"I can reply immediately and the content is detailed. Some matters we did not expect will be included in the reply." As a law enforcement comrade participating in the seventh round of summer supervision and assistance for air quality improvement in key areas in 2024, Ecology from Puyang City, Henan Province Peng Jianzhi of the Environment Bureau consulted the "Intelligent Question-Answering Model for Supervision and Assistance" about relevant environmental management policies and said it was too convenient.
To realize the goal of building a beautiful China, we need to fully unleash the superposition, aggregation, and multiplication effects of digital technology in ecological and environmental governance, and stimulate the potential and vitality of the "intelligence" of the ecological environment. This large intelligent question and answer model, which combines deep learning, natural language processing and other advanced technologies, is providing efficient and powerful analysis and decision-making support for supervision and assistance work.
"It's easy to climb the storage tank every day, but it's really difficult to get all the news in the group."
In June 2019, the Environmental Engineering Assessment Center of the Ministry of Ecology and Environment (hereinafter referred to as the Assessment Center) deployed 28 technical backbones to form a special atmospheric class to provide technical support for atmospheric supervision and assistance. Since then, the WeChat Q&A group has become a must-have for every round of work. .
"Generally speaking, the consultation issues include standard application, policy compliance, and inspection points. The standardized filling of supervision and assistance APPs is also a hot issue in the group." Hao Shaoyang, an engineer in the atmospheric class, told a reporter from China Environment News.
In 2019, Hao Shaoyang entered the assessment center and has been working in the atmospheric class for five years. "The WeChat group has a special class responsible for clearing up doubts. Because the consultation questions involve different fields, the special class will also contact experts to help answer the questions. Currently, the expert team has stabilized at more than 30 people."
Regarding the work of clearing up doubts in the WeChat group, Hao Shaoyang summed it up with the word "urgent". "Our Q&A time is 24 hours a day. As long as the supervision and assistance work continues, Q&A cannot stop. Most of the questions are consulted by the working group on site, so we are required to complete the answers quickly and accurately. Comrades from the atmosphere class I have a habit of checking the WeChat group before going to bed at night, and I will fall asleep after a period of time when no one asks me questions.”
In the past five years, the atmospheric class has established more than 100 Q&A groups, and has answered nearly 60,000 questions on various front-line technical problems. It has compiled a total of 7 Q&A manuals on VOCs supervision and assistance, industrial furnaces, etc., covering more than 400 questions. A typical question.
However, the problem of clearing questions through WeChat groups is also obvious. "Each batch of WeChat groups must support on-site work by dozens or even hundreds of groups in key areas, and there are a lot of messages in the group. However, the on-site group can only read the chat history to understand the key issues in the group, which is very inconvenient." Xu Haihong, a senior engineer in the atmospheric class, told reporters that some law enforcement comrades previously said, "It is easy to climb the storage tank every day, but it is really difficult to 'climb' all the news in the group."
In addition, repeated consultations on some norms and procedures also consume too much energy of the atmosphere team members. "Traditional manual processing methods are no longer able to meet the current high-efficiency requirements of environmental protection work, and the introduction of intelligent technology to improve the efficiency of supervision and assistance has become an inevitable choice." said Lu Xiaojun, head of the atmospheric class.
“Feeding” large models with massive ecological and environmental data
Use the generalization capabilities of large models to integrate knowledge, data and logic in the vertical field of ecological environment, allowing you to focus on ecological and environmental protection work. Therefore, in accordance with the requirements for summer supervision and assistance for air quality improvement in key areas, the Assessment Center began to take the lead in organizing the development and implementation of the "Intelligent Question-Answering Model for Supervision and Assistance".
However, there are many general knowledge models in society and few ecological and environmental professional models. The key to building a large model in the field of supervision and assistance lies in how to "feed" the large model. The large model learning content is very rich, and information such as Q&A manual data, ecological and environmental protection laws and regulations, ministers’ mailbox replies, and emission standards are all included in the learning scope. Massive ecological and environmental knowledge can help large models exercise logical thinking and generate accurate responses to questions.
In 2023, the prototype of the "Supervision and Assistance Intelligent Question-Answering Model" will take shape. The atmosphere class was not in a hurry to put the model into practice, but conducted a comprehensive test and analysis on it and asked the development company to further strengthen the training of the large model.
A screenshot provided to reporters by Song Lu, a member of the atmospheric class, is a response to a large model question and answer last year on "Is white latex a VOCs material?" "In the 2023 big model's answer, white latex was identified as a VOCs material. In this year's answer, it was clarified that if white latex is simply used without adding other high-VOCs ingredients, it generally does not need to be considered a VOCs material." Song Song Lu said that the change in answer content reflects that the large model's logic has become clearer through continuous learning and error correction, and its response to questions has become more accurate.
In this year’s supervision and assistance work, the “Supervision and Assistance Intelligent Question-Answering Model” was officially put into use. Data show that in the sixth round of supervision and assistance work, 32 professional groups used the Q&A model to answer questions, with a usage rate of 62%, a total of 636 times, and an average of 53 queries per day, with more than half of the working groups using it.
"The intervention of the big data model enables faster responses to normative and policy issues, making it easier for on-site working groups to consult. Relying on the powerful information summary capabilities of the large model, on-site comrades can expand problem-solving ideas, and comrades in the atmosphere class can also be more Focus on analytical support for complex problems,” Xu Haihong said.
In the future, it will be disclosed to ecological and environmental law enforcement personnel across the country.
In fact, the "Supervision and Assistance Intelligent Question-Answering Model" is still in the process of iterative upgrades. What is currently being planned is to improve the large model's learning and analysis capabilities for formulas, tables, pictures and other types of information.
Xu Haihong told reporters that the special class and supervision and assistance personnel found that there are many formulas, tables, and pictures in the standards and specifications related to ecological environmental protection, but there is a significant gap in the learning and processing capabilities of large models for this type of information compared with texts, which has a direct impact Accuracy of responses to relevant questions.
It is understood that the technical support company is currently improving the multi-modal capabilities of the model. By optimizing the conversion algorithm, the extracted picture data and formula information are converted into a format that is easier to understand and process by the large language model, thereby improving the overall processing efficiency and accuracy.
It is worth mentioning that the reporter learned from the assessment center that the “Intelligent Question-Answering Model for Supervision and Assistance” will be made public to ecological and environmental law enforcement personnel across the country in the future.
“Disclosure to ecological and environmental law enforcement officers across the country can improve the efficiency and accuracy of on-site inspections by law enforcement officers, solve some cases of errors in the application of standards, norms and laws during on-site law enforcement, and at the same time continue to improve the large model’s ability to understand and understand professional environmental issues. Professionalism." Lu Xiaojun said that the assessment center will continue to deepen the application of technology in the field of ecological environment, explore more intelligent solutions, and contribute to protecting the ecological environment and promoting green development.