Relation Extraction

Event-Argument Relation Extraction corpus and model (Open Source). We support three relations (has Agent, hasLocation, hasDate) between events and 21 entity types (see article).
Performance: WojoodHadath (93.99%) WojoodOutOfDoman (74.90%)

  • Method: We formulated the relation extraction task as a Natural Language Inference task and fine-tuned a BERT model using a large set of sentence pairs (NLI dataset) extracted from the WojoodHadath Corpus. (see article).

    WojoodHadath Corpus: We extended the Wojood nested NER corpus (550K tokens), by manually annotating event entities with 3 relations.

    WojoodOutOdDomain Corpus: New corpus with 80K tokens in MSA, manually annotated with entities and relations as WojoodHadath. It covers 10 domains (Economics, Finance, Politics, Science, Technology, Art, Law, Agriculture, History, and Sports).

    Relations:

    has Agent: participant(s) involved in the event (Domain: Event, Range: PERS, ORG, OCC, NORP)
    hasLocation: where the event occurred (Domain: event, Range: GPE, LOC, FAC)
    hasDate: when the event occurred (Domain: event, Range: TIME, DATE)


    Please email Prof. Jarrar (mjarrar AT birzeit.edu) for the annotation guidelines

  • SinaTools: Relation Extraction module as python library.

    GitHub: training source code + sample data (~20 sentences with event mentions).

    Hugging Face: fine-tuned BERT model using WojoodHadath.

    WojoodHadath (Corpus only)

    WojoodOutOfDomain (Corpus only)

  • Coming soon

  • Alaa Aljabari, Lina Duaibes, Mustafa Jarrar, Mohammed Khalilia: Event-Arguments Extraction Corpus and Modeling using BERT for Arabic. In Proceedings of the Second Arabic Natural Language Processing Conference (ArabicNLP 2024), Bangkok, Thailand. Association for Computational Linguistics.