Daejeon, July 27–31
Since the seventeenth-century, Vietnamese romanized scripts have undergone multiple and various phases of lexical standardization and enrichment (Rageau 1984; Pham 2018). The following centuries brought significant changes as local print culture began to flourish across Southeast Asia, especially in places such as Calcutta, Bangkok, and Singapore (Rageau 1984; Pham 2018). The following centuries brought significant changes as local print culture began to flourish across Southeast Asia, especially in places such as Calcutta, Bangkok, and Singapore (Duverdier 1980; Proudfoot 1993; Ghosh 2003; D’Souza 2004). The emerging urban areas of Hanoi and Saigon in French Indochina operated the same turn by the end of the nineteenth century (Duverdier 1980; Proudfoot 1993; Ghosh 2003; D’Souza 2004). The emerging urban areas of Hanoi and Saigon in French Indochina operated the same turn by the end of the nineteenth century (V. T. Huỳnh 1971; McHale 2004; Cao 2024). In this rapidly changing environment, political and intellectual efforts sought to articulate and disseminate new vocabulary through local and transnational circuits of knowledge (V. T. Huỳnh 1971; McHale 2004; Cao 2024). In this rapidly changing environment, political and intellectual efforts sought to articulate and disseminate new vocabulary through local and transnational circuits of knowledge (Meade et al. 2024). (Meade et al. 2024).
In 1895, the first volume of the Đại Nam quốc âm tự vị – Dictionnaire de la langue annamite, compiled by Huỳnh Tịnh Của, was published in Saigon. This work is one of the earliest modern dictionaries of romanized Vietnamese (quốc ngữ) produced and edited by a Cochinchinese locutor. From 1900 onwards, hundreds of didactic and teaching books for , compiled by Huỳnh Tịnh Của, was published in Saigon. This work is one of the earliest modern dictionaries of romanized Vietnamese (quốc ngữ) produced and edited by a Cochinchinese locutor. From 1900 onwards, hundreds of didactic and teaching books for quốc ngữ circulated in Indochina, favored by the education policy implemented by the French administration. In 1917, with the support of the Directorate of Political Affairs of Indochina, Vietnamese intellectuals published Nam Phong (The Eastern Wind), a trilingual periodical which incorporated lists of Chinese, Vietnamese, and French vocabularies in its early volumes Mostly edited by Phạm Quỳnh, a well-known scholar, Nam Phong engaged with a broad range of scientific, literary, and cultural topics. Nam Phong has since been digitized and made publicly accessible through the Internet Archive, enabling large-scale computational analysis.
The computational detection of neologisms and lexical semantic change has been explored in several NLP initiatives (Schlechtweg et al. 2020). More recently, word embedding approaches have also been applied to colonial discourse in digitized historical newspapers (Schlechtweg et al. 2020). More recently, word embedding approaches have also been applied to colonial discourse in digitized historical newspapers (Zhu et al. 2025). However, such approaches remain largely untested on non-European multilingual corpora, where orthographic instability, script diversity, and the absence of annotated resources pose particular challenges.(Zhu et al. 2025). However, such approaches remain largely untested on non-European multilingual corpora, where orthographic instability, script diversity, and the absence of annotated resources pose particular challenges.
This short paper introduces Rag’it, a collaborative project undertaken during a 2025 digital residency funded by the DISTAM DIgital STudies Africa, Asia, Middle East.
The pipeline relies on a vector space model in which multilingual text segments are transformed into high-dimensional embeddings. By mapping text segments into a shared semantic space, measures such as cosine similarity allow for the querying and retrieval of segments likely to contain neologisms and their variants (Camacho-Collados and Pilehvar 2018; Liu et al. 2020). For this project, we use (Camacho-Collados and Pilehvar 2018; Liu et al. 2020). For this project, we use paraphrase-multilingual-MiniLM-L12-v2, a sentence-transformer model producing 384-dimensional dense vectors, indexed with FAISS for efficient nearest-neighbor search. This model supports Vietnamese, French, and English, and handles code-switching across mixed-language documents – a frequent feature of the colonial corpus under study.
Although embeddings can provide interesting results in semantic and cross-lingual search, it remains difficult to assess the relevance of retrieved material for specific research goals (Michail et al. 2025). (Michail et al. 2025). Rag’It intends to enhance the capability of multilingual embeddings by implementing a RAG structure using Large Language Models (LLMs), which aims to generate contextualized summaries anchored in source texts and their metadata. This structure allows researchers to evaluate retrieved segments both quantitatively and in terms of their contextual relevance to the research question.
The project is structured around two main stages. The first stage involves the optical character recognition (OCR) of the Nam Phong trilingual lexicons. This corpus offers a rich material for identifying and analyzing neologisms in the early twentieth-century Vietnamese intellectual networks. Another corpus of translated philosophical texts, published by Tân Việt editions in Hanoi in the early 1940s, has also been OCRed This corpus includes Vietnamese translations of foundational works by popular Western thinkers such as Bergson, Kant, Nietzsche, or Freud. It has been digitized by the Bibliothèque nationale de France and is now available on Gallica. This step was made possible through the use of Calfa Vision, a web-based annotation tool for both images and texts. https://vision.calfa.fr
At its current stage, Rag’it has OCRed and published the Nam Phong trilingual lexicon dataset (Cao and Saudemont 2026). The dataset covers issues 1 to 10, 13, and 14 of (Cao and Saudemont 2026). The dataset covers issues 1 to 10, 13, and 14 of Nam Phong periodicals, representing 1,512 lexical entries. Each entry is structured across four fields corresponding to the journal’s trilingual lexicon: the romanized Vietnamese term (vi), its Han-Nôm equivalent (vi-hani), a Vietnamese-language definition (vi-def), and a French gloss (fr). Transcription was performed using a fine-tuned Vision Language Model (Qwen VL 2.5), followed by manual correction, and evaluated using KAMI (CER: 11.48%, Word Accuracy: 87.94%). This alignment across three linguistic registers constitutes the primary query interface for the RAG pipeline.
An experimental RAG structure has also been built within the project. The pipeline uses LangChain as the orchestration layer, coupled with locally hosted large language models (LLMs) including Mistral and PhoGPT, to handle natural language queries in Vietnamese, French, and English. We have defined a set of multilingual queries to reflect the source materials and to test the capacity of the RAG to retrieve coherent and contextually grounded results. Our evaluation criteria focused on the extent to which the RAG output included both the text segments containing potential neologistic forms and the corresponding metadata that would allow researchers to locate and interpret these segments within their original publications.
The RAG pipeline is structured around two complementary query modes. The first operates on attested lexical forms: given a term from the structured lexical dataset in romanized Vietnamese (vi), Han-Nôm script (vi-hani), or French (fr), the system retrieves passages in which that term appears and returns it alongside its semantic context. The second operates on definitional proximity: using the Vietnamese definition field (vi-def) as a query, the system identifies passages containing terms semantically close to that definition, even when the exact form is absent. This mode is designed to trace synonyms and orthographic variants that are particularly common in early twentieth-century romanized Vietnamese.
At its current stage, the pipeline has been tested on general multilingual queries across both corpora, yielding relatively coherent and contextually grounded results. The two lexicon-based query modes described above are currently under development. Preliminary observations nonetheless highlight several challenges inherent to this line of experimentation. Ambiguities in neologistic usage and orthographic variation in historical Vietnamese texts contribute to retrieval noise and hallucination. Building a robust evaluation framework to systematically test the two lexicon-based query modes constitutes the next phase of the project.
Rag’it mobilises computational methods to address enduring questions about lexical innovation, multilingual textuality, and historical semantic change in Vietnamese print culture. The project offers a workflow for exploring neologisms in multilingual and unannotated historical corpora, which could be scaled to larger corpora of lexicons and dictionaries.
The preliminary and exploratory stage has successfully generated clean and structured dataset of the Nam Phong trilingual lexicons. It also reveals the necessity to create a coherent and robust protocol for evaluating the RAG behaviour, which could require more human and technological resources. The challenge of correctly and consistently identifying neologisms or words that are absent from the lexicon and their usage still requires more testing and evaluation. This ultimately underscores the need to create transparent documentation to allow continuous development of adaptable, context-aware models for Asian studies and the digital humanities.