By Huang Xiaohui, People's Daily
Yang Hongxu (right) discusses AI training progress with a colleague. (Photo by Huang Xiaohui/People's Daily)
Opening the computer and starting an intelligent training software, artificial intelligence (AI) trainer Yang Hongxu began his work. His recent task is to train an AI model that can autonomously recognize bank wire transfer applications.
Yang works for Data Grand, a company located in Shanghai Pudong Software Park, specialized in intelligent text processing. The AI models developed by the company can read various types of text files, such as contracts, financial statements, and industry reports, and have been applied in scenarios such as bank transaction recognition, and contract auditing and comparison.
"Before training an AI model, the trainer needs to read relevant documents and extract and annotate key information," Yang said. He compares this process to a teacher preparing for a class, which is the foundation of training.
"We highlight and annotate important parts, and then 'feed' the AI models with 'useful' information, telling them which information to extract and in what format," said Yang, adding that AI trainers need to read tens of thousands of documents and train dozens of AI models each year.
Yang opened a scanned image of a remittance application form from a bank in the intelligent training software. He selected the fields such as "transaction type," "remitter name," "account number," "bank name," "contact address," one by one and labeled them accordingly. Then he clicked the mouse to generate a training model.
Once the model was generated, he raised some practice questions for the AI model. He input the field information and compared the results generated by the AI model with the original text to evaluate how well it had learned.
Whenever he noticed any discrepancies between the generated results and the original text, he recorded and compiled them into a list. For example, an image in the document was not recognized, and an upright image was recognized as sideways. These issues were then collectively provided as feedback to the developers of the AI model for future improvements.
"After training, the AI model can independently extract and read the key information from similar text files, enabling it to execute human instructions more efficiently," Yang said.
While others may perceive tasks like document reading, data annotation, and result testing as monotonous, repetitive, and somewhat dull, Yang finds them effortless. His greatest joy is witnessing the increasing intelligence of the models he trains.
Yang, born in 1993, has congenital hearing impairment and needs to wear hearing aids to hear sounds. In 2014, he graduated from a vocational and technical college in Shanghai with a major in environmental design. He worked different jobs such as home interior designer, accounting clerk, and warehouse administrator.
"When communicating with clients, I often felt frustrated because my speech was not clear and fluent enough," Yang said.
When he was feeling lost, the Disabled Persons' Federation in Pudong New Area recommended him for the AI trainer job at Data Grand. Despite having no programming skills and limited knowledge of AI, Yang decided to give it a try when he learned that the job mainly involved working with machines.
Yuan Shaoyang, head of operations and management of Data Grand, who recruited Yang, said that the first batch of AI trainers of the company mainly engaged in basic data annotation. The repetitive tasks and high volume of annotations require a great deal of patience, attention to detail, and focus, Yuan explained.
Over the course of five years, Yang, who is diligent, focused, and eager to learn, has earned a good reputation among his colleagues. Through training sessions organized by the company and his own efforts to learn, he gradually became proficient in tasks such as model processing, product testing, and client communication. In 2022, he became a certified Level-2 AI training instructor, among the first batch recognized by Shanghai.
To adapt to the rapid development of the AI industry, Yang has taken the initiative to learn mainstream programming languages for AI models so that he can write some simple code to address issues that arise with the models.
"For example, when I discover a problem with the trained model misinterpreting information, I can write a few lines of code to make some initial corrections instead of directly handing the problem over to the model developers," Yang said.
In order to improve himself, he is currently preparing for a certification exam for project manager of information systems.
AI is one of the three leading industries that Shanghai seeks to develop. In recent years, Shanghai has been strategically developing the industry in terms of computing power, language resources, models, testing, and application scenarios, while also increasing efforts in talent cultivation.
In 2022, Shanghai initiated the professional skill level certification for AI trainers, and so far, more than 450 people have become certified AI trainers.
Yang works for Data Grand, a company located in Shanghai Pudong Software Park, specialized in intelligent text processing. The AI models developed by the company can read various types of text files, such as contracts, financial statements, and industry reports, and have been applied in scenarios such as bank transaction recognition, and contract auditing and comparison.
"Before training an AI model, the trainer needs to read relevant documents and extract and annotate key information," Yang said. He compares this process to a teacher preparing for a class, which is the foundation of training.
"We highlight and annotate important parts, and then 'feed' the AI models with 'useful' information, telling them which information to extract and in what format," said Yang, adding that AI trainers need to read tens of thousands of documents and train dozens of AI models each year.
Yang opened a scanned image of a remittance application form from a bank in the intelligent training software. He selected the fields such as "transaction type," "remitter name," "account number," "bank name," "contact address," one by one and labeled them accordingly. Then he clicked the mouse to generate a training model.
Once the model was generated, he raised some practice questions for the AI model. He input the field information and compared the results generated by the AI model with the original text to evaluate how well it had learned.
Whenever he noticed any discrepancies between the generated results and the original text, he recorded and compiled them into a list. For example, an image in the document was not recognized, and an upright image was recognized as sideways. These issues were then collectively provided as feedback to the developers of the AI model for future improvements.
"After training, the AI model can independently extract and read the key information from similar text files, enabling it to execute human instructions more efficiently," Yang said.
While others may perceive tasks like document reading, data annotation, and result testing as monotonous, repetitive, and somewhat dull, Yang finds them effortless. His greatest joy is witnessing the increasing intelligence of the models he trains.
Yang, born in 1993, has congenital hearing impairment and needs to wear hearing aids to hear sounds. In 2014, he graduated from a vocational and technical college in Shanghai with a major in environmental design. He worked different jobs such as home interior designer, accounting clerk, and warehouse administrator.
"When communicating with clients, I often felt frustrated because my speech was not clear and fluent enough," Yang said.
When he was feeling lost, the Disabled Persons' Federation in Pudong New Area recommended him for the AI trainer job at Data Grand. Despite having no programming skills and limited knowledge of AI, Yang decided to give it a try when he learned that the job mainly involved working with machines.
Yuan Shaoyang, head of operations and management of Data Grand, who recruited Yang, said that the first batch of AI trainers of the company mainly engaged in basic data annotation. The repetitive tasks and high volume of annotations require a great deal of patience, attention to detail, and focus, Yuan explained.
Over the course of five years, Yang, who is diligent, focused, and eager to learn, has earned a good reputation among his colleagues. Through training sessions organized by the company and his own efforts to learn, he gradually became proficient in tasks such as model processing, product testing, and client communication. In 2022, he became a certified Level-2 AI training instructor, among the first batch recognized by Shanghai.
To adapt to the rapid development of the AI industry, Yang has taken the initiative to learn mainstream programming languages for AI models so that he can write some simple code to address issues that arise with the models.
"For example, when I discover a problem with the trained model misinterpreting information, I can write a few lines of code to make some initial corrections instead of directly handing the problem over to the model developers," Yang said.
In order to improve himself, he is currently preparing for a certification exam for project manager of information systems.
AI is one of the three leading industries that Shanghai seeks to develop. In recent years, Shanghai has been strategically developing the industry in terms of computing power, language resources, models, testing, and application scenarios, while also increasing efforts in talent cultivation.
In 2022, Shanghai initiated the professional skill level certification for AI trainers, and so far, more than 450 people have become certified AI trainers.