Listed here are the titles and abstracts for papers that will be delivered on May 31 2025, for the DO 2025 Virtual Conference on the theme “AI and the Digital Humanities for the Study of Asia, Africa, and Oceania.” The details here are subject to change. Speakers are listed in alphabetic order.
Keynote Speakers
Ephrem A. Ishac (Austrian Academy of Sciences), Who is the Public in a Digital Age? Democratization of HTR—Syriac as Example
Handwritten Text Recognition (HTR) technology promises extraordinary access to vast textual heritage in manuscripts. Yet, the question of who truly benefits from these advancements remains critical. This keynote aims to address this crucial question, prompted in part by the academic engagement and the definitions of “public” that can emerge even in discussions around outreach.
Moving beyond purely technical discussions, we examine the current landscape of HTR through the lens of “democratization,” a concept I explored in my previous Digital Orientalist post, “From Vienna to the World: Launching the First Public Syriac HTR Model on Transkribus.” We ask: does “the public” simply mean “everyone and everywhere,” as the digital realm theoretically suggests, or are there barriers—technical, conceptual, and institutional—that limit HTR’s reach?
A key demonstration of HTR’s democratizing principle lies in its ability to unlock inaccessible knowledge. We will illustrate this through practical examples, showcasing how the development of digital Syriac corpora, built through OCR, HTR and AI technologies, has been instrumental in the identification of fragments from the Syriac world of written culture, distributed in several libraries worldwide. These successes underscore the tangible impact of making advanced tools and resources more widely available.
Finally, this keynote will invite us to rethink and expand our understanding of “the public.” Thanks to public collaboration, the digital revolution has, since its inception, been crossing borders. The “public” with the internet’s digital network ideals is, in fact, everyone, with the right to access knowledge, everywhere. This foundational concept of democratization—beyond mere mottos—has been and continues to be a defining force shaping our current digital age.
Antonia Karaisl (Waseda University), Fast Cars need Good Brakes – OCR Quality Control in an Age of AI
Optical Character Recognition (OCR) technology is fundamental to the field of Digital Humanities. In a time of mass digitization of historic books, OCR helps unlocking the content of scanned texts and provides a first step towards surveying the sheer mass of materials now available to researchers, be that for the purpose of transcription, keyword search or corpus text mining. The crux of OCR processing, then as now, is the question of quality control: as a technology developed to precisely skirt the necessity to assess or transcribe a text word for word, selective close reading is still necessary to understand whether the output is good enough for the purpose envisaged.
During the past ten years, the technology has progressed rapidly with the advent of new forms of artificial intelligence. A decisive departure from rule-based systems, neural network-based OCR systems were made available around 2016; the quality of OCR results improved significantly, and a much greater range of documents can now be processed. Most recently, generative AI has provided a new avenue for experimentation with various sources, but also new kinds of errors.
The following talk will take a look at these developments with a particular view to the strategies developed for quality control of OCR output – and by extension, the quality of the quality control.
Elaine Lai (Stanford University), AI Ethics and the Humanities: A Perspective from Buddhist Studies
In universities across the United States, tech majors are soaring, while humanities majors have been plunging. These trends correlate with reports by PwC, MicKinsey, and the World Economic Forum, where it is estimated that AI will fundamentally transform the workforce by 2050, with up to 60% of jobs requiring significant adaptation due to AI. The humanities and digital humanities are already being impacted by these changes; as such, it is imperative for humanities scholars to critically consider how to ethically integrate AI–as a methodology or a topic of research–in order to meaningfully connect with the rest of the world.
In this paper, I consider different ethical questions around the development and adaptation of AI. Beginning with my field of Buddhist studies, I review the various ways that AI has been adapted for religious and academic purposes–from creating spiritual chatbots to emerging translation technologies. I highlight the benefits and potential risks in adopting these technologies. The second part of my paper introduces a few emerging AI tools that humanities scholars in various fields might consider testing in their own projects. In the third part of my paper, I zoom out to larger conversations and fears around AI. Here, I draw from my experience as an instructor for Stanford undergraduate students, most of whom are CS majors. I propose several ways that humanities scholars can contribute to the conversation on AI ethics more generally, so as to steer AI’s development in more inclusive ways that consider ethical obligations beyond a narrow definition of the human.
Participants
Lin Du (National University of Singapore)
From Information to Metaphor: Tracking Photographic Editing in Chinese Wartime Magazines Through Digital Historical Forensics
This study introduces Digital Historical Forensics (DHF) as an innovative methodological framework that integrates traditional media research with computer vision to reconstruct the editorial strategies of the Jinchaji Pictorial, a WWII-era photographic publication of the Chinese Communist Party (CCP). DHF addresses challenges posed by the vast corpus of 20th-century print media and the absence of textual records documenting editorial processes. By resituating photographs within their original publishing contexts and tracing their circulation, this research examines intertextuality both in the iterative reuse of images across publications and their relationship with adjacent texts. Editorial interventions—cropping, altering, retouching, repurposing, recaptioning, and textual reinforcement—transformed photographs from informational records into metaphors of socialist propaganda, aligning with shifting political agendas. The analysis highlights a preference for human subjects and textual elements, reflecting socialist realism’s emphasis on visualizing social relationships, crafting visual evidence, and ensuring textual authority over interpretation. The insistence on textual intervention—whether embedded within photographs or imposed editorially—not only preemptively determined meaning but also undermined photography’s documentary autonomy. This study demonstrates how computational methods extend visual studies beyond static content analysis, mapping hidden networks of image dissemination and transformation to reframe historical imagery as a dynamic site of contested meaning rather than fixed representations.
Gal Forer (UC Berkeley School of Law; Harvard Kennedy School)
Towards a “Beijing Effect” in AI Governance?
Building on the “Brussels Effect” theory, a term coined by Professor Anu Bradford, I will discuss the three major AI models, namely of the EU, the US, and China. I will then focus on the pros and cons of China’s AI ecosystem, expanding on Kai-Fu Lee’s comprehensive approach in his book “AI Superpowers”. Lee believes that AI is based on abundant data, hungry entrepreneurs, AI scientists, and AI-friendly policy environment. In this paper, I would like to propose that considering China’s recent regulatory efforts in AI, a “de jure” Beijing Effect might be possible for jurisdictions with significant trade relations with China, such as some developing nations in Africa, Latin America, and Southeast Asia. I think that Beijing will struggle to create a de facto influence, but there is still a possibility for a “de facto” Beijing Effect through Chinese firm exports. To examine this potential effect, I will use Kenya as a case study building on China-Africa relations and Chinese investments in Kenya’s ICT sector as part of the Belt and Road and Digital Silk Road initiatives.
Stephen Forrest (University of Massachusetts Amherst)
Labor, Ethics, and Automated Text Recognition in the Classical Japanese Classroom
The remarkable development in recent years of text recognition software for premodern Japanese — especially the openly accessible Miwo and NDLKotenOCR — has opened up new study and research opportunities for students starting out reading premodern texts. Undergraduates and graduates alike appreciate the assistance in accessing authentic editions of Edo-period printed texts (many of which are not available in modern print form), and they then often feel empowered to pursue their explorations beyond an initial course. We teachers know there is much to study in the field, and a relative shortage of students, so this must be good news.
However, the new methods have brought with them new problems and questions. How can beginner students recognize when the OCR is wrong? How can they solve the sections of text an OCR read incorrectly (or skips, as sometimes happens)? What is the appropriate timeline for introducing OCRs into paleographical pedagogy? Although this is not the case at present, could the software in future make training human readers unnecessary? And what if the OCR systems are put behind paywalls, even though some of them have been trained on databases made with a great deal of volunteer labor over many years?
These are not new questions, but in this paper I reflect on different types of classroom situations where I have used OCR software in the past couple of years, and draw some tentative conclusions that might point towards a productive coexistence.
Hong-Yu Hsien (Independent)
AI Bias-cancelling for the interpretation of a Late Qing Dynasty text
This paper describes a series of AI experiments performed on a Qing Dynasty examination essay by Wang Huiluan 汪噦鸞 in 1902. The essay, “The Grand Minister uses the method of land accounting to categorize the living things of the five types of land” 大司徒以土會之法辨五地之物生論 discusses how land in China should be categorized for agricultural production.
First, I read and translated Wang’s essay, comparing the text with quotations from the Rites of Zhou. Based on that comparison, I concluded that Wang questioned the Rites of Zhou. Wang transformed from a Qing official into a revolutionary, and I sought traces of his revolutionary future in his examination essays. Next, I treated my conclusions as hypotheses that I could double-check with AI, by prompting ChatGPT to analyze the essay text the same way I did. This method could be called bias-cancelling. Humans have biases; AIs have biases. Can they cancel each other out? In this case, the AI convinced me that Wang was reaffirming the Rites of Zhou, not doubting it. Finally, I checked secondary scholarship on the imperial examinations, which supported the AI’s conclusion.
Overall, ChatGPT was a useful tool for making me question a mistaken hypothesis. However, ChatGPT was also likely to magnify the original mistake. At first, ChatGPT denied that the essay was revolutionary, but as I “refined” my queries, ChatGPT’s answers began to cater toward my starting assumption that Wang was a revolutionary. This serves as a warning that human judgement is still the crucial final step of bias-cancelling.
Moeka Kiyohara (Kobe University)
Suggesting a language-specific gender perspective into AI translation research
The Japanese language features speech styles that are reserved for use by particular gender groups. While real-life Japanese speakers do not commonly employ these ‘gendered’ speech styles, these are prominently seen in fictional works and translated texts (Nakamura 2005, Furukawa 2024). Nakamura identified women’s language in Japanese (onnna kotoba) as an ideological construct historically shaped through discourse, highlighting the contribution of translated speech in sustaining gender ideologies.
But what if we apply this concept to AI translation? Do texts translated by AI also have the potential to contribute to the sustainment of gender ideologies? As discussed in Lakoff (1973), women’s speech in English is characterized by pragmatic features such as the more frequent use of hedges, tag questions, and modal constructions—elements that are relatively unlikely to appear unexpectedly in an AI translation process. Meanwhile, the representation of gender differences in Japanese also extends to lexical and grammatical elements, such as first-person pronouns and sentence-final particles. This highlights the challenge of AI translation, where it is up to the AI system to determine whether to incorporate gendered language in the translated text, as these linguistic components would be ‘added’ by AI during the translation process.
This study aims to summarize the recent research into AI translation and gender bias in other languages and to address the ethical concerns surrounding AI’s role in perpetuating gender ideologies through translation, specifically examining the English-to-Japanese language pair. By offering a language-specific perspective of gender, this presentation aims to expand the growing body of research on AI translation.
Prince Kumar (Université de Franche-Comté)
LLM-Assisted Geospatial Mapping of 17th-Century European Travel Writings on India
The geospatial analysis of historical travelogues typically requires multiple specialized computational steps: OCR/HTR processing of source materials, named entity recognition for locations, and geocoding to produce maps. While existing tools address each step individually, this paper explores the potential of large language models (LLMs) as an all-encompassing solution for such studies. Traditional methods face several key challenges like linguistic variability in ancient texts, interpretation of implicit spatial references, and handling narrative complexity in subjective accounts.
LLMs offer distinct advantages for this kind of study due to their promising ability to extract text from images and other documents and information retrieval from unstructured textual content subsequently. Their contextual understanding can potentially handle archaic spellings and multilingual content more robustly than dictionary-based methods. The models’ ability to interpret implicit references may provide solutions to ambiguous toponyms as opposed to conventional fuzzy matching approaches. Additionally, LLMs may better distinguish between substantive geographical content and narrative digressions – a crucial capability given travel writers’ tendency to combine factual observations with personal commentary.
This study develops a methodology following FAIR principles, using 17th-century European travel writings about India as a case study. While acknowledging that human verification remains essential, we examine how LLMs might streamline the entire analytical pipeline from text to map. The paper focuses on theoretical framework and methodological potential rather than specific results, providing a foundation for future implementations while critically assessing the technology’s limitations for historical geographic research.
Shuohong Lyu (University of East Anglia)
Artificial Intelligence Hegemony: How Researcher’s Independent and Critical Thinking Capabilities Are Eroded
The concept of Artificial Intelligence Hegemony (AIH) has yet to be explicitly theorised within academic discourse. In the first part, this paper starts by explaining what is hegemony and introduces AIH as an unprecedented, unusual, undetectable hegemonic process increasingly shaped by algorithmic systems with minimal human intervention. It discusses how AIH influences ideological reproduction, embedding bias into social perception, cultural narratives, and knowledge dissemination.
The second part of the paper explores the impact of AIH on research practices. AI-generated and retrieved content may promote algorithmically favoured perspectives, reducing critical engagement and epistemic diversity. In particular, AI tools can degrade the independent thinking capability of researchers, especially in comparative cultural and linguistic studies, which usually require the researcher to spot differences based on personal experiences and feelings before investigating them academically. For instance, in cross-cultural privacy studies, semantic nuances are not able to be identified by AI tools, such as the moral connotation of yin’si (privacy) in Mandarin Chinese, as the word yin’si (privacy) in Mandarin Chinese includes a strong emphasis on shame, while AI tool’s interpretation of the term is based on its English equivalent. In this process, the hegemonic nature of AI is reflected.
This paper concludes by arguing that the effect of AIH presents epistemological challenges to academic inquiry. It not only erodes academic integrity but also diminishes the researcher’s capacity for independent and critical thought, especially in fields that rely on human interpretation.
Peter Francis Smith (University of Oxford)
Handwritten Text Recognition of Chinese Manuscripts: Complexities, Human Input, and Machine Learning
This presentation will focus on Handwritten Text Recognition (HTR) in relation to the complexities of processing manuscripts, the human input required, and the iterative process of machine learning. It will explore the great potential of this work and some innevitable limitations. Details are based on a Britsh Library project aiming toward the mass transcription of Dunhuang manuscripts. The process uses models developed by CHAT (Chinese Historical documents Automatic Transcription) for binarisation, segmentation, and transcription – three key stages of the HTR process.
The first step, image binarisation, simplifies a manuscript image by removing extraneous details while preserving key textual information. Human evaluation of the process is essential to refining the binarisation model so that it retains character detail while handling variations in ink colour and paper condition. Layout segmentation defines text regions within a manuscript, distinguishing between body text, annotations, and images. Irregular layouts and varying line structures must be identified and the correct reading order established. Text recognition is the final stage. Pre-existing models, already trained on historical Chinese scripts, are effective in processing well-formed characters but the variation of scripts and character forms remain an ongoing challenge.
These advancements in HTR underscore its growing viability for large-scale manuscript digitisation. While challenges remain, continued refinement of these models, supported by human expertise, are strengthening their capacity to handle diverse historical materials. This will ensure broader accessibility of handwritten texts and expanded potential for digital humanites approaches to research.
Tiziana Pasciuto (University of Turin)
From Historical Encounters to Digital Records: The REDMIX Archive of Red Sea Mixedness
The REDMIX project aims to develop a multilingual digital archive to document and explore the cultural, historical, and social mixedness of the Red Sea from the 1800s to the 2000s. The archive will be built using open-source software and will collect a wide range of different digitized materials. It will also feature digital tools for visualizing resources such as geographical data, genealogies, bibliographies. Designed to be accessible, replicable, sustainable, and functional, the archive will serve users across diverse contexts, including those affected by the digital divide.
The project faces significant challenges, including ensuring interoperability between the archive and the project website, managing multilingual materials, and designing an intuitive user interface that works even with slow internet connections or on mobile devices. The choice of open file formats and standardized metadata schemas is critical for long-term preservation and interoperability. Additionally, the digitization of fragile and often hard-to-access materials requires portable, low-cost equipment while maintaining high-quality standards.
Beyond technical aspects, the project emphasizes decolonizing knowledge by contextualizing materials with historical and cultural information. It also addresses ethical considerations, such as representing traumatic events and handling sacred or sensitive materials, by obtaining consent from stakeholders.
To overcome these challenges, the project draws inspiration from existing initiatives and leverages open-source tools and Semantic Web technologies. By combining technical innovation with cultural sensitivity, REDMIX aims to create an inclusive and enduring resource for understanding the Red Sea’s complex and entangled histories.
Ziyi Qin (Beijing Foreign Studies University)
Comparing Sentiments in Yuan Zhen and Bai Juyi’s Poetry: A Computational Analysis Across Time and Space
Sentiment analysis of classical Chinese poetry has primarily relied on qualitative approaches, but existing quantitative studies remain limited and often suffer from coarse-grained sentiment classification. Although Yuan Zhen and Bai Juyi are usually categorized within the same literary school, they are mainly studied for their similarities, and comparative studies focusing on their differences are scarce. To bridge these gaps, this study constructs a database of 3,798 poems incorporating 14 metadata dimensions and a fine-grained sentiment lexicon with three primary categories and 16 subcategories. Sentiment words are identified through bidirectional matching, followed by manual verification. Specifically, this study compares the distribution of positive and negative sentiment words in the poems of the two poets across different periods and locations. The results indicate that Yuan Zhen expressed stronger negative sentiments. While both poets faced adversity, Yuan Zhen tended to dissipate sorrow, whereas Bai Juyi was more inclined to transcend it. Further analysis reveals that socio-cultural contexts, individual personalities, and life experiences simultaneously shape poetic sentiment. By analyzing the co-occurrence of the 20 most frequently used imagery words with 16 subcategories of sentiment words, the study finds that the emotional associations of the same imagery differ between the two poets, and their preferences in leveraging imagery to express similar sentiments also vary. This study enriches insights into the comparison between Yuan Zhen and Bai Juyi and achieves an efficient and fine-grained sentiment analysis of Chinese classical poetry, which can be applied to larger datasets.
Stephanie Santschi (University of Zurich) and Drew Richardson (University of California, Santa Cruz)
From Print to Place: Creating an Integrative Geolocalization Workflow using Citizen Scientists and Computer Vision AI
In early 2025, we launched the prototype for Drawing from the Crowd, a platform combining computer vision (CV) with crowdsourced human expertise to explore the relationship between Edo period (1603–1868) Japanese woodblock prints (ukiyo-e) and their topographical reality. The project geolocates viewpoints and landmarks in prints to understand the mechanisms of shared spatial imagination in late Edo period. To this end, computational image analysis (CLIP / SigLIP) first helps organize and classify the digital images by identifying recurring visual elements to create clusters according to visual characteristics like composition or content, before citizen scientists localize the selected prints using the Smapshot (HEIG-VD) workflow.
The integration between CV and citizen science is central to reconstructing the multiple spatial realities in these historical landscapes. Both approaches share important characteristics: they process datasets exceeding individual capacity, expand research participation, and distribute analytical tasks across contributors. However, they also face parallel challenges in data verification and quality control.
This paper presents initial findings from Drawing from the Crowd, examining how combining CV and citizen science ensures accuracy and reliability of generated data while contextualizing the findings within Humanities’ ambiguous “ground truths.” It argues that while these methods independently face validation issues, their integration creates a productive workflow where machine pattern recognition and human contextual understanding enhance insights into early-modern landscape images and their perception by producers and consumers.
Enes Yılandiloğlu (University of Helsinki)
Mapping the Orient: Computational Approaches to Eighteenth-Century Travel Writing
This study analyzes the representation of the Orient in eighteenth-century English travel writing through a dual approach combining computational methods and literary analysis. The OCR’d texts from Eighteenth-Century Collections Online (ECCO) (Gale, n.d.) are leveraged as a source of data. The travel books are identified via both metadata and keyword searches. This process yields 971 English travel works from the eighteenth century. The location entities, the main focus of the study, are extracted via a transformer model (Franklin et al., 2024) for the Named Entity Recognition (NER) task. Following NER and post-processing, the location entities are georeferenced via Pleiades (Gillies, Turner, & Elliott, 2025) and are visualized using interactive maps enriched with bibliographic metadata, enabling spatial and textual analysis. The visualization goes beyond basic GIS mapping and engages with the deep mapping paradigm within digital humanities (Bodenhamer, Corrigan, & Harris, 2015). Three dominant narrative modes are identified via the visualization: (I) depictions of the Orient as uncanny, obscene, and unlawful—linked to Orientalist discourse; (II) realistic and empirical portrayals, influenced by Enlightenment rationality and negotiation with the East; and (III) intentional distortion of reality aimed at narrative embellishment, considered through the lenses of Romanticism, the travel lie (Batten, 1978), and reception theory (Jauss, 1982). Overall, the study explores how computational methods leveraging big data can be applied to reveal the dynamics of the eighteenth century travel writing.”
The Digital Orientalist 