Will the real-person counter be replaced by AI? Shanghai's first government service model was launched in Xuhui
"I want to start a company, what are the procedures?" "Are there any requirements for registered capital for setting up a company?" "One of my company's shareholders is a legal entity, how do I do real-name registration?" In the comprehensive reception hall for legal person affairs on the first floor of the Xuhui Administrative Service Center, the reception staff at 18 comprehensive windows repeatedly answer the above questions almost every day.
Now, after more than three months of model training and tuning, the "Xuhui District Government Service Big Model" developed based on Tongyi's government big model technology and Xuhui District's intelligent computing platform has officially been launched. This is also the first vertical big model application of government services in Shanghai.
Focusing on the consulting business of the entire life cycle of government services, the Xuhui Government Affairs Big Model brings together 39 business-related data sets, more than 500 systematized knowledge points and more than 2,300 high-frequency consulting corpora. In business-related service scenarios such as government service processing procedures, application materials, service elements, special situation handling, etc., it can achieve more than 10 rounds of natural language consulting capabilities, and the consultation accuracy rate for high-frequency matters exceeds 90%.
What do frontline government service workers think of the increasingly "intelligent" big models? Will big models replace humans?
"The company's legal representative has committed a crime. How do I change the legal representative?" When Liu Lei typed this question into the dialog box, she decided to try asking the big model in a more colloquial way: "What should I do if the company's legal representative has been in jail?"
Liu Lei is a staff member of the comprehensive reception hall for corporate affairs in Xuhui District Administrative Service Center. Every day, she and her colleagues are responsible for the consultation and acceptance work of the 18 comprehensive windows in the hall, so they are called "Eighteen Arhats". Since Xuhui pioneered the "zero-difference" comprehensive reception window in Shanghai in 2018, each of the "Eighteen Arhats" is familiar with more than 1,000 corporate affairs "one-window comprehensive handling" matters.
Since there are skilled window staff, why do we need to develop a government affairs big model? Hu Bing, deputy director of the Xuhui District Big Data Bureau, said that Xuhui's accumulated digital foundation, the increase in the number of cases brought about by the new company law, and the emergence of new industries and new tracks have all driven Xuhui District to sort out and forge industry-specific government affairs vertical big models from massive government affairs information.
Xuhui Administrative Service Center Legal Person Affairs Comprehensive Reception Hall
"On the one hand, in the six years since the comprehensive window was opened, the processing data of more than 1,000 matters have been digitized and archived, forming a digital base. On the other hand, about 90% of the cases handled in Hall A every day are about business changes of enterprises. On the day when the single-day case volume hit a record high of 800 cases, about 85% were related to this type of matters, and the repetitiveness was prominent." Huang Wei, deputy chief of the Public Service Section of the Xuhui District Administrative Service Center, said that whether it is the new Food Registration Management Regulations issued in May or the new Company Law to be implemented on July 1, as long as the relevant laws and regulations are updated or changed, the front-line window staff must learn and master them as soon as possible.
Therefore, relying on the rapidly developing and mature big model technology to assist window personnel in improving service efficiency and accuracy of receipt, thereby reducing the number of times companies run around and waiting time, is the original intention of Xuhui to spend a lot of effort to develop vertical big models of government affairs.
At the end of last year, based on the digital base advantages of the "three centers merged" of Xuhui District Administrative Service Center, District Big Data Center and District Urban Operation Center, Xuhui launched the research and development of the government affairs big model. The technical support came from the Alibaba Cloud team. From corpus sorting, technical deployment, localized resource configuration such as servers, actual training to test launch, the development cycle was less than 4 months.
Finally, after more than 30 rounds of fine-tuning and model link optimization, and more than 20,000 rounds of dialogue testing, at the end of March this year, the Legal Person Affairs Comprehensive Acceptance Hall was finally confirmed as the first government service big model application scenario. On May 24, the business combing was completed, and the government service big model officially "started its internship".
Liu Lei uses a large model to assist in handling business at the window
Before the May Day holiday, Liu Lei started using the Xuhui government service model in the testing phase. At first, she asked the model about the most common window handling matters in the dialog box, such as "how to cancel a company." But the model's responses were very general, either irrelevant or listing all the related handling operations.
However, Liu Lei's task is not just to "chat" with the big model, but to guide the big model to output the most accurate answer through advanced questions and answers. A common training method is to give the big model more concrete keywords. For example, when asking questions, tell the big model whether to consult the company about "simple deregistration" or "ordinary deregistration." "After several trainings, the big model learned to ask back which type of deregistration it is, and even took the initiative to ask whether it is to deregister a domestic company or a foreign company. It learns very quickly."
In addition to the daily "training" of window staff, the amount of data also determines the upper limit of the capabilities of the large model.
Last year, Xuhui District took the lead in launching the "Online Government Service Hall" in the city, and then launched the "Processing Center" module. This allowed Xuhui to be the first to have a "problem discovery model" in the field of government services. After conducting intelligent self-inspections on 930,000 pieces of processing data in four categories, including timeouts, multiple corrections, multiple trips, and material specifications, it was converted into usable and queryable large model basic data.
The "digital foundation" accumulated by the Xuhui District Big Data Center covers high-frequency business, legal policies, and daily case handling data, so that when the Xuhui government service big model was "born", the basic data scale reached 12.31PB. In addition, Xuhui currently supports the large language model computing power of 72B parameters, and the "infrastructure" of the government big model is guaranteed.
"We have innovatively combined Tongyi's big model technology with Xuhui's government service items. Through the comprehensive application of intelligent agent technology and RAG retrieval enhancement generation technology, we have improved the government service big model's ability to learn business knowledge, as well as the accuracy of intent classification and recognition and professional question answering."
Zhang Peng, solution architect of Alibaba Cloud and project leader of the Xuhui government service big model, said that after experiencing the intuitive changes brought about by ChatGPT and facing the industry truth that gradually emerged as the "100 Model War" receded, many practitioners found that it was not enough to just increase the parameters of the big model, and that the big model must be deeply integrated with the industry.
"As Shanghai's first national 'Internet government service' demonstration zone, Xuhui has both a solid big data foundation and a group of window reception staff who are highly sensitive to digitalization and can 'teach' new knowledge and new scenarios to the big model," said Zhang Peng.
The Alibaba team did three things in Xuhui: sort out the knowledge points that the big model needs to learn; summarize the high-frequency questions and answers of the comprehensive window for legal person affairs; and deploy the basic big model of Alibaba Cloud Tongyi Qianwen 72B in Xuhui. When the government affairs big model "thinks", it will mobilize the basic knowledge points, high-frequency questions and answers, and the general big model at the same time, and synthesize them into one answer through the algorithm.
In actual use, the window staff corrects the knowledge points that are missing or used incorrectly in the big model, fills in the correct answers and submits them to the background. The algorithm will apply the above feedback to the model library. After the big model learns by itself, it can give more accurate answers when encountering similar problems again.
In this process, it is the Alibaba Cloud Bailian platform that helps government service scenarios achieve "business intelligence", playing a key role in linking industry data and computing power.
"The government consultation big model product provided by the Alibaba Cloud Digital Government team can help new window staff to be as confident as 'old masters' who have worked for many years." Huang Wei found that the big model has an unexpected advantage: it helps new window staff grow quickly. Previously, it took at least 4-5 months of training for a comprehensive window staff to go from joining the company to "working independently", and senior window staff had to sacrifice some time to "pass on the knowledge and skills" to new window staff.
Li Zian, born in 1995, is one of the first window staff to use the government affairs big model. Recalling that he had just transferred to the comprehensive window more than a year ago, the young man said frankly that it was somewhat "painful" to have to remember, recite, understand and master a lot of training content. For example, shareholders in limited liability companies and partnerships are respectively called "shareholders" and "partners", and the expressions should be switched at any time when dealing with companies of different natures. Even if you can open the book and read it at the window, you still need to memorize a lot of content in advance.
Now, all of these contents can be provided with the assistance of large models, and window personnel need to do more to characterize, classify and provide feedback on the problems.
In early March, Li Zian conducted a Q&A session with the government affairs big model for the first time and soon discovered that the other party would "answer questions irrelevantly". The reason was precisely that the big model "knew everything" and even had "knowledge hallucinations" - the scope of the corpus it mastered was too wide and the quality of the corpus was uneven, which often hindered the big model from being more "vertically focused".
During this year's National People's Congress, many delegates and members mentioned that the global large-scale model field is generally facing a data bottleneck. my country also has problems such as few public data sources in vertical fields and scarcity of professional vertical data, and there is still a clear gap in high-quality corpus.
"The government service center has a natural supply of high-quality corpus." Zhang Peng introduced that the basic corpus of the Xuhui government service model comes from the company registration management regulations, as well as the corresponding material specifications, disassembly specifications and filling specifications of 39 business situations in the market supervision field. Based on these scenarios, more than 200 knowledge points were sorted out, and finally combined with the high-frequency problems accumulated by the windows, nearly 2,400 subdivided knowledge points were derived.
As the counter staff use it day after day, the learning effect of the big model is quite significant. "At the beginning, when asked 'How to change the equity when the shareholder is in police station', the big model would be entangled in the colloquial expression, but now it has understood that the essence of this question is to ask about equity change, and can find the corresponding knowledge points to answer it." Li Zian said.
Xuhui Administrative Service Center Legal Person Affairs Comprehensive Reception Hall
At present, the Xuhui government service big model has built an application scenario of "the service user asks questions - the window staff converts - the big model gives detailed replies", which effectively improves the efficiency of window staff in reading laws and comparing specific numbers. Will "real" window staff be replaced by big models?
"In the future, there may be 18 digital people sitting in the comprehensive window, but at present, the big model is still in the process of learning and evolving, and the warmth and efficiency of face-to-face communication between people cannot be replaced." Huang Wei said. "During the test, we tried to urge the big model to answer quickly, but it was asked to 'burn out the CPU' and crashed." Liu Lei found from this example that humans should not be obsessed with whether they will be replaced, but use the time and energy freed by artificial intelligence for themselves in more advanced places.
It is reported that the "Eighteen Arhats Digital Humans" based on the Xuhui government service model are being planned to be launched, and the corresponding model algorithms will also be extended to various streets, towns and functional areas in Xuhui.
For companies with the ability to develop general large models, developing industry-specific large models is not something that a single company can do, and the entire industry ecosystem needs to move forward. Currently, Alibaba Cloud Magic Build, the largest open source community in China, has landed in Xuhui Binjiang. In the same area, the country's first large model innovation ecological community "Mosu Space" has gathered more than 80 large model companies just 9 months after its completion.
"China's artificial intelligence industry is full of talents, and ecological teams of big models are emerging frequently." Zhang Peng said that when more and more basic models are open source, more teams will focus on the industry to develop vertical models and empower the industry, so that big models can truly be used by people and realize the vision of cultivating and expanding new quality productivity.