By Salah-Dean Satouri*
As technology continues to evolve, artificial intelligence and machine learning will undoubtedly play a central role in advancing the study of Islamic law. This report will (1) propose potential applications of AI in canon categorization and (2) examine the major limitations and challenges that may be encountered in its application.
1. Future Uses
(a) Automating Canon Tagging
Moving forward, it would be highly beneficial to automate canon tagging by employing linguistic pattern matching and keyword detection, as AI excels at matters of pattern recognition and categorization.[1] Can on tagging involves two main core components: (1) identifying patterns and (2) assigning classifications. Both actions are "fundamental to the concept in Artificial Intelligence," as it is "what allows AI systems to make sense of the vast amounts of data they're trained on."[2] For instance, when encountering the word "contract" in a canon, an AI, like humans, could associate it with the Sales and Contracts category. While a single word does not definitively determine the category, it provides a starting point for the categorization to occur.
By combining keyword detection with phrase recognition and contextual analysis, AI systems can effectively categorize canons. For example, the canon the denial of the contract by one party and the other's intention not to dispute it results in dissolution (juḥūd aḥad al-mutaʿāqidayn al-ʿaqd wa-ʿazm al-ākhar ʿalā ʿadam al-khuṣūma faskh; جحود أحد المتعاقدين العقد وعزم الآخر على عدم الخصومة فسخٌ),[3] was correctly categorized by AI under Sales and Contracts and its subfield Contracts because of its ability to recognize linguistic patterns and contextual meanings.
2. Limitations
(a) Linguistic Complexity
A major limitation in both OCR and canon tagging lies in the linguistic complexity of the languages these canons are written in. Islamic legal canons are often found in Perso-Arabic scripts such as Arabic, Urdu, Persian, and Ottoman Turkish, which pose unique challenges for modern language models and AI systems. Transformer-based Models for Arabic, such as AraBERT or ALLaM, help resolve this issue.[4] However, even with advanced models, understanding classical Islamic texts can be quite difficult. One issue is the calligraphic and dialectical variation that exists amongst classical texts, which are often written in ornate fonts and can be difficult for OCR to process.[5] Another issue is that Arabic script relies on "diacritics and ligatures [known as harakat] to indicate short vowels and certain consonant combinations," and OCR systems often struggle "to recognize and process these diacritics and ligatures correctly."[6]
(b) Front-end Labor
Further, for AI to effectively tag and gather canons using ML and OCR, significant upfront labor is required to digitize and scan the Islamic texts. Further, it would require that the texts that are uploaded are not worn or damaged and can be read by an OCR. However, if the AI can be trained on the current extent of the SHARIAsource Lab canons that have been manually gathered, then once the labor-intensive job of digitizing the works is done, the AI could extract canons at an extremely rapid rate.
Conclusion
Although I loved tagging canons, I am afraid if it was a competition between me and an AI, I would lose. By leveraging AI's skills in pattern recognition and OCR canon gathering, we can streamline the processes of tagging and gathering, making Islamic legal resources more accessible and efficient.
Notes:
* Salah-Dean Satouri is a J.D. Candidate at Harvard Law School with a focus on international human rights, democratization, and advocacy for Muslim Americans. He graduated magna cum laude and with first-class honors from the Joint Degree Program at the College of William & Mary and the University of St. Andrews in 2022, earning a B.A. International Honors in International Relations. A former Global Law Scholar at Georgetown University Law Center and a Lebanese-Tunisian American raised in Northern Virginia, Salah-Dean has conducted extensive research and advocacy on constitutional legitimacy and democratic backsliding in the Middle East and North Africa, contributing to publications like The Washington Post, Brookings Institute, and Voice of America.
[1] I do not know how difficult it would be to leverage AI to the SHARIAsource programming. But it would be something to consider, even if just by using open-source generative models like ChatGPT to make tagging easier.
[2] Alexander Stahl, "How Pattern Recognition is Improving Lives," Medium, March 22, 2024, https://medium.com/@stahl950/how-pattern-recognition-is-improving-lives-bf9ee7e988d2#:~:text=Recognizing%20these%20patterns%20%E2%80%94%20both%20as,events%2C%20whether%20positive%20or%20negative.
[3] SHARIAsource CnC Database Canon No. 2638 (citing Muḥammad Ṣidqī Būrnū, Mawsūʿat al-qawāʿid al-fiqhiyya (3d ed., 2015), 3:13).
[4] See M. Saiful Bari et al., "ALLaM: Large Language Models for Arabic and English," arXiv, July 22, 2024, https://arxiv.org/abs/2407.15390; Wissam Antoun et al., "AraBERT: Transformer-based Model for Arabic Language Understanding," ACL Anthology (2020): 9–15.
[5] See Safiullah Faizullah et al., "A Survey of OCR in Arabic Language: Applications, Techniques, and Challenges," Applied Sciences 13, no. 7 (2023); Atique Ur Rehman and Sibt ul Hussain, "Large Scale Font Independent Urdu Text Recognition System," arXiv, May 15, 2020, https://arxiv.org/pdf/2005.06752.
[6] Faizullah et al., "A Survey of OCR in Arabic Language."
No comments:
Post a Comment