Alignment of French Resources

This repository contains resources and codes to run experiments on aligning the Trésor de la Langue Française informatisé and the French Wiktionary -- Wiktionnaire.

Data Extraction

In addition to codes/main.py, two of the provided scripts can be used within other applications and are described as follows:

codes/dbnary.py retrieves data from Wiktionnaire through DBnary's SPARQL endpoint
codes/tlfi.py extracts relevant information from TLFi by reaching out to specific tags and paths in the XML files. To reduce complexity in the hierarchy of senses, all senses are flattened and provided at the same level. Extraction of definitions was particularly challenging due to lack of distinction between the definition of the entry versus that of the locution/idioms which also appear in the same entry. For example, mur (noun) comes with over 60 definitions in (TLFi)[https://www.cnrtl.fr/definition/mur] among which only a couple of them define the entry and the other, focus on locutions, idioms or multi-word expressions as in "mur des Lamentations, des Pleurs." and "mur ossaturé".

In both resources, the microstructure of each entry is reshaped as follows:

puissamment from TLFi (See [https://www.cnrtl.fr/definition/PUISSAMMENT])

{
    "gender": "",
    "id": "69866",
    "lemma": "puissamment",
    "pos": "adverb",
    "senses": {
        "A.": "Très, à un degré élevé, avec force.",
        "B.": "Avec efficacité, grandement.",
        "C.": "Avec de la force physique; de façon marquée.",
        "D.": "Avec des moyens puissants, fortement."
    }
}

puissamment from Wiktionnaire (See [https://fr.wiktionary.org/wiki/puissamment])

{
    "http://kaiko.getalp.org/dbnary/fra/puissamment__adv__1": {
        "lemma": "puissamment",
        "pos": "adverb",
        "gender": "",
        "senses": {
            "http://kaiko.getalp.org/dbnary/fra/__ws_2_puissamment__adv__1": "Avec force.",
            "http://kaiko.getalp.org/dbnary/fra/__ws_4_puissamment__adv__1": "(Par extension) très.",
            "http://kaiko.getalp.org/dbnary/fra/__ws_1_puissamment__adv__1": "Avec puissance.",
            "http://kaiko.getalp.org/dbnary/fra/__ws_3_puissamment__adv__1": "Extrêmement."
        }
    }
}

Data Annotation

Based on the lemmata used in the MWSA datasets (which are based on the English WordNet), we retrieve the synsets associated to the word in the French language based the Open Multilingual WordNet. The is beneficial to carry out cross-lingual word-sense alignment tasks in the future.

The relevant information are then extracted for these words in French as described in the previous section. Given a lemma with identical part-of-speech and gender in TLFi and Wiktionnaire, a combination of all possible definition/sense matches are provided. This way, the annotator can select one of the following relations between every pair of senses in the two resources:

none: the two senses/definitions are not referring to the same concepts, i.e. meanings
exact: the two senses/definitions are semantically identical. In our meeting on September 29th, we referred to this category as strict exact indicating the fact that the two senses/definitions should strictly be identical.
broader: the first sense/definition completely covers the meaning of the second one and is applicable to further meanings
narrower: the first sense/definition is entirely covered by the second one, which is applicable to further meanings.
related: There are cases when the senses may be equal but the definitions in both dictionaries differ in key aspects despite being semantically related

These types of relations are based on SKOS's semantic relations. It should be noted that the proposed relations, except broader and narrower, are symmetric. Therefore, R(A, B) equals with R(B, A), where R() refers to the relation between A and B. Regarding broader, it is the inverse relation of narrower, i.e. if R(A, B)=broader, then R(B, A)=narrower.

Here are a few examples:

livre (noun, masculine) (TLFi, Wiktionnary)
- proposed relation: none
  - Ouvrage constituant un volume imprimé. (Wiktionnaire)
  - Registre destiné à recueillir les signatures et les commentaires des visiteurs. (TLF)
- proposed relation: exact
  - Assemblage de feuilles manuscrites ou imprimées destinées à être lues. (Wiktionnaire)
  - Assemblage de feuilles en nombre plus ou moins élevé, portant des signes destinés à être lus. (TLF)
- proposed relation: narrower
  - Ouvrage qui fait référence à un texte sacré (Wiktionnaire)
  - Ouvrage imprimé, relié ou broché, non périodique, comportant un assez grand nombre de pages. (TLF)
- proposed relation: broader
  - Ouvrage constituant un volume imprimé. (Wiktionnaire)
  - Ouvrage en vers ou en prose, d'une certaine étendue. (TLF)
- proposed relation: related
  - Chacune des parties principales de certains ouvrages. (Wiktionnaire)
  - Ensemble de feuilles de parchemin ou de papier écrites des deux côtés et rassemblées en cahiers (TLF)

The following, is another example that we annotated together for the word bannissement (noun) in TLF and Wiktionnaire:

TLF	Relation	Wiktionnaire
Peine politique criminelle infamante consistant dans la défense, pour le condamné, de résider sur le territoire national pendant une durée déterminée	none	Action de bannir, d’expulser une personne en lui interdisant de revenir ou résultat de cette action.
Simple mesure d'éloignement prise à l'égard de certaines personnes, notamment des membres de familles ayant régné	related	Action de bannir, d’expulser une personne en lui interdisant de revenir ou résultat de cette action.
Éloignement imposé de quelque chose	related	Action de bannir, d’expulser une personne en lui interdisant de revenir ou résultat de cette action.
Action de bannir; état qui en résulte	exact	Action de bannir, d’expulser une personne en lui interdisant de revenir ou résultat de cette action.

Inter-annotator agreement

First attempt

Initially, we took only three semantic relations into account with slightly different definitions: related, exact and none. For this annotation task, Krippendorff's alpha of the two groups of annotators is as follows:

Group 1 (annotated by Mathieu, Hee-Soo and Bruno): 0.29. 228 pairs of definitions are annotated.
Group 2 (annotated by Karen and Sina): 0.39. 227 pairs of definitions are annotated.

The following figure illustrates heatmaps of how each two annotators perform comparatively:

Here are the comparison of the annotations of group_1 and group_2.

New attempt

In our new attempt, the annotation task is carried out according to the new guidelines where five semantic relations are taken into account.

The Krippendorff's alpha of the two groups of annotators is as follows:

Annotators	5-class	2-class
Group 1 (HS - MC - BG) - Batch 1	0.39	0.484
Group 1 (HS - MC - BG) - Batch 2	0.492	0.61
Group 2 (SA - KF) - Batch 1
Group 2 (SA - KF) - Batch 2	0.83	0.88

In addition to the five classes, the inter-annotator agreement is calculated for alignable (0) versus non-alignable (1).

Here is the comparison of the annotations of group_1.

sinaahmadi/fr-align