Reporter: Hello, Mr. Zhong, we know that you did not work as an artificial intelligence trainer from the very beginning. Can you briefly introduce to us, under what circumstances did you choose to become an AIT? What do you think is the biggest difficulty for you in working on artificial intelligence?

Zhong Liang: Before working as an AI trainer, I tried to work in customer service, sales, finance and many other industries, and I was in a state of confusion about my future career planning. It wasn’t until 2016 that I received an interview call from Zhuiyi Technology, and stepped into the field of AI ignorantly. At that time, the country had no definition of this position, and I didn’t even know how to explain my profession to my family. At first, I did data labeling. Later, with the development of the business, I gradually undertook knowledge base construction, task-based configuration, test tuning, customer training, etc., and accumulated experience and skills in project practice, and gradually grew into a qualified AI trainer.

I think the biggest difficulty in transformation lies in the change of thinking, from human thinking to machine thinking. The robot itself has no logical thinking. When the model algorithm remains unchanged, the effect of the reply depends on the quality of the artificially fed corpus. A good trainer needs to have a certain understanding of machine learning technology and algorithms, and understand what kind of questions and corpus are helpful for model training.

Reporter: We know that you are mainly engaged in the training of intelligent service products in the home appliance industry. The construction of the knowledge base is one of the most critical parts. Could you please share with us how you built the knowledge base of intelligent services based on the characteristics of the industry and how to optimize it? What about?

Zhong Liang:  The way the knowledge base is built depends on what kind of data the company has. The way we use it is to clean the logs, cluster them, extract questions, and modify user questions as similar questions. If there is no historical log, you can refer to the internal customer service knowledge documents of the enterprise, use the machine reading comprehension function to split the documents, modify them into standard questions and answers that can be displayed to customers, and then train the model by manually writing corpus.

The advantage of logs is that they are real questions asked by users, and hot issues can be found through clustering. The advantage of customer service documents is that the knowledge coverage is relatively comprehensive, but the workload of splitting is heavy, and there may be some discrepancies with the actual online questions, resulting in a situation where the amount of knowledge is large but the response effect is not good.

The knowledge base is not static. No matter which method is used, it is necessary to continuously adjust the knowledge granularity and check for omissions and fill in gaps based on the data after the launch. In addition to FAQ, you can also sort out some large-scale and fixed-flow multi-round conversation scenarios, as well as multi-dimensional table structure knowledge. Combined with the characteristics of business scenarios, choose the most suitable knowledge display form.

Reporter: The positioning and boundaries of intelligent service products are issues that trainers need to determine before training. The former tells us what artificial intelligence should do? The latter is to help us determine what should not be done? How do you feel about this issue.

Zhong Liang: First of all, we need to clarify the purpose of using robots. Is it diversion? Is it marketing? or something else. The second is to understand which problems the robot can solve and which problems it cannot. Many people have high expectations for intelligent customer service. Some problems are very simple for humans, but the algorithm technology behind them is actually very complicated. Smart services are not omnipotent, and their positioning should be to solve simple, high-frequency, repetitive, and mechanical problems, and to obtain maximum benefits at the lowest cost.

Reporter: As you mentioned earlier, you are also responsible for the formation, training, and management of the AI ​​trainer team in your daily work. Could you please share with us how you build and train the AI ​​trainer team?

Zhong Liang: In my opinion, elementary AIT needs to have the ability of knowledge editing, data labeling and system testing. The advanced AIT is more like a cross position of product manager, data analyst, and project manager. In the initial stage of team creation, the requirement for members is to meet the primary standard. Our team members are all excellent front-line customer service. Some of them are good at statistics, some are good at speech editing, and some are good at recording analysis. Everyone is proficient in the professional knowledge of the home appliance industry. Through training and job rotation, they can find a suitable position for everyone and be responsible for the categories, channels, and modules they are good at. In the way of brainstorming, regularly share the problems encountered, tips, and AI industry knowledge, and formulate and improve clear operating procedures and rules. In addition to internal training, the company leaders also support us to participate in some external training and competitions, hoping to communicate with more peers and learn from experience and new ideas.

Reporter: Without assessment, there is no progress. How did you design the performance indicators for the assessment of the artificial intelligence trainer team?

Zhong Liang: The performance evaluation indicators should be quantifiable and linked to the robot’s operating goals. Our current assessment is mainly divided into workload and quality of work. For workload assessment, it is necessary to set different conversion coefficients according to the difficulty level of the work content. The assessment of work quality, for example, the text robot mainly assesses the conversation resolution rate of the responsible channel and the improvement effect of satisfaction, and the voice robot mainly assesses the optimization effect of the order success rate and reassignment rate. Combined with the operating focus of robots at different stages, adjust the assessment plan. In addition, we also attach great importance to whether we can put forward innovative suggestions and improvement plans, and whether we can effectively coordinate with internal and external.

Reporter: He Yonglan

Written by: Qian Yi

Planning: Su Yu

Time: July 2022

Location: Online

Team introduction: Captain: Zhong Liang; Team members: Qin Huarong, Zhao Lanyan, Lin Xikun

The team was established in March 2021. The members are all excellent front-line customer service representatives with unique insights into customer service and intelligent interaction. At present, the team is responsible for the operation of text and voice robots. Through data analysis, speech optimization, process optimization, system iteration, etc., the robot’s conversation resolution rate, order success rate, and satisfaction rate are improved, and customer diversion and agent performance can be improved. achieved excellent results.