The Ministry of Digital Development held a press conference today (15th) on "Taiwan's Sovereign AI Training Corpus Launches Private Corpus Collection" to officially promote the private corpus collection plan. (Photo by reporter Qiu Qiaozhen)
[Reporter Qiu Qiaozhen/Taipei Report] Generative AI is developing rapidly. One of the keys to whether the model can truly understand Taiwanese culture and viewpoints is the training corpus. Wang Fuzhongjin (15), Director of the Data Innovation Department of the Ministry of Digital Development, said that since the "Taiwan Sovereign AI Training Corpus" was launched in December last year, the number of data sets has increased from about 2,000 to more than 5,000, the number of Tokens has increased from 600 million to 2.1 billion, and the size of the corpus has more than tripled. As government data accumulates to a certain extent, the next stage will further expand the collection of private corpus to include high-quality original content from more publishers and creators.
Wang Fuzhong said that as AI develops faster and becomes more powerful, the outside world has begun to pay attention to what data AI uses for learning and training, and how well AI understands different languages and cultures. Therefore, all countries attach great importance to autonomous, safe, reliable and resilient AI capabilities.
Please read on...
One of the purposes of the Ministry of Digital Development's establishment of the "Taiwan Sovereign AI Training Corpus" is to hope that when faced with a work, AI can not only understand the content through external comments or a few summaries, but also more fully understand the core ideas expressed in the work, and further understand Taiwanese culture and viewpoints.
Wang Fuzhong pointed out that since the launch of Taiwan’s sovereign AI training corpus in December last year, it has adopted the strategy of “first the government, then the private sector; first the central government, then the local governments”. The number of data sets has increased from about 2,000 to 5,000, and the number of tokens has increased from 600 million to 2.1 billion. The overall corpus size has more than tripled.
The current corpus content covers folklore, language, biology, tourism, publications, law, art, culture, economy, geography, transportation, medical care, history and other fields, and also includes local languages. The Digital Development Department hopes to continue to expand the corpus to provide applications needed for the development of generative AI industry and technical teams, making AI more relevant to Taiwanese culture and content in different fields.
As government data has accumulated to a certain scale, the Ministry of Digital Development has also further expanded the source of corpus from the government to the private sector. Wang Fuzhong said that the private sector has quite a wealth of resources, especially books owned by publishing houses, which are valuable in terms of depth and quality of content. Therefore, collections are currently being conducted in different forms such as e-books, books, book content, and book introductions, hoping to include more high-quality and original works.
The folk corpus collection mechanism adopts the principle of "voluntary participation, clear authorization, and withdrawal". Clear authorization is obtained on the premise that the data provider is voluntary. The relevant high-quality corpus can be used by government agencies, civil society, academic research institutions and other data applications in the future, including model training, fine-tuning, retrieval enhanced generation (RAG) and knowledge extraction, etc., and can be further used in educational learning, administrative assistance, smart customer service and professional consultation.
As for whether the corpus provided by the outside world has been subject to prior content review, Wang Fuzhong explained that the publishing house is mainly responsible for screening and quality control of private collections. The platform itself will check the content structure, whether the data format is correct, and whether the data can be successfully used for AI training, etc., and will not interfere with the content itself.
In terms of participation methods, Wang Fuzhong said that the website has designed a three-stage process. After the data provider logs in and applies for an account, the data can be put on the shelf. After completion, the platform will check the data content, format and related information, and the overall process will be simplified as much as possible (https://taic.moda.gov.tw/content/application-guidelines?utm_source=chatgpt.com).
Grasp the economic pulse with one hand. Click here to subscribe to Free Finance Youtube Channel
