The analysis of ancient Chinese texts tends to disregard their phonological parameters, largely for reasons that may be described as practical. Scholarly work generally relies on transcriptions into contemporary standard Mandarin, whether in earlier studies such as Pulleybank (1995) or in more recent research such as Sun (2020). At the same time, the lack of consensus regarding the reconstruction of Archaic Chinese discourages the incorporation of phonological information into linguistic or philological research. As a result, studies of pre-Han Chinese typically proceed without systematic attention to sound structure, despite its relevance for various morphological and syntactic phenomena, as proven by Jacques (2022). This is illustrated by (1), where 壞 is read kwɛjH (not ɣwɛjH), indicating transitivity:

Transcription into a contemporary language is not in itself an obstacle to philological inquiry, but remains insufficient for analyses relying on phonological evidence. Although Archaic Chinese reconstruction draws on diverse sources (rhyming, phonetic series, etc.), morphological alternations are rarely represented in the script; phonetic glosses thus provide virtually the only direct evidence. Middle Chinese, closer to earlier stages, offers a practical intermediary for capturing such phenomena, with well-established reconstruction standards (Baxter / Sagart 2014).

To date, no corpus of ancient Chinese texts annotated with Middle Chinese glosses is available. Such annotation is therefore well motivated, especially since most Archaic Chinese alternations leave no trace in Middle Chinese, and representing them directly at the Archaic level would be misleading. However, creating such a resource presents two main challenges. First, fully manual annotation is prohibitively time-consuming, while fully automatic annotation cannot realistically be achieved. Some characters exhibit multiple alternative readings, and textual commentaries are not always sufficient to identify which one is appropriate. Second, preexisting high-quality tools and datasets, such as the Baxter-Sagart reconstruction (Baxter / Sagart 2014) or SinoPy (List 2018), do not provide sufficient coverage to enable large-scale pre-annotation character by character. This lack of resources constitutes a significant gap not only in sinology but also more broadly within digital humanities, where phonologically enriched corpora remain rare, especially for non-alphabetic traditions.

To address these limitations, we propose a processing pipeline designed to produce richly annotated corpora of pre-Han Chinese texts, with phonetic glosses. It currently consists of two components: a database of character transcriptions and a Python library dedicated to automatic pre-annotation and corpus generation.

The database PHONO-ML (Delaporte / Jacques 2025) was built from the data of the rhyme dictionary Guangyun 广韵, which is widely used for Middle Chinese reconstruction. We developed a Python program that applies to the contents of the Guangyun the principles used to manually reconstruct Middle Chinese pronunciation, as described by Baxter (1992) and Jacques (2021). The resulting database contains 26,224 character/transcription associations, covering a set of more than 20,000 distinct characters, substantially expanding the range of phonologically annotated data available for digital research. A significant subset of approximately 2,700 characters has been validated against The Baxter-Sagart reconstruction of Old Chinese (Baxter / Sagart 2020), thus strengthening the reliability of the dataset.

PHONO-ML is organized into three CSV files: a character-centered table listing all attested readings for each character, a transcription-centered table listing all characters associated with a given reading, and a table of character/reading pairs, where each row represents a single association. While this three-file structure may seem redundant at first glance, it accommodates diverse research workflows while minimizing the need for data reformatting and preparatory handling. The simplicity of the CSV format enhances accessibility and makes the database easily reusable across projects. Since the sound glosses were reconstructed automatically, we additionally provide the fanqie 反切 and the sixtuples extracted from the Guangyun to enable manual verification in cases of doubt.

Developed as a continuation of the PHONO-ML database, the Python library YunPy (Delaporte / Xie 2025) serves as both an exploration tool for PHONO-ML and as the next stage of the processing pipeline leading to corpus creation. Its exploration functionalities enable rapid and direct querying of the database: checking whether a character is included, retrieving its known transcriptions, and obtaining detailed information on single characters or characters from a string. These features improve usability for researchers who may use the data sporadically and do not need to automatically annotate long texts.

More importantly, the annotation functionalities of YunPy make it possible to generate fully pre-annotated corpora. The library produces a character-level (pre-)annotation in which each character is associated with all of its attested readings in the database. Batch processing enables the annotation of entire textual collections, facilitating large-scale corpus construction. Output formats currently include Webanno TSV 3.3 (Figure 1), compatible with widely used annotation tools such as INCEpTION (Klie et al. 2018), and a simple ad-hoc XML format designed for easy conversion into other formats. These outputs are intended to support the subsequent stage of human annotation, during which annotators select the appropriate reading in context whenever multiple possibilities exist.

Figure 1: Sample output from YunPy

Although the processing pipeline remains incomplete, our ongoing work aims to integrate human validation into the workflow. Ultimately, the resulting corpora will present ancient Chinese texts in which each character is assigned an exact Middle Chinese phonetic transcription. Such corpora will constitute a new type of digital resources: phonologically enriched datasets that provide researchers with transcriptions closer to the texts’ original pronunciation. Beyond their obvious relevance for linguistics and philology, these resources may support broader objectives within digital sinology, such as the reconstruction of Archaic Chinese and the development of models capable of fully automatic phonological annotation via machine learning.

By expanding the range of available annotated materials and by addressing an understudied area of phonological and historical research, this project contributes to increasing the diversity of topics and perspectives represented within digital humanities. At the same time, its open, extensible, and computationally oriented design aims to support researchers from a wide range of linguistic and disciplinary backgrounds.