This paper reports on an ongoing project that develops an AI-assisted workflow for the qualitative annotation of Chinese Buddhist monastic biographies encoded in TEI. The project builds on a previously completed qualitative coding study of the Gaoseng zhuan 高僧傳 (compiled early sixth century) and extends the same analytical framework to later biographical corpora, most notably the Xu gaoseng zhuan 續高僧傳 (compiled mid seventh century). Its central aim is to examine, using a hybrid BERT–LLM pipeline, whether social and religious patterns identified in early medieval monastic biographies persist, shift, or dissolve in later historical contexts.

The project takes as its point of departure a manually coded dataset of 227 Gaoseng zhuan biographies, annotated with a fixed codebook of forty social and religious traits. That earlier study showed that qualitative coding, when combined with statistical and network analysis, can reveal structural patterns in monastic culture that are difficult to discern through close reading alone (Lee 2020; Lee 2021; the dataset itself is under preparation for web publication). In particular, it identified a division between elite scholar-monks associated with aristocratic and imperial support, institutional authority, and doctrinal activity, and ascetic practitioners characterized by bodily discipline, devotional practice, meditation, miracle-working, and closer interaction with non-elite lay communities. These findings aligned with emic classificatory distinctions found in medieval sources, including the opposition between “masters” (shi 師) and “ascetics” (kujie 苦節) in the now-lost Mingseng zhuan 名僧傳.

The present project builds on this qualitative framework in two respects. First, it addresses scale. Manual qualitative coding is effective but labor-intensive, which limits its application to later and larger biographical collections (the XGSZ for example, contains around 470 biographies of Chinese monks, and an even later collection Song gaoseng zhuan 宋高僧傳 around 500). Second, it addresses diachrony. The XGSZ reflects a period in which Buddhist institutions, practices, and prestige—most notably those associated with meditation and emerging Chan 禪 traditions—had undergone significant transformation. A central research question is whether patterns observed in the GSZ remain stable over time. For example, in the GSZ dataset, engagement in meditation exhibited a strong negative association with scholarship. The project asks whether this relationship persists in the XGSZ, even after meditation had acquired greater cultural and institutional prestige with the rise of the Chan tradition.

To make such comparisons feasible, the project develops a hybrid annotation workflow that combines a fine-tuned BERT model with a large language model within a TEI-based environment. The workflow applies the same forty qualitative traits used in the earlier manual study, ensuring conceptual continuity between datasets. In the first stage, a BERT-based classifier (hfl/chinese-roberta-wwm-ext), fine-tuned on the manually coded GSZ corpus, identifies candidate passages corresponding to specific traits at the sentence or short-span level. In the second stage, a large language model accessed via API evaluates these candidates by considering their broader narrative context. It filters implausible attributions of qualitative codes and adjusts the range of the passages to be applied with these codes. After the BERT and LLM stages, the researcher reviews the proposed annotations, confirming, revising, or rejecting them as necessary. All accepted annotations are encoded directly in TEI, preserving transparency and enabling later inspection, reuse, and comparison.

The output of this workflow is a set of TEI-encoded biographies annotated with social and religious traits that are directly comparable to the manually coded GSZ dataset. This enables controlled diachronic comparison using the same statistical and network-analytic methods applied in the earlier study.

The proposed pipeline differs from earlier approaches such as Nelson (2020) in that it builds on a gold-standard dataset produced through human qualitative annotation—specifically, the manually coded GSZ—rather than relying on machine learning to inductively discover patterns. It also differs from work such as Jiang et al. (2023) and Xu et al. (2025) in that it aims to identify complex semantic and narrative units, rather than named entities.

The paper presents the design of the pipeline, the structure of the codebook, and the historical questions guiding the comparison between the GSZ and XGSZ, and seeks to open discussion among scholars working on digital biography, qualitative annotation, and AI-assisted humanities research.