Mandarin is replacing Cantonese. Offbeat AI fights back as Big Tech looks away

Preserving Cantonese has been challenging due to the dominance of Mandarin, limited learning resources and a lack of a standard written form. With a declining number of young learners, the language faces an uncertain future. Artificial intelligence (AI) – seen by some as an existential threat to humanity – may become the hope for saving the language, and many others, along with the distinct cultures they embody. This is the mission of Hong Kong-based deep-tech company Votee AI: to use large language models (LLMs) to preserve languages, especially those overlooked by tech giants. Leo Ma, Votee’s chief scientist for the Asia-Pacific, said in an interview that AI could create living records of many languages beyond English – the primary language for most LLMs today. “While mainstream AI models excel in English, they remain ‘functionally illiterate’ for 99 per cent of the world’s languages,” Ma said, adding that this limited AI access for the billions of non-English speakers. Unlike Mandarin, English or Spanish, which have vast troves of digital text for AI training, Cantonese lacks written text that accurately reflects how it is spoken. “Cantonese is a ‘low-resource language’ because there is not much written text that we can use for training,” Ma said. Votee AI, which says it is the first company to open source a Cantonese LLM, went through the exacting challenges of recording the language spoken by 85 million people worldwide. “Neglected languages should not be gated. Open source is a critical path to ensuring linguistic diversity survives in the age of AI,” Ma said. “Neglected languages are constantly disappearing. “For us in the technology industry, we look to the future – but at the same time, we hope to preserve our past, which is the foundation of the present.” To collect training data, the team built a digital dictionary by gathering data from the internet, such as on social media and forums, and partners such as universities, then annotating the cultural context of words. “Words can have many meanings in different contexts. Every word, corresponding to a specific context, will be represented by its own independent item in the data set we create,” Ma said. For example, the word for “water” is often used to refer to money, in addition to its literal meaning as a substance, and has many related usages in Cantonese. According to Ma, this level of deep understanding is crucial because models trained on other languages cannot capture the unique idioms, cultural nuances and identity of Cantonese speakers through translation. “This is not merely a matter of inaccurate translation; it is a problem of cultural misrepresentation,” he said. Linguistic precision is also important for the real-world application of LLMs. For models deployed in sectors such as finance, medicine, law and government, having “half-fluent foreign interns” with 50 per cent accuracy represented a liability, Ma said. To ensure models reflected the rich cultural nuances, a Votee AI researcher created a benchmark with peers, including those from The Education University of Hong Kong, to elevate the language proficiency and cultural knowledge, as well as reasoning and problem-solving of Cantonese LLMs in the context of Hong Kong. “While proprietary models generally outperform open-weight models, significant limitations remain in handling Cantonese-specific linguistic and cultural knowledge, highlighting the need for more targeted training data and evaluation methods,” the team wrote in a paper published by the Association for Computational Linguistics last year. “Addressing the segregation of regional and linguistic knowledge is crucial for developing culturally and linguistically adaptive LLMs. This issue extends beyond Cantonese to other under-represented language communities,” they added. Ma said the initial data cleansing for the Cantonese data set took the team four to five months, but the company had built a system using autonomous agents that shortened the process to two weeks. “Training a new model used to be a months-long process involving data cleansing, persona creation and data synthesis. We needed data representing various speech patterns – for example, how young trend hoppers speak, the language used by lawyers or the way celebrities talk.” The new AI-powered process could be replicated for other low-resource languages in Southeast Asia and Africa, with the company setting up an AI research lab in Toronto and regional hubs in Kuala Lumpur and Ho Chi Minh City, Ma said. “We are currently scaling our ‘foundry’ model to support other Southeast Asian and emerging market dialects. This platform is not just for one or two languages; its goal is to dramatically lower the technical and cost barriers to building high-quality, native-language models,” he added. Around 1,500 languages are endangered globally, according to a 2021 study led by the Australian National University. They are at high risk of no longer being spoken by the end of this century. Ma pointed to the centuries-long inability to understand and pronounce Egyptian hieroglyphs to illustrate the importance of preserving languages. The meaning of the language remained a mystery until the Rosetta Stone, discovered in 1799, finally allowed scholars to translate Egyptian hieroglyphs thanks to its inscriptions showing the same content in three scripts, including Ancient Greek, which served as the translation guide. We believe Al’s role is not only to create the future but also to safeguard the past Leo Ma, chief scientist for Asia-Pacific, Votee AI “When a language disappears – how can we know its pronunciation and the true meaning of its words? No one can be 100 per cent certain, so parts of that culture are likely lost,” Ma said. “To understand the true nature of a culture, we must study its language and texts.” Despite automation, Ma said humans remained central to ensuring the performance of new models and passing down knowledge. For building every new language library, linguists and native speakers would help refine the grammar, vocabulary and phonetics that AI alone cannot grasp. The models built can help people learn languages. “The passing down of languages in communities is incredibly important,” Ma said. “If there is nobody left to speak a language, the language is already lost. We believe Al’s role is not only to create the future but also to safeguard the past.”

Mandarin is replacing Cantonese. Offbeat AI fights back as Big Tech looks away

.css-1c6uqr6{color:inherit;font-weight:inherit;font-size:inherit;font-family:inherit;line-height:inherit;overflow-wrap:break-word;}Hong Kong-based tech firm Votee AI aims to replicate its model for other low-resource languages in Southeast Asia and Africa

Hong Kong-based tech firm Votee AI aims to replicate its model for other low-resource languages in Southeast Asia and Africa