Konooz
收藏资源简介:
Konooz是一个新颖的多维度命名实体识别(NER)语料库,旨在解决阿拉伯方言在NLP任务中的低资源问题。该语料库由伯利兹大学、哈马德·本·哈利法大学和巴勒斯坦理工大学-卡多里的研究人员创建,涵盖了10个领域和16种不同的阿拉伯方言,共计160个独立语料库。语料库包含约777k个标记,经过人工收集和标注,标注了21种实体类型。该语料库对于基准测试跨领域和跨方言的NER模型非常有用,并且已经通过使用Konooz对四个阿拉伯NER模型进行基准测试来展示其价值。此外,该语料库还用于深入分析不同领域和方言之间的词汇相似性,揭示了语言变体对模型性能的影响。Konooz是一个开源数据集,可供公众在https://sina.birzeit.edu/wojood/#download上访问。
Konooz is a novel multi-dimensional named entity recognition (NER) corpus designed to address the low-resource issue of Arabic dialects in NLP tasks. This corpus was created by researchers from Birzeit University, Hamad bin Khalifa University, and Palestine Technical University-Kadoorie, covering 10 domains and 16 distinct Arabic dialects, totaling 160 independent sub-corpora. The corpus contains approximately 777k annotated tokens, which were manually collected and labeled with 21 entity types. It is highly valuable for benchmarking cross-domain and cross-dialect NER models, and its utility has been demonstrated by benchmarking four Arabic NER models using Konooz. In addition, this corpus has also been used to conduct in-depth analyses of lexical similarity across different domains and dialects, revealing the impact of language variation on model performance. Konooz is an open-source dataset accessible to the public at https://sina.birzeit.edu/wojood/#download.




