camembert-base-edda-span-classification

This model is designed to identify and classify named entities (such as Spatial, Person, and MISC), nominal entities, spatial relations, and other relevant information such as geographic coordinates within French encyclopedic entries. It has been trained on the French Encyclopédie ou dictionnaire raisonné des sciences des arts et des métiers par une société de gens de lettres (1751-1772) edited by Diderot and d'Alembert (provided by the ARTFL Encyclopédie Project). Dataset: https://huggingface.co/datasets/GEODE/GeoEDdA

Class labels

The tagset is as follows:

• NC_Spatial: a common noun that identifies a spatial entity (nominal spatial entity) including natural features, e.g. ville, la rivière, royaume.
• NP_Spatial: a proper noun identifying the name of a place (spatial named entities), e.g. France, Paris, la Chine.
• Relation: spatial relation, e.g. dans, sur, à 10 lieues de.
• Latlong: geographic coordinates, e.g. Long. 19. 49. lat. 43. 55. 44.
• NC_Person: a common noun that identifies a person (nominal spatial entity), e.g. roi, l'empereur, les auteurs.
• NP_Person: a proper noun identifying the name of a person (person named entities), e.g. Louis XIV, Pline, les Romains.
• NP_Misc: a proper noun identifying entities not classified as spatial or person, e.g. l'Eglise, 1702, Pélasgique.
• Head: entry name
• Domain-Mark: words indicating the knowledge domain (usually after the head and between parenthesis), e.g. Géographie, Geog., en Anatomie.

Model Description

• Developed by: Ludovic Moncla and Hédi Zeghidi in the framework of the GEODE project.
• Model type: CamemBERT token-classification
• Repository: https://github.com/GEODE-project/ner-bert
• Language(s) (NLP): French
• License: cc-by-nc-4.0
• Dataset: https://zenodo.org/records/10530177

Bias, Risks, and Limitations

This model was trained entirely on French encyclopedic entries and will likely not perform well on text in other languages or other corpora.

How to Get Started with the Model

Use the code below to get started with the model.

from transformers import pipeline
import torch
from datasets import load_dataset


pipe = pipeline("token-classification", model="GEODE/camembert-base-edda-span-classification", aggregation_strategy="simple", device=device)

content = "* ALBI, (Géog.) ville de France, capitale de l'Albigeois, dans le haut Languedoc : elle est sur le Tarn. Long. 19. 49. lat. 43. 55. 44."

print(pipe(content))


# Output
[{'entity_group': 'Head',
  'score': 0.9918331,
  'word': 'ALBI',
  'start': 2,
  'end': 6},
 {'entity_group': 'Domain_mark',
  'score': 0.9260238,
  'word': '(Géog.',
  'start': 8,
  'end': 14},
 {'entity_group': 'NC_Spatial',
  'score': 0.99029493,
  'word': 'ville',
  'start': 16,
  'end': 21},
 {'entity_group': 'NP_Spatial',
  'score': 0.9919335,
  'word': 'France',
  'start': 25,
  'end': 31},
 {'entity_group': 'NC_Spatial',
  'score': 0.9903319,
  'word': 'capitale',
  'start': 33,
  'end': 41},
 {'entity_group': 'NP_Spatial',
  'score': 0.9919644,
  'word': "l'Albigeois",
  'start': 45,
  'end': 56},
 {'entity_group': 'Relation',
  'score': 0.98715705,
  'word': 'dans',
  'start': 58,
  'end': 62},
 {'entity_group': 'NP_Spatial',
  'score': 0.9919502,
  'word': 'le haut Languedoc',
  'start': 63,
  'end': 80},
 {'entity_group': 'Relation',
  'score': 0.98698694,
  'word': 'sur',
  'start': 92,
  'end': 95},
 {'entity_group': 'NP_Spatial',
  'score': 0.9921453,
  'word': 'le Tarn',
  'start': 96,
  'end': 103},
 {'entity_group': 'Latlong',
  'score': 0.99200517,
  'word': 'Long. 19. 49. lat. 43. 55. 44',
  'start': 105,
  'end': 134}]

Training Details

Training Data

The model was trained using a set of 2200 paragraphs randomly selected out of 2001 Encyclopédie's entries. All paragraphs were written in French and are distributed as follows among the Encyclopédie knowledge domains:

Knowledge domain	Paragraphs
Géographie	1096
Histoire	259
Droit Jurisprudence	113
Physique	92
Métiers	92
Médecine	88
Philosophie	69
Histoire naturelle	65
Belles-lettres	65
Militaire	62
Commerce	48
Beaux-arts	44
Agriculture	36
Chasse	31
Religion	23
Musique	17

The spans/entities were labeled by the project team along with using pre-labelling with early models to speed up the labelling process. A train/val/test split was used. Validation and test sets are composed of 200 paragraphs each: 100 classified as 'Géographie' and 100 from another knowledge domain. The datasets have the following breakdown of tokens and spans/entities.

	Train	Validation	Test
Paragraphs	1,800	200	200
Tokens	132,398	14,959	13,881
NC-Spatial	3,252	358	355
NP-Spatial	4,707	464	519
Relation	2,093	219	226
Latlong	553	66	72
NC-Person	1,378	132	133
NP-Person	1,599	170	150
NP-Misc	948	108	96
Head	1,261	142	153
Domain-Mark	1,069	122	133

Training Procedure

For full training details and results please see the GitHub repository: https://github.com/GEODE-project/ner-bert

Evaluation

• Overall micro-average model performances (token-based)

	Precision	Recall	F-score
	91.5	94.8	93.1

• Token-based model performances (Test set)

	Precision	Recall	F-score	Support
NC-Spatial	96.7	95.1	95.9	592
NP-Spatial	95.9	95.5	95.7	717
Relation	89.8	95.6	92.6	452
Latlong	97.0	98.5	97.7	789
NC-Person	70.4	78.4	74.2	222
NP-Person	88.6	90.4	89.5	198
NP-Misc	69.0	82.9	75.3	175
Head	97.3	98.0	97.6	254
Domain-mark	99.0	100.0	99.5	392

Acknowledgement

The authors are grateful to the ASLAN project (ANR-10-LABX-0081) of the Université de Lyon, for its financial support within the French program "Investments for the Future" operated by the National Research Agency (ANR). Data courtesy the ARTFL Encyclopédie Project, University of Chicago.