laitimes

Wang Shen: New Possibilities for Generative Artificial Intelligence and Historical Research|202407-106 (No. 2789)

Source: China Social Science Network

Generative AI and New Possibilities for Historical Research

Text / Wang Shen

Assistant Researcher, Institute of Ancient History, Chinese Academy of Social Sciences

Wang Shen: New Possibilities for Generative Artificial Intelligence and Historical Research|202407-106 (No. 2789)

Generative AI technology has advanced rapidly in recent years and has had a profound impact on both the technical and business sectors. New technologies and related products bring new technological experiences such as large language models, pre-training, and interactive responses to historians. Although it is still in the exploratory stage in the field of history, it is likely to greatly promote the upgrading of the construction and application of historical databases in the future, and promote the positive transformation of historical topics. While new technologies bring new possibilities to historical research, they will also generate technical risks and challenges, which deserve to be taken seriously by scholars.

Promote the construction of historical databases and the upgrading of application methods

Historians have been using databases to carry out research for decades, and obtaining historical materials through keyword search has become the basic way of historical research, and database construction is in the ascendant. Although the nature of the historical materials covered by the databases and the research groups they face are different, the types of most databases are very similar: the builders integrate and enter historical materials such as ancient books, inscriptions, and folk documents in the form of texts or pictures and establish catalogs; Researchers extract the required historical information through browsing or keyword search, manually select the results, and carry out follow-up research. Due to the relatively simple reading and retrieval methods, the quality of database use depends almost exclusively on whether the researcher can set the appropriate keywords. The database only plays a passive role in providing text and basic search functions, and most of the data and logical relationships that are difficult to detect through keyword search are still in a "sleeping" state. Replacing quality with quantity, subscribing with accumulation, and severing the connection between keywords and context and even historical background are the critical criticisms of retrieval research.

Integrating generative AI products into a database and pre-training them for a period of time before the database is made public (e.g., allowing them to learn research papers in the relevant fields) can greatly expand the analysis capabilities of the database, so that it can output logical answers, and the analysis capabilities will become more intelligent as the number of uses increases. This means that the database has enhanced its initiative in historical research, and can better discover the materials that truly meet the needs of researchers from massive data, understand the relationship between researchers' instructions and data content, and provide researchers with more personalized and in-depth analysis results.

Scholars can use the new database as a search tool for general keywords, and they can also use the database as a research assistant through questions and answers, which is revolutionary for historical research. One of the major features of generative AI is that it greatly enhances the ability of models to judge and analyze natural language. On the one hand, the text in the database can be understood by the model in some way; On the other hand, researchers can send instructions to the database through natural language instead of programming language, and continue to ask questions to continuously clarify requirements, which greatly reduces the threshold for the application of new technologies. When interacting with the database, researchers need to describe the demands in as much detail and accuracy as possible, so as to obtain hierarchical and systematic answers. The Q&A process can also help the AI learn from the user's ideas and improve the analysis conclusions for better feedback. Interactive Q&A transforms the process of historical researchers searching databases to obtain information into a comprehensive research scenario with the "three-in-one" of information extraction, academic discussion and logical examination, and human-computer interactive academic research is likely to be realized.

Promote the positive transformation of historiography issues

Generative AI can collect and analyze massive amounts of text data in its entirety and generate responses in natural language in a short period of time. When the model is trained to a certain extent, much of the research and interpretation work can be done by artificial intelligence, and the speed and accuracy are likely to greatly surpass that of anthropologists. For example, some domestic and foreign research teams have begun to use machine learning technology to deconjugate and interpret oracle bone inscriptions and papyrus documents. The professional barriers of related work will be broken through by artificial intelligence, and the analysis of anthropologists is no longer irreplaceable. Although this does not mean that scholars can give up their judgment and become "hands-off", the weight of critical issues will inevitably decline, and the "scissors and paste" methods of argumentation such as listing historical materials and summarizing conclusions will also be eliminated at an accelerated pace. There is an urgent need for scholars to actively change the topic and promote the development of historical studies to a higher level of speculation under the stimulation of technology.

Focusing on the generation of historical texts is a research path that fully demonstrates the basic skills and critical thinking ability of scholars. As it stands, generative AI products are not only highly dependent on the number of corpora, but also weak in distinguishing between authentic and indistinguishable historical narratives and complex historical writing intentions, and almost indiscriminately regard the texts in the database as "true". In contrast, in the field of medieval history and Liaojin history, which are not rich in historical materials, researchers have made breakthroughs through historical writing, historical source identification and analysis, and used limited historical materials to make breakthroughs, which have refreshed people's understanding of the formation process of historical records. Scholars no longer easily assert the authenticity of historical materials or only value the authenticity of historical materials, but focus on analyzing the complex factors behind the formation of texts.

The importance of setting topics will be strengthened. While the database initiative for access to generative AI has increased, the starting point and assumptions of research must still be initiated by the researcher. Since new technologies can liberate human beings in basic work, and the rapid growth of computing power in computer hardware has greatly reduced the time cost of theoretical trial and error, the value of scholars will be more reflected in how to initiate a research with high academic significance. This requires scholars to ask questions and construct theories in a more rigorous, comprehensive, and logical way.

Digital humanities will become a strong growth point for the interdisciplinary development of historical research, and there may be breakthroughs in the presentation and use of historical materials. Updating historiography databases is not something that can be achieved simply by connecting to generative AI. Due to the huge differences between historical and contemporary texts, database builders must transform historical texts into a style suitable for AI analysis, which requires the introduction of advanced digital humanities tools. Promising progress has been made in the development of tools such as ByteDance's "Reading Ancient Books" platform, which uses optical character recognition (OCR) and automated algorithms to convert images of ancient books into text and automatically punctuate them. The team of Melissa Dell, a professor at Harvard University, · specializes in extracting complex and irregularly arranged historical texts, and has broad prospects in automatic identification of archives and folk documents. The Japan Center for Open Data Utilization in the Humanities (CODH) uses machine learning technology to read glyphs in ancient Japan books and develop mobile apps with character recognition capabilities. The value of using similar techniques to process historical texts is clearly not limited to database construction.

Generate technical risks and challenges

The issue of academic ethics in introducing generative AI into academic research is a cliché. It is difficult to truly solve the problem by relying on the efforts of the academic community alone, and it depends on the "multi-pronged approach" of legal construction, government supervision, and cooperation of market entities. In contrast, it is the technical risks and challenges that new technologies may pose that are the primary issues that historians need to face. After all, technical problems directly affect "use", and ethical dilemmas can only arise after "use".

The results of AI output are not always objective and neutral, but are affected by factors such as the amount of data, the frequency of training, and the training method, and it will lead astray if used without discrimination. AI products that use the same model may still give different answers to the same question because they are connected to databases with different amounts and contents, and are trained in different ways. In addition, the researchers' thinking tendencies, research paths, and subjective motivations will intentionally or unconsciously cause the output to be artificially shaped, resulting in misleading "new history writing" that is difficult to distinguish. A recent paper by researchers at the University of Oxford· Mrinank Sharma and Anthropic, "Understanding Flattery in Language Models," shows that flattery is widespread in AI models based on human feedback, as users prefer flattering responses.

Sharma et al.'s paper uses sophisticated methods to interact with a variety of AI models to reach these conclusions. This brings us to another challenge posed by generative AI: the black-box processing of information makes it difficult to verify the results of AI outputs. This not only does not conform to current academic norms, but also makes it difficult for researchers to test the results. Manual verification is still useful in the face of simple logical deduction or a small number of historical data collation results, but it is powerless in judging the black box processing of massive data by artificial intelligence. Even if the authors submit the AI-enabled part of the study to the editorial office or publisher of a journal or publisher as an attachment, how to review this lengthy and unfamiliar text of digital information can be challenging for reviewers and editors in the field of history. It is not yet known to what extent people will be able to intellectually accept the involvement of generative AI in historical research.

In short, we can certainly choose to use new technologies to retrieve historical materials in the traditional way, or even refuse to use them for academic research. However, while avoiding risks and challenges, we have also missed a good opportunity to make leapfrog progress in this ancient discipline and strengthen dialogue and cooperation with emerging disciplines. The progress of related technology in other fields has fully demonstrated its great potential, and how to apply it more appropriately to historical research is worthy of continuous exploration by the academic community. Every great technological innovation has broken the existing relationship between humans and machines, and between humans and algorithms. But as the title of the book of Norbert Wiener, the father of cybernetics, says, "man has a useful man".

(This article is the interim result of the National Social Science Foundation project "Research on the Construction of Monetary and National Fiscal System in the Song Dynasty" (22CZS024))

1. Song Dynasty History Research Information 1

E-mail: [email protected]

Read on