Tafsir Dataset: A Novel Multi-Task Benchmark for Named Entity Recognition and Topic Modeling in Classical Arabic Literature

Proceedings of the 29th International Conference on Computational Linguistics, pages 3753–3768 October 12–17, 2022. 3753 Tafsir Dataset: A Novel Multi-Task Benchmark for Named Entity Recognition and Topic Modeling in Classical Arabic Literature Sajawel Ahmed1,2,3, Rob van der Goot3, Misbahur Rehman2, Carl Kruse2, Ömer Özsoy2, Alexander Mehler1, Gemma Roig1 1Faculty for Computer Science and Mathematics, Goethe University Frankfurt 2Faculty for Linguistics, Cultures, and Arts, Goethe University Frankfurt 3Department of Computer Science, IT University of Copenhagen {sahmed}@em.uni-frankfurt.de Abstract Various historical languages, which used to be lingua franca of science and arts, deserve the attention of current NLP research. In this work, we take the first data-driven steps to- wards this research line for Classical Arabic (CA) by addressing named entity recognition (NER) and topic modeling (TM) on the ex- ample of CA literature. We manually anno- tate the encyclopedic work of Tafsir Al-Tabari with span-based NEs, sentence-based topics, and span-based subtopics, thus creating the Tafsir Dataset with over 51,000 sentences, the first large-scale multi-task benchmark for CA. Next, we analyze our newly generated dataset, which we make open-source available, with current language models (lightweight BiL- STM, transformer-based MaChAmP) along a novel script compression method, thereby achieving state-of-the-art performance for our target task CA-NER. We also show that CA-TM from the perspective of historical topic mod- els, which are central to Arabic studies, is very challenging. With this interdisciplinary work, we lay the foundations for future research on automatic analysis of CA literature. 1 Introduction All languages deserve equal technologies. Named entity recognition (NER) and topic modeling (TM) are a crucial part of various downstream tasks in natural language processing (NLP), such as Entity Linking, Relation Extraction, and ultimately Ques- tion Answering. For such tasks, many research in- stitutes and individual scholars put their emphasis on popular, high-resource languages like English, where there is already a large amount of previ- ous work and resources available (Rajpurkar et al., 2018; Dzendzik et al., 2021; Cambazoglu et al., 2021). This definitely accelerates the progress of the ongoing data-driven NLP. However, many his- torical languages, such as Ancient Egyptian, An- cient Greek, and especially Classical Arabic (CA), which used to be the lingua franca of science and arts, have been mostly neglected by the NLP com- munity. These languages possess large volumes of historical literature (CA: e.g. Liber Algebrae et Almucabola, Canon Medicinae, Tafsir Al-Tabari), which were and still are to this date relevant for many communities and societies, lay their foun- dations and even shape their further development. In order to perform historical analysis which are relevant for our modern age, we need to let these forgotten low-resource languages benefit from the wave of machine learning (ML) progress, thus mak- ing historical texts accessible to modern studies and approaching ethically an egalitarian state of NLP research. To this end, within the project Linked Open Tafsir (Ahmed et al., 2022), firstly, we create the Tafsir Dataset by annotating the CA encyclope- dic books of Tafsir Al-Tabari on exegetical studies of law, ethics and philosophy. This is …

Content truncated to 500 words