This case study presents challenges and solutions for building calibrated
syntactically annotated diachronic corpora of French. Through a series of integrated
projects that developed methods for syntactic annotation of medium and large
collections of non-literary texts over a long period of time (13th – 19th centuries),
we consider how the availability of tools and the emergence of research questions
shape and direct corpus-building activity. Supported by the experience of an earlier
project, a semi-automated workflow for syntactic annotation was developed to allow
processing of two corpora calibrated by provenance and genre, with a focus on
parts-of-speech tagging in three annotation frameworks. The workflow combines the use
of a neural dependency parser, a lexicon of lemmata and forms and a Python library.
This experience, in turn, gave rise to a new initiative that aimed at formulating and
testing best practices for creating syntactically annotated
A case study exploring the building of a calibrated syntactically annotated diachronic corpora of French, focusing on collections of non-literary texts between the thirteenth and nineteenth centuries.
Journée d’études Retours d’expériences en édition numérique des textes, organised by Pierre-Jean Suriac, Angela Goebel and Morgane Pica on 22 June 2023 at Université Jean Moulin Lyon 3.
The study of the evolution of language over several centuries is facilitated by
access to digital collections of syntactically annotated texts. Such
- are usually limited to a particular period, which impedes comparative work
across the history of the language;
One notable exception is the Frantext corpus that, as of June 2026, contained 275 million words in French from the 10th to the 21st centuries from a range of genres. The corpus is lemmatized and annotated in parts of speech (https://www.frantext.fr/, accessed 1 June 2026). Among existing resources for the medieval period, we can cite the Base du Français Médiéval (URL: http://bfm.ens-lyon.fr/, accessed 8 June 2026 , MCVF-PPCHF and PhraseoRoche . For the 16th-17th centuries, SERMO corpus (URL: http://sermo.unine.ch/SERMO/, accessed 8 June 2026). - consist predominantly of literary material which exhibits conservative traits
and is believed to be removed from the everyday language
; - lack accurate annotation of syntactic functions, which means that even after
the extraction of desired configurations of parts of speech considerable
investment is often needed to review the data manually.
For Old and Middle French two important exceptions are two corpora of literary texts, MCFV-PPCHF, annotated according the UPenn syntactic framework, and Profiterole (formerly known as SRCMF) corpus annotated in Universal Dependencies (UD) .
In the present contribution, we discuss how a series of integrated projects conducted
in 2018-2024 at the CRISCO Laboratory at the University of Caen (France) set out to
address these challenges making use of language processing technologies available to
researchers at each stage.
Supporting research into the history of French syntax, the three projects
presented here sought to integrate existing tools and develop methodologies for
successful management and syntactic annotation of medium to large diachronic
The different stages of processing of the data, from image to annotated text, are described below. The major methodological and practical challenges include transcription and digitization of large amounts of textual data from heterogeneous original sources, lemmatization and syntactic annotation. Appropriate tools need to be selected to process the data at each stage. Different types of software use different data formats; therefore, conversion scripts need to be provided to ensure interoperability. In addition, choices need to be made about when linguists intervene to correct outputs of automatic processing.
All corpora were digitized from reproductions of manuscripts or scans of printed
editions using OCR and HTR technologies made available via the Transkribus portal.
The diversity of sources from across several centuries, ranging from medieval
manuscripts to modern critical editions, not to mention early printed books, meant
that the texts of the
Once digitized, the texts of
The chosen format allowed reflecting the structure of the original (i.e., the
manuscript or printed version that had been digitized) in the structure of the XML
document. According to the XML-TEI hierarchy, a text can be divided into books, a
book can be divided into sections, a section into chapters, and a chapter into
paragraphs; the corresponding tags made it possible to indicate where paragraphs
and larger sections of the text begin and end.
For 2_1_34_3_1
. This number
would, in turn, allow finding the place of the sentence in the XML-TEI structure
after parsing is completed.
Prior to any further treatment, the texts were tokenized and, in the case of
Considering the nature of the data and of the research needs as well as the availability of annotation tools, the sequence of projects followed two different yet complementary axes:
1) increased accuracy of the annotation and formalization of the correction process;
2) introduction of more layers of annotation.
Where
Following an extensive campaign of digitization, conversion to XML-TEI and
tokenization, the texts of the
While the use of the Presto lexicon and the script allowed matching the form of
the word with a possible lemma and PoS, it did not make use of contextual
information. It neither disambiguated when there was more than one lemma or PoS
speech possible, nor offer a solution when the form was not listed. In order to
address such cases, a programme of semi-manual correction was put in place to
reduce manual intervention.
Even though, given the size of the corpus (just under four and a half million
tokens), a comprehensive manual correction was never conducted, it proved to be
a reliable resource for tracing grammatical change
The challenges identified during work on
The workflow developed and tested by the two project teams relied on a
succession of automatic stages followed by manual checks (Figure 2).
Among existing dependency parsers, HoPS parser was selected because in 2021 the
developers of this tool had made available a model trained on a corpus of Old
and Middle French literary texts, and a large proportion of the data to be
annotated dated from the medieval period. Contextual analysis provided by HoPS
allowed successfully distinguishing between forms that otherwise would have
remained ambiguous. To take examples cited before, HoPS was very successful in
distinguishing in Normandy
) and pronouns (and of-it received the king hommage
, an example
from a 1373 text from I say that he proved this fact well
it was possible
to correctly tag in case of
grievance the adjournment is done in this manner
the same word was
tagged as a past participle of the verb
The absence of annotation of syntactic functions in
Modern neural network parsers that provide statistics-based predictions not only for
annotation in PoS (
At the center of our preoccupations when setting up and implementing the diachronic
Finally, the annotation process and the scripts developed as part of
Broader implications of the projects discussed here are found in highlighting
shortcomings in the current approaches to
- Lack of training
corpora for creating models for all periods of the history of the French language.The texts of
Chroniques andMICLE-Fr present considerable diachronic variation, ranging as they are from the 12th to the 19th century. As far as the choice of annotation model is concerned, in the UD collection, only the medieval and the contemporary period have referencecorpora for French language.For the comparative statistics of the Modern French UD treebanks, see https://universaldependencies.org/treebanks/fr-comparison.html (accessed 8 June 2026). Profiterole Old French and Profiterole Middle French corpora (https://universaldependencies.org/treebanks/fro_profiterole/index.html and https://universaldependencies.org/frm/index.html, accessed 8 June 2026) were formerly known as SRCMF. HoPS parser has a pre-trained model based on the previous version of the Old and Middle French corpus (https://github.com/hopsparser/hopsparser/blob/main/docs/models.md, accessed 8 June 2026). - Lack of consistency in the annotation principles for Medieval and Modern French
which impedes training of models and analysis. In the UD collection, Old/Middle
French and Modern French are treated as different languages with slightly
divergent annotation guidelines, e.g., where the annotation of modal verbs is
concerned. For example, for Old/Middle French, modal verbs (
devoir ,pouvoir ,souloir )In the Profiterole corpus, are treated as auxiliaries of the main lexical verb in the infinitive (e.g.,vouloir is no longer considered as a modal verb/auxiliary.Il peut travailler He can work
). For modern French, on the other hand, the modal verb is the head of the clause, and the infinitive of the lexical verb is the head of an open clausal complement (xcomp ). For reasons of consistency of annotation and given that the majority of our texts dates from before the 17th century, we used the pre-trained model for Old French supplied by the creators of the HoPS parserand we have thus annotated modal verbs as auxiliaries across the corpus but more discussion about homogenizing guidelines for different periods of French (and other Romance languages) is needed. https://zenodo.org/records/7708976 (accessed 8 June 2026).
In addition to HoPS, several graph-based syntactic parsers are currently available to
researchers, including UDify
Projects presented in this paper have received funding from Normandy Region (RIN:
Réseau d’Intérêts Normands): ConDÉ, High-Tech, and from the ANR: Agence Nationale de
Recherche (