Skip to content
Permalink
master
Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?

UD_Old_French-SRCMF/README.md

Go to file
 
 
Cannot retrieve contributors at this time
Summary Introduction Structure Deviations from UD Acknowledgments References Changelog
177 lines (140 sloc) 9.75 KB
Raw Blame
Open in GitHub Desktop

Summary

UD_Old_French-SRCMF is a conversion of (part of) the SRCMF corpus (Syntactic Reference Corpus of Medieval French srcmf.org).

Main development happens on the GitLab of the Profiterole Project.

Introduction

UD_Old_French-SRCMF consists in 10 texts spanning from 9th to 13th century. It includes 18 029 sentences and 199 699 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
    • date
    • form : verse and/or prose

The following table lists the texts used in this treebank :

ID Name of the text Author Tokens Trees
Strasbourg_842_prose Serments de Strasbourg anonymous 131 3
StLegier_1000_verse Vie de saint Léger anonymous 1665 189
StAlexis_1050_verse Vie de saint Alexis anonymous 5662 572
Roland_1100_verse Chanson de Roland anonymous 34803 3890
Lapidaire_mid12_prose Lapidaire en prose anonymous 5494 524
QuatreLivresReis_late12_prose Quatre livres des reis anonymous 15030 1509
BeroulTristan_late12_verse Tristan de Beroul Beroul 32596 3310
TroyesYvain_1180_verse Yvain de Chrestien Chrestien de Troyes 47964 3880
Aucassin_early13_verse-prose Aucassin et Nicolet anonymous 11639 1038
Graal_1225_prose Queste del Saint Graal anonymous 44715 3114
Total 199699 18029

Structure

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) :

ID Train Test Dev
Strasbourg_842_prose 131 0 0
StLegier_1000_verse 1665 0 0
StAlexis_1050_verse 5662 0 0
Roland_1100_verse 22593 6080 6130
Lapidaire_mid12_prose 0 2802 2692
QuatreLivresReis_late12_prose 15024 6 0
BeroulTristan_late12_verse 20404 6060 6132
TroyesYvain_1180_verse 47964 0 0
Aucassin_early13_verse-prose 11639 0 0
Graal_1225_prose 33538 5578 5599
Total 158620 20526 20553

Note 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 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
  • advmod:obl : contracted advmod + obl (e.g. sin = si + en)
  • aux:pass : passive auxiliary
  • case:det : contracted case + det (e.g. del = de + le)
  • cc:nc : non coordinating conjunction (e.g. et at the beginning of a sentence)
  • mark:advmod : mark and advmod (e.g. coment at the beginning of a subordinate clause)
  • nsubj:advmod : contracted nsubj + advmod (e.g. jon = jo + en)
  • nsubj:obj : contracted nsubj + obj (e.g. quil = qui + le)
  • obj:advmod : contracted advmod + obj (e.g. sis = si + les)
  • obj:advneg : contracted negation + obj (e.g. nes = ne + les)
  • obj:obl : contracted obl + obj (e.g. oul = ou + le)
  • obl:advmod : the double labelling accounts for the difficulty to decide between obl and advmod relations (en and i).

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.

Acknowledgments

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 (e.g. 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.

A significant import of data from the Base de français medieval has been done by Loïc Grobol, Alexei Lavrentiev and Sophie Prévost on the occasion of the UD 2.9 release. Most notably, this release adds punctuation tokens for most trees as well as around 350 new trees, consisting mostly of averbal sentences and fixes a number of conformity bugs with the UD guidelines. See the full changes on the upstream repository

References

Changelog

  • 2021-10-29 v2.0
    • Significant bugfix and data addition
    • Trees now have punctuation
    • See the upstream release for more details
    • Punctuation addition HAS a significant impact on automatic performances in most settings, please take this into account when comparing the results of parsers on this release with older releases
  • 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 ; Regnault, Mathilde
Contributing: elsewhere
Contact: sophie.prevost@ens.psl.eu
===============================================================================