How do chatbots do? One of the common ways is through deep learning; the other is the use of information theory, that is how to let robots chat and how to feedback. The use of deep learning is now widely used, while the second approach is currently in use and is still at an exploratory stage.
At the Global AI and Robotics Summit, academicians of the Royal College of Canada, Professor of the University of Waterloo, ACM and IEEE Fellow Li Ming mentioned that using the Context Model for deep learning as a chat robot, the robot’s answer to the dialogue is too general. The pain point of the product to be solved is to make the machine's answer more targeted. So Li Ming and his research team added a CNN encoder based on this, gave the robot a contextual question, and then conducted hundreds of millions of question and answer trainings on it, and finally trained 40 to make the answer more specific. And the accuracy is above 80%.
It works by first accessing two Contet Models. Through different connections, the CNN classification results are input to the RNN, so that the problems can be more accurately understood and the correct answers can be made.
Siri is an application behind natural language processing. There are still some problems. When the user asks Siri “What does the fish eat?†Siri's internal system extracts the two keywords “fish†and “eat†, and understands that the user’s intention is to eat seafood, so the answer is to list many seafood restaurants. . If you do not use deep learning to use template matching, you will also have problems. The flexibility of template matching is poor. You can ask the question, "How is the weather today?" It can be answered, but if you change to " What's the weather today?", problems will arise.
Recently Li Ming made a study to identify how many substances are in a cell. The flow is first given to a cell, the cell is crushed and separated, and after the separation, the black spot is taken out, and the spectrum is generated after shaking with a mass spectrometer. The spectrum is its mass spectrum, and a very simple CNN is written according to Li Ming. Model, connected to the LSTM to complete the identification. In this case, deep learning does not work on its own due to noise problems. In addition, it also needs to do dynamic planning, through countless CNNs, and eventually solve it with dynamic programming.
One of the major technical challenges for question and answer robots is the moderate feedback system. Alpha Dog's feedback system, in layman's terms, means that if you lose one game, you lose one point, and vice versa. Compared to Alpha Dogs' simple feedback on wins and losses, chats and quizzes cannot be fed back with simple right and wrong, lose and win, so there is a need for a suitable metric system. Li Ming proposes a measurement system concept. Based on information theory, they need to measure the similarity of two sentences or the similarity of one question and one answer, and finally find the semantic distance between them. Semantic distance is not calculable, but semantics can be approximated, and the approximate implementation is compression. The approximation of language is measured by compression.
Li Ming uses deep learning from several other perspectives to solve natural language problems and provide researchers with a new perspective.
We are manufacturer of Ev Car Chargers in China, if you want to buy Level 2 Ev Charger,Evse Charger,Portable Charging Station please contact us.
Ev Car Chargers,Level 2 Ev Charger,Evse Charger,Portable Charging Station
Shenzhen Hongjiali New Energy Co., Ltd. , https://www.hjlcharger.com