We have a new working paper at SocArXiv “Office, Rank, and Mobility: Decomposing Historical Chinese Bureaucratic Titles with Language Models”, lead authored by Bruce (Yue) Yu, introducing a language model based workflow for parsing the job titles of Qing officials in the China Government Employee Dataset-Qing (CGED-Q JSL).
Once of the challenges we have faced with studying careers is that the job titles recorded for many officials recorded in the CGED-Q are concatenations of the title of their substantive post with various bestowed or purchased titles, positions for which they are eligible to be promoted into, temporary assignments, and other information, without any punctuation. To measure promotion, we need to isolate the substantive post from all the other information so that it can be matched to a bureaucratic rank (pinji) that can be compared across editions, but isolating substantive posts manually would be so time consuming as to be impractical. Our efforts at automation using keywords, regular expressions etc., have also proven unreliable. At the same time, the other information in the compound titles, like the honorary and bestowed titles, is potentially useful as a measure of another dimension of status in the hierarchy.
In response to this, Bruce developed a workflow involving adapting a language model to the specific task of parsing the job titles, via training. General language models are not adequate since they do not have Qing job titles in their training data, at least not yet. He was supported in this work by two experts on Qing history, co-authors Chen Jun and Michael Chung. The workflow does a very good job of parsing the job titles. The working paper shows that it can be used in other situations, for example, bureaucratic titles of Joseon Korean officials recorded in hangul. We expect it should work well in other situations where a recorded job title is a compound of multiple pieces of information, like in the CGED-Q.
To demonstrate the usefulness of the approach, the working paper measures time trends in the promotion rates of officials based on their ‘clean’ substantive posts and the associated bureaucratic ranks, and shows that promotion rates declined after the Taiping rebellion. At the same time, the numbers of officials with purchased or bestowed titles, and specified eligibility, actually increased dramatically.
Here is the abstract:
We introduce a procedure that employs an adapted small language model to decompose
compound job titles in bureaucracies and other large organizations into substantively mean-
ingful components using a schema grounded in institutional context. This addresses an issue
commonly encountered in historical sources for China and neighboring countries according to
which recorded job titles include diverse additional attributes beyond the post, such as location,
eligibility, temporary assignments, additional titles, or honors, with no consistent format. This
makes it difficult to infer the status of officeholders at scale, because their job titles can’t be
matched to a canonical list that also specifies ranks or other measures of status. It may be
unclear whether a changed job title reflected a promotion, a demotion, or a lateral transfer. We
demonstrate the procedure by applying it to personnel records of 19th century officials in the
China Government Employee Database–Qing (CGED-Q) Jinshenlu (JSL). In the Qing bureau-
cracy, compound titles were common for senior officials and purchased and awarded honors,
temporary assignments, and locations were routinely recorded in job titles. The model achieves
average agreement with manual annotation of 94 percent across fields, outperforming larger
general-purpose models. The model works for job titles written in other scripts: we apply it to
the hangul form of the job titles in publicly available career records of Joseon dynasty Korean
officials and show that it has high accuracy in identification of departments and posts. Our
decomposition thus enables career mobility researchers not only to extract names of posts that
can be mapped to administrative rank, but also to separate and make use of other components
of job titles relevant to their interests. As a case study, we show that analysis of posts, addi-
tional titles, and eligibility extracted from job titles of late Qing officials yields novel insights
into dynamics of bureaucratic careers and the composition of officialdom. An analysis of offi-
cials from 1830 to 1905 reveals divergent trends after the Taiping Rebellion (1850–1864). While
promotion to a higher office within three years declined from 12% to 8%, eligibility for higher
office more than tripled, and the share receiving a higher additional title nearly doubled.