Command Palette
Search for a command to run...
Louis Martin Benjamin Muller Pedro Javier Ortiz Suárez Yoann Dupont Laurent Romary Éric Villemonte de la Clergerie Djamé Seddah Benoît Sagot

摘要
预训练语言模型如今已在自然语言处理领域广泛应用。尽管取得了显著成功,但大多数现有模型要么仅基于英语数据训练,要么基于多种语言数据的拼接进行训练。这使得这些模型在除英语以外的其他语言中的实际应用受到极大限制。本文探讨了为其他语言(以法语为例)训练单语Transformer架构语言模型的可行性,并在词性标注、依存句法分析、命名实体识别和自然语言推理四项下游任务上评估了所构建的语言模型。研究结果表明,使用网络爬取的数据相较于维基百科数据更具优势。更令人意外的是,仅使用较小规模的网络爬取数据集(4GB)即可取得与使用更大规模数据集(130GB以上)相当甚至更优的性能。我们提出的最优模型CamemBERT在上述四项任务中均达到或超越了当前最优水平,展现了强大的语言建模能力。
代码仓库
Karthik-Bhaskar/Context-Based-Question-Answering
tf
GitHub 中提及
hbaflast/bert-sentiment-analysis-pytorch
pytorch
GitHub 中提及
anaishoareau/french_preprocessing
pytorch
GitHub 中提及
huggingface/transformers
官方
pytorch
GitHub 中提及
pwc-1/Paper-8/tree/main/camembert
mindspore
bourrel/French-News-Clustering
tf
GitHub 中提及
hbaflast/bert-sentiment-analysis-tensorflow
tf
GitHub 中提及
基准测试
| 基准 | 方法 | 指标 | 
|---|---|---|
| dependency-parsing-on-french-gsd | CamemBERT | LAS: 92.47 UAS: 94.82  | 
| dependency-parsing-on-partut | CamemBERT | LAS: 92.9 UAS: 95.21  | 
| dependency-parsing-on-sequoia-treebank | CamemBERT | LAS: 94.39 UAS: 95.56  | 
| dependency-parsing-on-spoken-corpus | CamemBERT | LAS: 81.37 UAS: 86.05  | 
| named-entity-recognition-on-french-treebank | CamemBERT (subword masking) | F1: 87.93 Precision: 88.35 Recall: 87.46  | 
| natural-language-inference-on-xnli-french | CamemBERT (large) | Accuracy: 85.7  | 
| natural-language-inference-on-xnli-french | CamemBERT (base) | Accuracy: 81.2  | 
| part-of-speech-tagging-on-french-gsd | CamemBERT | UPOS: 98.19  | 
| part-of-speech-tagging-on-partut | CamemBERT | UPOS: 97.63  | 
| part-of-speech-tagging-on-sequoia-treebank | CamemBERT | UPOS: 99.21  | 
| part-of-speech-tagging-on-spoken-corpus | CamemBERT | UPOS: 96.68  |