UD_Old_French-SRCMF is a conversion of (part of) the SRCMF corpus (Syntactic Reference Corpus of Medieval French srcmf.org).
UD_Old_French-SRCMF consists in 10 texts spanning from 9th to 13th century. It includes 17678 sentences and 170 741 tokens.
Sentences are annotated with the following metadata:
- sent_id : a unique id for each sentence in the treebank
- text : the sentence
- newdoc id : a unique id for each of the texts. This id can be split on underscores to get back :
- name of the text
- form : verse and/or prose
The following table lists the texts used in this treebank :
|ID||Author||Name of the text||Number of tokens|
|Strasbourg_842_prose||anonymous||Serments de Strasbourg||115|
|StLegier_1000_verse||anonymous||Vie de saint Léger||1,388|
|StAlexis_1050_verse||anonymous||Vie de saint Alexis||4,750|
|Roland_1100_verse||anonymous||Chanson de Roland||28,752|
|Lapidaire_mid12_prose||anonymous||Lapidaire en prose||4,708|
|QuatreLivresReis_late12_prose||anonymous||Quatre livres des reis||12,949|
|BeroulTristan_late12_verse||Beroul, Tristan||Tristan de Beroul||26,766|
|TroyesYvain_1180_verse||Chrestien de Troyes, Yvain||Yvain de Chretien de Troyes||41,256|
|Aucassin_early13_verse-prose||anonymous||Aucassin et Nicolet||9,838|
|Graal_1225_prose||anonymous||Queste del Saint Graal||40,219|
The corpus results from a conversion from the SRCMF corpus.
In the SRCMF projet, texts with less than about 40,000 words were entirely annotated, while texts with more than 40,000 words were sampled in three parts (beginning, middle and end of the text) to reach a total amount of about 40000 words.
As a result, UD_Old_French-SRCMF includes 8 full texts (Strasbourg, StLegier, StAlexis, Roland, Lapidaire, BeroulTristan, TroyesYvain, Aucassin) and 2 sampled ones (QuatreLivresReis and Graal).
The treebank is split as follows (in number of tokens) :
It is to be noticed that most of Dev and Test data is taken from texts that also exist in Training data, but these are large texts, so it does not make the testing too easy… One text (Lapidaire) is represented only in Dev/Test but not in train. Pre-1100 texts are only in Train, because they seem too small to reserve anything for testing. The rest of the data (136 021 /170 741 tokens) is for Train corpus.
Deviations from UD
We added some more specific relations (subtypes), either to specify a relation, or in the case of tokens entering a double dependency relation (typically : relative pronouns and contracted forms) :
acl:relcl: relative clause
obl(eg. sin = si + en)
aux:pass: passive auxiliary
det(eg. del = de + le)
cc:nc: non coordinating conjunction (eg. et at the beginning of a sentence)
advmod(eg. coment at the beginning of a subordinate clause)
obj(eg. relative pronoun que)
obl(eg. relative pronoun cui / qui )
advmod(eg. jon = jo + en)
obj(eg. quil = qui + le)
obj(eg. sis = si + les)
obj(eg. nes = ne + les)
obj(eg. oul = ou + le)
obl:advmod: the double labelling accounts for the difficulty to decide between obl and advmod relations (
We added some features:
Morph=VFin: finite verb
Morph=VInf: non finite verb
Morph=VPar: verbal participle
Consult the language specific documentation for further details.
UD_Old_French-SRCMF results from the conversion of (part of) the SRCMF corpus (Syntactic Reference Corpus of Medieval French srcmf.org).
The SRCMF corpus results from the SRCMF project which took place in 2008-2012, funded by the ANR (France) and the DFG (Germany), and supervised by Sophie Prévost and Achim Stein.
The SRCMF project consisted in the manual syntactic annotation of 15 texts (251,000 tokens) from the 9th to 13th C. Part-of-speech tags were for most of them retrieved from the already existing tagging of the texts (stemming from: Base de Français Medieval, Lyon, ENS de Lyon, IHRIM Laboratory http://txm.bfm-corpus.org, and the Nouveau Corpus d'Amsterdam http://www.uni-stuttgart.de/lingrom/stein/corpus#nca)
The contributors to the SRCMF project were: Stein, Achim; Prévost, Sophie; Rainsford, Tom; Mazziotta, Nicolas; Bischoff Béatrice; Glikman, Julie; Lavrentiev, Alexei; Heiden, Serge; Guillot-Barbance, Céline; Marchello-Nizia, Christiane.
The whole SRCMF corpus (251,000 tokens) was converted into UD dependencies, but only 172,000 tokens have so far undergone a significant checking: the remaining 80,000 tokens will be added to UD_Old_French-SRCMF in a future release.
The conversion from the original SRCMF annotation to the SRCMF-UD annotation was done automatically both for the POS and the syntactic relations, thanks to a set of elaborated rules. Some 1,200 syntactic relations left unlabelled were then manually annotated (Sophie Prévost), and significant spot-checking occurred, focusing on potential difficulties (eg. conj relation).
This conversion was achieved by Aurélie Collomb, in the frame of a internship funded by lab Lattice (Paris, CNRS, ENS & Université Sorbonne Nouvelle Paris 3, PSL & USPC), and supervised by Sophie Prévost, Isabelle Tellier and Kim Gerdes. Marine Courtin achieved the deposit of the files, and especially took in charge the validation of the corpus through the successive steps of the process.
A significant review of this initial release has been done on the occasion of the UD 2.6 release by Loïc Grobol and Sophie Prévost in the frame of the ANR PROFITEROLE project in order to improve the compliance of the corpus to UD guidelines. This includes both automatic corrections using the graph rewriting system GREW (Bonfante et al., 2018) and extensive manual corrections.
- Stein, Achim, and Sophie Prévost. 2013. ‘Syntactic Annotation of Medieval Texts: The Syntactic Reference Corpus of Medieval French (SRCMF)’. In New Methods in Historical Corpora, edited by Paul Bennett, Martin Durrell, Silke Scheible, and Richard J. Whitt, 275–82. Corpus Linguistics and International Perspectives on Language. Gunter Narr Verlag.
- Bonfante, Guillaume, Bruno Guillaume, and Guy Perrier 2018. Application of Graph Rewriting to Natural Language Processing. Vol. 1. ISTE Wiley.
- 2020-05-15 v2.6
- Significant bugfix release (more than 9k changed tokens)
- 2018-04-15 v2.2
- Initial release
=== Machine-readable metadata (DO NOT REMOVE!) ================================ Data available since: UD v2.2 License: CC BY-NC-SA 3.0 Includes text: yes Genre: nonfiction legal poetry Lemmas: not available UPOS: converted with corrections XPOS: manual native Features: automatic Relations: automatic with corrections Contributors: Prévost, Sophie; Collomb, Aurélie; Gerdes, Kim; Tellier, Isabelle; Courtin, Marine; Lavrentiev, Alexei; Guillot-Barbance, Céline; Grobol, Loïc Contributing: here source Contact: email@example.com ===============================================================================