Biographical information constitutes a cornerstone of historical research. When synthesized with event data—encompassing temporal and spatial dimensions—it forms the fundamental nexus of historical inquiry: person, event, time, place, and object. As a result, biographical materials represent a vitally important genre in historiography. These biographies have been transmitted to later generations in various forms. Official dynastic histories, such as the Biographies (列傳) section of the Old Book of Tang (舊唐書) (Liu n.d.), provide official records of prominent Tang Dynasty figures. Local gazetteers (n.d.) offer administrative records of regional information, incorporating biographical sections such as “Local Worthies” (鄉賢), “Loyal Martyrs” (忠烈), and “Literary Figures” (文苑). Furthermore, epitaphs collected in the Qing Dynasty’s Complete Tang Prose (全唐文) (Dong 1814) supplement formal official records with details often lacking elsewhere, such as comprehensive family lineage information. Modern compilations like the Compilation of Tang Dynasty Epitaphs (唐代墓誌彙編) (Zhou / Zhao 1991), collated from excavated inscriptions, further expand this corpus.

The Sequel to the Taipei City Gazetteer (續修臺北市志, STCG henceforth) (Chen / Fan 2017) includes two volumes of biographies, comprising 162 figures in politics and economics, and 157 in society and culture. Using the biographies of Wu Jin-chuan (吳金川) and Lin Qing-hui (林清輝) (Wu and Lin, respectively, henceforth) as case studies, we demonstrate how advanced AI facilitates data analysis and integration tasks previously difficult to execute. By establishing links between scattered data regarding biographical subjects, this approach offers new opportunities for verification, comparison, and textual research. The ultimate objective is to establish a framework for generating high-quality, integrated biographical information.

The biographies in STCG contain many organized paragraphs. With the assistance of Large Language Models (LLMs) and appropriate prompt engineering, we extracted information regarding the subjects across different years to establish basic chronologies. While there are many ways to organize life history data, using time as the primary axis is most conducive to integrating information from multiple sources. An individual’s achievements may vary in volume from year to year. To create a demonstration suitable for this extended abstract, we used prompts to request the LLM to simultaneously summarize the content of years with heavier detailed records, as shown in the simplified chronology of Wu (Figure 1). Without the assistance of LLMs, further summarizing biographical data—which is inherently a summary itself—presents a non-trivial Natural Language Processing (NLP) challenge (Liu et al. 2019).

Figure 1. A chronology of Wu Jin-chuan

LLMs excel not only at summarization but also at classification. We instructed the LLM to categorize the major stages of the subjects’ lives, such as Family, Birth, Education, Career, and Retirement. Additionally, we separated careers with different attributes into sub-timelines. Figure 2 shows Wu’s chronology with added category tags, while Figure 3 displays Lin’s chronology. In Lin’s timeline, we can observe that he held concurrent positions in different career tracks during certain periods, i.e., Education and Educational Administration after 1946 and Party, Government, and Public Enterprises after 1952.

Figure 2. A thematic chronology of Wu Jin-chuan

Figure 3. A thematic chronology of Lin Qing-hui with parallel career tracks

Generative AI also demonstrates a high proficiency in synthesizing potential connections between scattered data points, thereby assisting historians in linking vast quantities of related information. The authors of STCG often provide data regarding relationships between biographical subjects; for instance, the text notes a co-authorship relationship between Wang Shi-lang (王詩朗) and Fei Xi-ping (費希平), a father-daughter relationship between Cheng She-wo (成舍我) and Cheng Lu-hsi (成露茜), and the assistance Zhu Zhao-yang (朱昭陽) received from Wu San-lian (吳三連). Identifying and mapping these pieces of information can be instrumental in helping historians reconstruct the causal nexus of historical events.

Most subjects included in STCG also possess dedicated Wikipedia pages. We can merge data from these personal pages into our dataset. Data from different sources may be consistent, complementary, or contradictory, thus providing opportunities for verification, supplementation, and textual collation. Table 1 lists specific discrepancies regarding Wu found between Wikipedia and STCG; for example, his father is listed variously as a Xiucai (licentiate) and a Gongsheng (tribute student), and there are conflicting records regarding his position in Northeast China in 1945 (Wikipedia n.d.). These discrepancies necessitate further textual research tasks, such as sourcing additional information about Wu from relevant specialized books (Sima 2005). Furthermore, we can employ web crawling techniques to extract additional information from the content of URLs listed in the Wikipedia references section.

Table 1. A Comparison of divergent biographical details for Wu Jin-chuan

Powerful AI data analysis software provides unprecedented services, allowing us to complete OCR, data organization, summarization, and classification. The results presented herein were generated using Gemini 3.0 Pro, though other Large Language Models could also be utilized. A comparative analysis of outputs from different models can further assist in identifying potential errors or hallucinations.

The workflow demonstrated above substantiates the efficacy of AI technologies in organizing, analyzing, synthesizing, and proofreading biographical data from diverse sources. Once the scale and quality of biographical data are enhanced, more advanced analytical paradigms become feasible. Regarding Named Entity Recognition (NER n.d.), our goal extends beyond identifying names within a biography to synthesizing records from diverse historical materials and diaries to map the connections between individuals across different events. Furthermore, the persistent challenge of name disambiguation in automated reading can be addressed by leveraging the contextual data provided by linked historical sources, so it is now easier to construct Social Network Analyses (SNA n.d.) that cover broader dimensions of time and space. The summarization and identification of the subjects’ careers demonstrated earlier illuminate opportunities for deeper analysis atop these social networks.

By restricting the LLM to source-grounded extraction and categorization, we mitigate the hallucination risks inherent in open-ended generation. This ensures AI serves as a reliable instrument for rigorous data organization, ultimately opening broader and more advanced horizons for historical research (Arthur 2017).

Our work was supported in part by grants 113-2221-E-004-008 and 114-2221-E-004-001-MY2 from National Science and Technology Council of Taiwan.