Relation with LASLA Corpus ?

[PhVerkerk]

My initial concern was to find out how a tmesis is encoded in conllup. So I searched my favourite LASLA APN file for an occurence of "ante <quam>" (the usual way to code a form of "antequam" split into two words, I have to use here the < and > equivalents because these signs do not show up in the preview if I use their ASCII form). It occurs, for instance, in Caesar_BellumCivile_CaesBC1.APN in line 137 (from the repos https://dataverse.uliege.be/dataset.xhtml?persistentId=doi:10.58119/ULG/QJJ0SA) :
A03&0007# ante <quam> 1,2,2 0
The second half appearing in line 144 as :
A03&0007ANTEQVAM <ante> quam 1,2,2 T

Then I went to sentence 7 in Caesar_BellumCivile_CaesBC1.conllup. The relevant part being : "... ante de ea re ad senatum referri quam dilectus tota Italia habiti...". The corresponding words "ante" and "quam" have the analyses :
15 ante ante ADV M Degree=Pos _ _ _ CitationHierarchy=Liber_1,Capitulum_2,Paragraphus_2 i CaesBC1-A-03-7 140 lilaLemma:89165 TokenURI=http://lila-erc.eu/data/corpora/Lasla/id/corpus/CaesarBellum%20Civile/Caesar_BellumCivile_CaesBC1.BPN_t_0000140
...
22 quam quam SCONJ T _ _ _ _ CitationHierarchy=Liber_1,Capitulum_2,Paragraphus_2 i CaesBC1-A-03-7 147 lilaLemma:90912 TokenURI=http://lila-erc.eu/data/corpora/Lasla/id/corpus/CaesarBellum%20Civile/Caesar_BellumCivile_CaesBC1.BPN_t_0000147

Surprizingly, the positions in the LASLA file are 140 and 147, instead of 137 and 144. This is due to a second type of errors that will be described later. In these two lines, we see that the third field (LEMMA) is now "ante" and "quam". However, the lilaLemma:90912 corresponds to "antequam", but the spurious lemma "ante" = lilaLemma:89165 is an insult to the philologist that had annotated this text at LASLA.

The shift in items numbering is due to the erroneus lemmatisation of one "re publica" and two "rei publicae" that appear in the file before the "antequam" I was looking for. As a matter of fact, "rei publicae" is a single item in APN/BPN :
A03&0004RESPVBLICA rei publicae 1,1,2 A551
A03=0004QVE que 1,1,2 S
But it becomes two items in the conllup :
5 rei res NOUN A5 Case=Dat|Gender=Fem|InflClass=IndEurE|Number=Sing _ _ _ CitationHierarchy=Liber_1,Capitulum_1,Paragraphus_2 n5 CaesBC1-A-03-4 38 lilaLemma:121868 TokenURI=http://lila-erc.eu/data/corpora/Lasla/id/corpus/CaesarBellum%20Civile/Caesar_BellumCivile_CaesBC1.BPN_t_0000038
6-7 publicaeque _ _ _ _ _ _ _ _ _ _ _
6 publicae publicus ADJ C1 Case=Dat|Degree=Pos|Gender=Fem|InflClass=IndEurA|Number=Sing _ _ _ CitationHierarchy=Liber_1,Capitulum_1,Paragraphus_2 n6 CaesBC1-A-03-4 39 lilaLemma:120358 TokenURI=http://lila-erc.eu/data/corpora/Lasla/id/corpus/CaesarBellum%20Civile/Caesar_BellumCivile_CaesBC1.BPN_t_0000039
7 que que CCONJ S _ _ _ _ CitationHierarchy=Liber_1,Capitulum_1,Paragraphus_2 i CaesBC1-A-03-4 40 lilaLemma:131416 TokenURI=http://lila-erc.eu/data/corpora/Lasla/id/corpus/CaesarBellum%20Civile/Caesar_BellumCivile_CaesBC1.BPN_t_0000040
Here, no hint at all of the presence of the lemma "respublica" : just a noun "res" and an adjective "publicus".
As the v.2 of CoNLL-U allows space characters is the FORM field (https://universaldependencies.org/v2/conll-u.html), the proper encoding should probably have been :
5-6 rei publicaeque _ _ _
5 rei publicae respublica NOUN A5
6 que que CCONJ S
I have truncated the lines mainly because I don't know which "respublica" identifier I should choose : LiLa has two of them 76484 (first decl.) and 122962 (invariable !?!).

What a hell are these conllup files? The title and the filenames suggest that they are a conversion of the original BPN files, but they are not! A mere unchecked relemmatisation of a reconstructed text ?

Yours,
Philippe.

PS : "knowledge" is better with the "n" and "w" in the right order (twice in the readme file).

[FedeIure]

How LASLA and LiLa team jointly worked to the conversion of LASLA files for the linking of the texts to the LiLa Knowledge Base is described in two scientific publications:
 
Fantoli, Margherita, Passarotti, Marco, Mambrini, Francesco, Moretti, Giovanni and Ruffolo, Paolo. Linking the LASLA Corpus in the LiLa Knowledge Base of Interoperable Linguistic Resources for Latin. In Proceedings of the 8th Workshop on Linked Data in Linguistics within the 13th Language Resources and Evaluation Conference, Marseille, France. European Language Resources Association, 2022, pp. 26–34. https://aclanthology.org/2022.ldl-1.4/
Fantoli, Margherita, Passarotti, Marco and Longrée, Dominique. "Lemmas in Dialogue: Linking the L.A.S.L.A. Corpus to the LiLa Knowledge Base". Recent Trends and Findings in Latin Linguistics: Volume I: Syntax, Semantics and Pragmatics. Volume II: Semantics and Lexicography. Discourse and Dialogue, edited by Concepción Cabrillana, Berlin, Boston: De Gruyter, 2024, pp. 297-314. https://doi.org/10.1515/9783110722116-017 
 
This collaboration was the ground for the work that originated the production of the conllup files in this repository. Conllup format is a standard file format used in the Universal Dependencies framework, reference framework for cross-linguistic morphosyntactic annotation. No surprise over shift of positions from annotation choices in one format to standard annotation choices required from a standard format - what you call errors. In particular, changes are mostly due to requirements about tokenization and word segmentation within UD format: for a general overview on this, see https://universaldependencies.org/u/overview/tokenization.html 

Not all cases have been handled; as a consequence, the code used for the conversion is subject to changes. Therefore, it has not been released yet.

We encourage academic exchanges in a register appropriate to this kind of communication; any further communication in this spirit will be welcome.

[Stormur]

Since I wrote the conversion script I can comment more in detail.

There are no errors here, only unavoidable adjustments between two quite different annotation standards. The main point is that LASLA performs a sort of hybrid analysis which sometimes goes at the phrase level instead of sticking to (syntactic) words. This is not possible (and not even desired) in UD, and could belong to a different, more lexicological level, as tackled e.g. by the PARSEME project, which cannot be mixed with the morphosyntactic one (and is particularly untenable when two words are not even contiguous, like ante ... quam).

In a sense, also LASLA recognises the presence of multiple words when the annotation consists of two lines even in the case of the contiguous rei publicae. By the way, the indices 140 and 147 noted in the LILA:LINE field of the conllup file are exactly the line indices of the respective tokens in the original BPN file, so I don't understand what should be surprising, and to what the indices 137 and 144 should refer.

In UD the annotation takes place per synatctic word, and so one annotation has been given to each word in these expressions, including a distinct lemma, part of speech, and set of morphological features. There is a rather big (and, admittedly, not so elegant) "module" in the conversion script hardcoding these annotation re-integrations. There can be no talk of "a mere unchecked relemmatisation" as these annotations are not at all trivial, such as assigning all the correct adjectival features to publicae while rei publicae as a whole is tagged as a noun in LASLA. At the same time it is incomprehensible to me how ante can be defined a spurious lemma, as it is such a commonly attested and frequent term in Latin, and annotated also by LASLA, and the spelling antequam clearly represents a graphical univerbation of two elements. So these analyses are all perfectly valid, in line with UD and as much as possible with LASLA, too; if LASLA's annotation is preferred, then the original files are always available, and the insulted philologists can still take pride in them, while the linguists can work on a meaningful annotation.

Leaving aside any discussion about the soundness of some annotation choices in LASLA, one of the proposed solutions to treat rei publicaeque as a multiword token is not even possible at a formal level, since there can be no spaces inside a token in UD (unless a language has particular, and regular, spelling conventions, like Vietnamese, but it is surely not the case for Latin).

So yes, these are conversions of the original LASLA files, and of good quality, too. I know there might be some errors here and there, and the conversion script needs to be published with some documentation. I also could envision some annotation somewhere (MISC) noting the LASLA phrasal lemma antequam or respublica, but it would not impact on the overall required conversion adjustments.