Month: September 2026

Can AI Understand Buddhist Concepts?

~ Victor ~ Leave a comment

[This post is based on Arjuna Hellingman’s Bachelor AI Thesis]

Religious texts are often rich in metaphor, symbolism, and implicit meaning, making them challenging for modern AI systems. While recent advances in natural language processing have improved concept extraction in domains such as news and scientific literature, much less is known about how these techniques perform on religious texts.

In his Bachelor’s thesis, Arjuna Hellingman investigated whether modern semantic AI models can identify key Buddhist concepts in the Dhammapada, one of the most influential texts in Theravada Buddhism. The study compared a traditional rule-based lexical approach with two embedding-based semantic models, BGE-base and BGE-m3.

To create an evaluation benchmark, three Buddhist monks annotated 120 verses from the Dhammapada using a predefined set of 32 core Buddhist concepts, including suffering, craving, enlightenment, wisdom, and compassion. These annotations were combined into two gold standards: a liberal version that included concepts identified by at least one monk, and a stricter majority version requiring agreement from at least two monks.

The results revealed an interesting trade-off. On the liberal gold standard, the semantic models performed best, identifying a broader range of concepts and achieving higher overall F1 scores. On the stricter majority gold standard, however, the simple rule-based method achieved the highest performance due to its greater precision. Semantic models generally found more relevant concepts but also produced more false positives.

One particularly surprising finding was that the English-focused BGE-base model outperformed the larger multilingual BGE-m3 model on several abstract Buddhist concepts, including craving, wisdom, rebirth, and enlightenment. This suggests that multilingual models do not necessarily have an advantage when working with English translations of religious texts.

Beyond the specific results, the thesis demonstrates the importance of evaluation design in digital humanities research. The choice of gold standard significantly influenced how the models were judged, highlighting that benchmark construction is as important as model selection when applying AI to low-resource cultural and religious domains.

Reference: Arjuna Hellingman. Concept Extraction under Domain Shift: A Comparative Study of Lexical and Semantic Methods in the Dhammapada. Bachelor’s Thesis, Vrije Universiteit Amsterdam, 2026

All data and supplementary materials can be found here: https://zenodo.org/records/20794173

Concept_Extraction_under_Domain_Shift__a_Comparative_Study_of_Lexical_and_Semantic_Methods_in_the_Dhammapada___Download

Share This:

Making Oral WW2 History More Searchable with Knowledge Graphs and AI

~ Victor ~ Leave a comment

[This post is based on Feruza Bakhtiyorova‘s Bachelor AI thesis]

Oral history interviews are rich sources of personal experiences, but they are often difficult to search. A single interview segment may refer to specific wartime locations, events, organizations, or people, while also touching on broader themes such as persecution, migration, or survival. Making these interviews accessible therefore requires high-quality metadata and meaningful concept annotations.

In her Bachelor’s thesis, Feruza Bakhtiyorova investigated how knowledge graphs can help improve the automatic annotation of oral history interviews from the WO2Net project. The existing workflow uses large language models (LLMs) to link interview segments to concepts from the WO2 Thesaurus, a controlled vocabulary of World War II-related concepts. Earlier work had shown that these links were often too broad or failed to capture important specific concepts such as camps, events, or organizations.

To address this problem, the thesis explored a knowledge graph retrieval-augmented generation (KG-RAG) approach. Instead of asking the language model to choose only from existing candidate concepts, the system first retrieves additional concepts from the WO2 Thesaurus and then asks the model to select the concepts best supported by the interview segment.

The study compared three approaches: baseline prompting, refined prompting, and KG-RAG prompting. Five real oral history segments were evaluated by domain experts, who judged both the correctness and specificity of the assigned concepts.

The results showed a nuanced picture. KG-RAG did not outperform the other methods in overall correctness, but it consistently produced more specific concepts. In several cases it successfully recovered important missing concepts such as Birkenau, Reichenbach, Kristallnacht, and Bergen-Belsen, which were unavailable to the prompting-only approaches. At the same time, retrieving more candidate concepts also increased the risk of introducing less relevant suggestions.

The thesis demonstrates that knowledge graphs can complement large language models by providing richer domain knowledge. While retrieval alone does not guarantee more accurate annotations, it can help generate more precise and historically meaningful concept links, improving access to large oral history collections and supporting future research into cultural heritage archives.

The project’s code can be found at https://github.com/ferufera/wo2-segmentation-optimization

Feruza Bakhtiyorova. Improving Concept Specificity in LLM-Based Concept Linking for Oral History Segments Using the WO2 Thesaurus. Bachelor’s Thesis, Vrije Universiteit Amsterdam, 2026.

Feruza_Bakhtiyorova_2740961Bachelor_ThesisDownload

Share This: