propp-fr_NER_camembert-large_FAC_GPE_LOC_PER_TIME_VEH

INTRODUCTION:

This model, developed as part of the propp-fr project, is a NER model built on top of camembert-large embeddings, trained to predict nested entities in french, specifically for literary texts.

The predicted entities are:

• mentions of characters (PER): pronouns (je, tu, il, ...), possessive pronouns (mon, ton, son, ...), common nouns (le capitaine, la princesse, ...) and proper nouns (Indiana Delmare, Honoré de Pardaillan, ...)
• facilities (FAC): chatêau, sentier, chambre, couloir, ...
• time (TIME): le règne de Louis XIV, ce matin, en juillet, ...
• geo-political entities (GPE): Montrouge, France, le petit hameau, ...
• locations (LOC): le sud, Mars, l'océan, le bois, ...
• vehicles (VEH): avion, voitures, calèche, vélos, ...

MODEL PERFORMANCES (LOOCV):

TRAINING PARAMETERS:

• Entities types: ['PER', 'LOC', 'FAC', 'TIME', 'VEH', 'GPE']
• Tagging scheme: BIOES
• Nested entities levels: [0, 1]
• Split strategy: Leave-one-out cross-validation (28 files)
• Train/Validation split: 0.85 / 0.15
• Batch size: 16
• Initial learning rate: 0.00014

MODEL ARCHITECTURE:

Model Input: Maximum context camembert-large embeddings (1024 dimensions)

• Locked Dropout: 0.5

• Projection layer:

  • layer type: highway layer
  • input: 1024 dimensions
  • output: 2048 dimensions

• BiLSTM layer:

  • input: 2048 dimensions
  • output: 256 dimensions (hidden state)

• Linear layer:

  • input: 256 dimensions
  • output: 25 dimensions (predicted labels with BIOES tagging scheme)

• CRF layer

Model Output: BIOES labels sequence

HOW TO USE:

Propp Documentation

TRAINING CORPUS:

PREDICTIONS CONFUSION MATRIX:

CONTACT:

mail: antoine [dot] bourgois [at] protonmail [dot] com