IndEL: Indonesian Entity Linking Benchmark Dataset for General and Specific Domains

In recent years, there has been a surge in natural language processing research focused on low-resource languages (LrLs), underscoring the growing recognition that LrLs deserve the same attention as high-resource languages (HrLs). This shift is crucial for ensuring linguistic diversity and inclusivity in the digital age. Despite Indonesian ranking as the 11th most spoken language globally, it remains under-resourced in terms of computational tools and datasets. Within the semantic web domain, Entity Linking (EL) is pivotal, linking textual entity mentions to their corresponding entries in knowledge bases. This process is foundational for advanced information extraction tasks, including relation extraction and event detection. To bolster EL research in Indonesian, we introduce IndEL, the first benchmark dataset tailored for both general and specific domains. IndEL was manually curated using Wikidata, adhering to a rigorous set of annotation guidelines. We used two Named Entity Recognition (NER) benchmark datasets for entity extraction: NER UI for the general domain and IndQNER for the specific domain. IndQNER focused on entities from the Indonesian translation of the Quran. IndEL comprises 4765 entities in the general domain and 2453 in the specific domain. Using the GERBIL framework, we use IndEL to evaluate the performance of various EL systems, such as Babelfy, DBpedia Spotlight, MAG, OpenTapioca, and WAT. Our further investigation reveals that within Wikidata, a significant number of NIL entities remain unlinked due to the limited number of Indonesian labels and the use of acronyms. Especially in the specific domain, transliteration and translation processes performed to create the Indonesian translation of the Quran contribute to the presence of entities in a descriptive form and as synonyms.