that have different properties. For verbal discourse markers (I mean), we use discourse instead of parataxis.

[amir-zeldes]

@sylvainkahane Thanks from me as well, this is a useful document and I like the rectional unit approach, as well as the option to have split-by-speaker and 'collaborative syntax across speakers' versions.

The issue that @dan-zeman mentioned is indeed in contradiction to the UD guidelines, and I think trying to decide on a case by base basis whether an unusual sentence final (or initial) speech verb is parataxis will lead to annotator agreement problems and cross-linguistic inconsistencies; these are a few of the reasons why the guidelines treat these as heads of ccomp.

Another important reason why I think the ccomp analysis is better is because it aligns more closely with valency/argument structure, which is a major goal of UD as a syntactic formalism. This is not to say that prosody is not interesting, I certainly think it is, but if we treat an extraposed "say" as parataxis, we make it appear like we have an intransitive speech verb, which is not really true. Treating the speech verb as the head even when it is postponed reflects the dependency grammar principle that argument structures can be realized in a variety of word orders. This is especially sensible for languages with relatively free word order, but even in languages like English or French, I would like to know where a speech verb can be realized relative to its reported speech.

For verbal discourse markers (I mean), we use discourse instead of parataxis.

This is the only other point I noticed that contradicts the current guidelines, at least for English. I think if you want to say that a verb has become a discourse marker, for example something like "Let's buy, say, 20 books", then you should tag that verb as INTJ and attach as discourse. But if you keep the transparent analysis where "I mean" is a VERB and its subject, then I think parataxis (in the parenthetical sense) is better, since it's really a full clause. These kinds of parentheticals can have relatively complex structure ("They drank, I don't know, 10 beers" has "know" with negation and an auxiliary!), and they are not a totally closed list (unlike lexical discourse markers in general), so analyzing them as "live", multi-part syntax, rather than lexicalized discourse markers seems safer and less prone to disagreements, while also ensuring that heads with a clausal structure get clausal deprels.

[ellepannitto]

We're not suggesting to tag "say" as INTJ, and also 'I mean', it stays a "PRON/nsubj - VERB/root" structure. But the way it is connected to the other clause it's not parataxis because it's not autonomous per se. Parataxis implies that the two parts are autonomous and coordinated at discourse level, it's not the case of these markers... Il Dom 5 Lug 2026, 17:55 Amir Zeldes ***@***.***> ha scritto:

…

*amir-zeldes* left a comment (UniversalDependencies/docs#1273) <#1273 (comment)> @sylvainkahane <https://github.com/sylvainkahane> Thanks from me as well, this is a useful document and I like the rectional unit approach, as well as the option to have split-by-speaker and 'collaborative syntax across speakers' versions. The issue that @dan-zeman <https://github.com/dan-zeman> mentioned is indeed in contradiction to the UD guidelines, and I think trying to decide on a case by base basis whether an unusual sentence final (or initial) speech verb is parataxis will lead to annotator agreement problems and cross-linguistic inconsistencies; these are a few of the reasons why the guidelines treat these as heads of ccomp. Another important reason why I think the ccomp analysis is better is because it aligns more closely with valency/argument structure, which is a major goal of UD as a syntactic formalism. This is not to say that prosody is not interesting, I certainly think it is, but if we treat an extraposed "say" as parataxis, we make it appear like we have an intransitive speech verb, which is not really true. Treating the speech verb as the head even when it is postponed reflects the dependency grammar principle that argument structures can be realized in a variety of word orders. This is especially sensible for languages with relatively free word order, but even in languages like English or French, I would like to know where a speech verb can be realized relative to its reported speech. For verbal discourse markers (I mean), we use discourse instead of parataxis. This is the only other point I noticed that contradicts the current guidelines, at least for English. I think if you want to say that a verb has become a discourse marker, for example something like "Let's buy, say, 20 books", then you should tag that verb as INTJ and attach as discourse. But if you keep the transparent analysis where "I mean" is a VERB and its subject, then I think parataxis (in the parenthetical sense) is better, since it's really a full clause. These kinds of parentheticals can have relatively complex structure ("They drank, I don't know, 10 beers" has "know" with negation and an auxiliary!), and they are not a totally closed list (unlike lexical discourse markers in general), so analyzing them as "live", multi-part syntax, rather than lexicalized discourse markers seems safer and less prone to disagreements, while also ensuring that heads with a clausal structure get clausal deprels. — Reply to this email directly, view it on GitHub <#1273?email_source=notifications&email_token=ADEAM7UOC2N32HHU2L4RFRT5DJ275A5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBYGY3DIMBUGA32M4TFMFZW63VKON2WE43DOJUWEZLEUVSXMZLOOSWGM33PORSXEX3DNRUWG2Y#issuecomment-4886640407>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADEAM7WJ5TEYHKUOOIRNPBT5DJ275AVCNFSNUABEKJSXA33TNF2G64TZHMYTQNZZGU3DOMZ3JFZXG5LFHM2DQMBYGQ4DSMRSHCQXMAQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[amir-zeldes]

Parataxis implies that the two parts are autonomous and coordinated at discourse level

That's often true, but it's not strictly required, since parataxis is also used for parentheticals. The guidelines specifically mention "and other elements, such as a sentential parenthetical":

https://universaldependencies.org/u/dep/parataxis.html

The problem with treating these as discourse is that there is no sharp line between literal uses of "I mean", "you know", "I don't know" etc. and on the other end of the spectrum, "empty" uses where they don't really mean anything. It might be more feasible if it was a closed list, but these kinds of parentheticals are productive. From a phrasal head perspective, I also think that a medially inserted "you know" is syntactically similar to other more contentful parentheticals, such as "they deliberated (I use that term loosely) whether they should do it" - the two are similar in that they can be inserted freely, and the normal syntactic structure will resume right after the parenthetical is complete.

[ellepannitto]

What do you mean there's no difference? They are prosodically very different, often phonetically reduced, and most of all syntactically they do not have scope on the rest of the sentence. When "I mean" is a discourse marker, you cannot change it to "I believe", you cannot add an explicit "that" to introduce the ccomp and so on... If you consider the sentence "they deliberated (I use that term loosely || I mean) whether they should do it", "I use that term loosely" is an autonomous clause with an indexical that refers to discourse context, "I mean" on its own is just nothing without a complement. Il giorno dom 5 lug 2026 alle ore 19:13 Amir Zeldes < ***@***.***> ha scritto:

…

*amir-zeldes* left a comment (UniversalDependencies/docs#1273) <#1273 (comment)> Parataxis implies that the two parts are autonomous and coordinated at discourse level That's often true, but it's not strictly required, since parataxis is also used for parentheticals. The guidelines specifically mention "and other elements, such as a sentential parenthetical": https://universaldependencies.org/u/dep/parataxis.html The problem with treating these as discourse is that there is no sharp line between literal uses of "I mean", "you know", "I don't know" etc. and on the other end of the spectrum, "empty" uses where they don't really mean anything. It might be more feasible if it was a closed list, but these kinds of parentheticals are productive. From a phrasal head perspective, I also think that a medially inserted "you know" is syntactically similar to other more contentful parentheticals, such as "they deliberated (I use that term loosely) whether they should do it" - the two are similar in that they can be inserted freely, and the normal syntactic structure will resume right after the parenthetical is complete. — Reply to this email directly, view it on GitHub <#1273?email_source=notifications&email_token=ADEAM7VI2464M2FHNGPDBMT5DKEDDA5CNFSNUABFM5UWIORPF5TWS5BNNB2WEL2JONZXKZKDN5WW2ZLOOQXTIOBYGY4DMMJZHE42M4TFMFZW63VHMNXW23LFNZ2KKZLWMVXHJLDGN5XXIZLSL5RWY2LDNM#issuecomment-4886861999>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADEAM7UA3DXCLH4IQFNXU3D5DKEDDAVCNFSNUABEKJSXA33TNF2G64TZHMYTQNZZGU3DOMZ3JFZXG5LFHM2DQMBYGQ4DSMRSHCQXMAQ> . You are receiving this because you commented.Message ID: ***@***.***>