Low-cost, high-IQ, knowledgeable AI assistants take up jobs
Innovation story
◎Our reporter Liu Yan
Zhao Hanyu, an algorithm engineer of Zhiyuan Large Model Research Center, came to Beijing Zhiyuan Artificial Intelligence Research Institute (hereinafter referred to as Zhiyuan Research Institute) as soon as he graduated.
One of his patented technologies has played a key role in boosting the cooperation between the Wudao model and the OPPO Xiaobu assistant.
Zhao Hanyu said: "The moment my research results came to fruition, I felt that my value was recognized, and I was proud that I could truly make a contribution to the industry and society."
In the cooperation with OPPO's open dialogue virtual voice assistant Xiaobu, the "generative answering system" based on the Wudao model has solved the long-tail problem of industry commonality to a certain extent.
Today, AI voice technology has a relatively mature implementation plan for intelligent assistants. There are countless application cases in scenarios such as smart homes, smart vehicles, news broadcasts, and smart terminals. However, many people who have "fighted" with AI customer service have I have had the experience of wanting to scoop it out of the phone, and even deeply doubted the "intelligence" of AI customer service.
According to incomplete statistics, at least 100 million users have asked questions that cannot be answered by AI voice assistants.
Even though almost all smartphones today have a voice assistant, less than 5% of people use it.
At this stage, "retrievable answer" is one of the main human-computer interaction methods of most virtual assistants, and the content of the dialogue is limited to the model self-built database and Internet data.
This means that if the answer to a question such as "the length of the Yangtze River or the length of the Yellow River" cannot be found in the database or search engine, the virtual assistant will not be able to answer it.
From the enterprise side, a single question from a user is large in volume and cannot be exhausted. Answering a user's question based on Internet data has a low hit rate. However, the self-built corpus not only has a long period of time, but also the cost of manual data labeling is high. This is a common situation in the entire industry because The inclusion of large models has been improved.
Tang Jie, academic vice president of Zhiyuan Research Institute and professor of Tsinghua University, introduced that the Wudao open platform provides two sets of solutions.
First of all, the Wudao model has powerful small-sample learning and text understanding and generation capabilities, and can generate answers to questions already raised by existing users of Xiaobu Assistant.
In addition, for the questions that users may ask in the future, the Wudao large model combines continuous learning with knowledge graphs to form a continuous knowledge pre-training model, and then generates a question-and-answer corpus.
Liu Haifeng, president of OPPO's Digital Intelligence Engineering Division, said that the construction cost of OPPO Xiaobu's single answer has been reduced by 99%.
"I hope to create a more intelligent knowledge-based, interactive and chat-based intelligent assistant in the future, which can be more open and have more diverse styles when answering human questions." Tang Jie said, "Of course, the goal of humanized interaction is very important. Natural language understanding and generation of models has put forward higher requirements, and there is no better solution yet, but the Enlightenment Big Model is moving in this direction."
On the path to general artificial intelligence paved by large models, AI assistants have made great progress in listening, seeing and feeling, becoming more and more "knowledgeable", with few functions, inaccurate speech recognition, and voice wake-up. Difficulties and other manifestations of unintelligence are being eliminated one by one.