Electronic enthusiasts eight o'clock: Natural language processing has always been a huge challenge on the road to artificial intelligence development. Previously, most research has taught machine learning models to be trained on a large number of tagged data sets. Recently, Baidu Research Institute has proposed a new approach in which researchers let artificial intelligence systems learn natural language and knowledge through spoken conversations with “teachersâ€. This method of learning a language process like a baby shows great potential. The heart of the machine compiled the article and its abstract of the research paper. The original link is at the end of the article.
In early April, Baidu's research team successfully taught artificial intelligence agents to navigate through the maze by issuing a natural language command from a virtual teacher (see: Using Artificial Language Education Artificial Intelligence: Baidu's New Algorithm Developed Zero-shot learning ability). Today, the Baidu research team is pleased to announce that its artificial intelligence agents have successfully learned to speak through interaction with virtual teachers.
Speaking, as well as other basic human abilities, are indispensable on the road to creating universal artificial intelligence. Although simple conversations with machines today are common, Baidu's research team teaches machines to talk differently than traditional methods.
Baidu's artificial intelligence agent learns to speak in a way similar to baby interaction. In contrast, traditional methods rely on supervised training, using static corpora containing a large number of pre-collected training sets, making it difficult to capture dynamic interaction attributes in language learning. As a result, systems trained by traditional methods primarily reflect behavior in the data set, with limited adaptability and generalization capabilities. Baidu's artificial intelligence agents learn to speak through interaction, aiming at acquiring language learning and understanding capabilities rather than just capturing statistical patterns in the data.
When a baby learns to speak, he interacts with people and learns through imitation and feedback. Infants initially generate verbal behavior by imitating their narrators and master the skills of word generation. Babies also make sounds to their parents and adjust their speech behavior based on parental corrections and encouragement.
Research overviewBaidu researchers have proposed an interactive approach based on natural language learning, in which artificial intelligence agents learn natural language by interacting with virtual teachers (professors) and gaining feedback, thereby learning and improving natural language skills to achieve the degree of participation in dialogue. . Here, there is no supervised learning in the form of tagged data to guide learners; instead, the system must learn to speak by constantly trying to speak, and the professor will provide verbal feedback (if yes/no) and non-verbal feedback (such as nodding/ smile).
The figure below shows several different forms of dialogue in training. In the beginning, the agent can only generate meaningless sentences, it can only improve their skills in pure dialogue. At the end of the day, the agent can correctly use the natural language to answer the questions raised by the professor.
Experiments on the other hand further prove that the new method has the ability to learn natural language. Researchers have shown that trained AI agents can answer new questions that are formed by concepts in known knowledge or problems, but are reorganized. For example, in training, the "avocado, east" combination never appeared in the question and answer; and orange was only described, never asked by the professor. In the test, the agent can answer questions about "avocado" in "east" or questions about "orange", as shown above.
Baidu researchers said they will further increase the complexity of the language learning environment in the future to train more complex language behavior. In addition, they plan to explore the knowledge modeling and rapid learning of machine learning systems, allowing artificial intelligence agents to interact naturally with humans and enable them to learn effectively from the physical world.
Thesis: Listen, Interact and Talk: Learning to Speak via InteracTIon
Link to the paper: https://arxiv.org/abs/1705.09906
Abstract: A long-term goal of artificial intelligence is to build an agent that can interact with humans in natural language. However, most of the current research in natural language learning relies on a large number of annotated tagged data sets for training, which causes the task of the artificial intelligence agent to become a statistical grab of the external data set. Since the training data is essentially a static representation of the knowledge by the annotator, the adaptability and expandability of the artificial intelligence agent after learning is limited. In addition, this training method is very different from the process of human learning natural language, which is a process of communication, which is done by speaking and getting feedback.
In this paper, we present an interactive form of natural language learning. Among them, the artificial intelligence agent learns and improves language skills in the conversation by interacting with the natural language with the teacher. To achieve this goal, we constructed a model that included imitation and reinforcement learning methods to compare sentences and professor feedback. We conducted experiments to prove the effectiveness of this method.
Disclaimer: The electronic reprinted works of E-Commerce Network are as far as possible to indicate the source, and all rights of the owner of the work are not transferred due to the reprint of this site. If the author does not agree to reprint, please inform the site to delete or correct it. Reprinted works may be subject to change in title or content.
Absolute Linear Encoders,Custom Absolute Encoder,Rotary Encoder Magnetic,Miniature Absolute Encoder
Yuheng Optics Co., Ltd.(Changchun) , https://www.yhencoder.com