Wojood: Nested Arabic Named Entity Corpus and Recognition using BERT

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 3626–3636 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 3626 Wojood: Nested Arabic Named Entity Corpus and Recognition using BERT Mustafa Jarrar, Mohammed Khalilia, Sana Ghanem Birzeit University Palestine {mjarrar, mkhalilia, swghanem}@birzeit.edu Abstract This paper presents Wojood, a corpus for Arabic nested Named Entity Recognition (NER). Nested entities occur when one entity mention is embedded inside another entity mention. Wojood consists of about 550K Modern Standard Arabic (MSA) and dialect tokens that are manually annotated with 21 entity types including person, organization, location, event and date. More importantly, the corpus is annotated with nested entities instead of the more common flat annotations. The data contains about 75K entities and 22.5% of which are nested. The inter-annotator evaluation of the corpus demonstrated a strong agreement with Cohen’s Kappa of 0.979 and an F1-score of 0.976. To validate our data, we used the corpus to train a nested NER model based on multi-task learning using the pre-trained AraBERT (Arabic BERT). The model achieved an overall micro F1-score of 0.884. Our corpus, the annotation guidelines, the source code and the pre-trained model are publicly available. Keywords: Named Entity Recognition, Multi-Task Learning, Nested Entities, BERT, Arabic NER Corpus 1. Introduction Named Entity Recognition (NER) is integral to many Natural Language Processing (NLP) applications such as chatbots and question answering (Shaheen and Ezzeldin, 2014), information retrieval (Guo et al., 2009), machine translation (Hassan and Sorensen, 2005), word-sense disambiguation (Al-Hajj and Jarrar, 2021a), clustering (Abuaiadah et al., 2017), interoper- ability (Jarrar et al., 2011), topic modeling and event discovery (Feng et al., 2018), among others. NER is the task of identifying named entity mentions in un- structured text and classify them to a predefined cate- gories such as person, organization, event, location, or date. For example, in a sentence like ”Sami works at Jimmy Carter Center”, ”Sami” would be tagged as a person, and ”Jimmy Carter Center” as an organization, as shown in Figure 1. We use the IOB2 tagging scheme, where B indicates the beginning of the entity mention, I indicates the inside token, and O refers to outside en- tity (Sang and Veenstra, 1999). The NER task becomes a sequential labeling task, where each token is tagged with a label based on the contextual information sur- rounding it (Sohrab and Miwa, 2018). Figure 1: Flat NER Example Many approaches were proposed for NER in many lan- guages, including Arabic, which we will review in Sec- tion 2. However, less attention is given to nested named entities (Finkel and Manning, 2009), which are entity mentions contained inside other entity mentions. For example, ”Jimmy Carter” is a person that is inside the ”Jimmy Carter Center” organization, and ”Birzeit” is a geopolitical entity inside the organization mention ”Birzeit University”. More examples in Arabic are shown in Figure 2. There are a few reasons why nested NER is understud- ied. First and most important is lack of data, specially for …

Content truncated to 500 words