Chatbot is a very hot word or an application in recent times. Not only are the major news media hot-selling the concept of bots, but the giants are also investing huge resources in research and development. It is also common for arxiv to brush out bot-related paper. . Hype is hype, PR is PR, and it must be said that the fact that it is really difficult to find a really good bot on the market. The bot is divided into open-domain and task-oriented bots according to the fields involved. The open domain has a lot to do, more like a platform that can do anything, no matter what kind of requirements you can, it can solve, a little trueAI, and task-oriented bots focus on one thing , book tickets, order meals, do passports, and more.
When it comes to open domain bots, the most popular ones are the entertainment bots that answer very inconspicuously. For example, the little yellow chickens that were active on major social networking sites many years ago are now active in the market and claim to have mastered the bot technology. The bot companies that use deep learning to solve bot technology are all this. They can't solve any practical problems. They can talk to you two times, and in many cases, the answer is that the bulls are not right, it is very ridiculous.
Let's say task-orientedbot, the most on the market is customer service robots , banks or e-commerce. If you don't want to answer the user's questions repeatedly, you can use a customer service robot to deal with it, and don't say how it works, develop a specific task bot. It takes a lot of work, and it requires a lot of maintenance in the later stage. Because too many handcraftedfeatures are used, the horizontal scalability of the entire bot framework is relatively poor. In other cases, it is necessary to re-develop a set of labor costs. Too high.
The ideal of the bot is very full, and the scene depicted by the big company is really beautiful, but the actual bot has poured a cold water down. The higher the expectation, the greater the disappointment. If the media touts the bot blindly, it seems that the whole world will be a bot tomorrow, and it will not be beneficial to the development of the bot. The killing will only bring bubbles, and after the break, everything will be the same.
A powerful, open-domain bot is difficult to achieve in the short term, but if you lower your expectations, it should be a technical level revolution, but an interactive level of innovation is a rational attitude. An entry, maybe everyone no longer needs a terminal to carry around, just need to find a identifiable, networkable hardware, such as a mirror, you can perform a lot of tasks, book tickets, buy things, and so on. At this time, the bot plays the entrance of an operation and the black box behind the various tasks. We don't need to see the whole execution process, and we don't need to know what the principle is. Some simple language interaction can complete some complexities. The task of the terminal is to feed back the results and receive the input. The execution process is in the cloud, and various bot clouds.
The key to all this is to solve the task-orientedbot and replace the traditional artificial features and templates with more data-driven solutions.
Problem Description
The bot is a comprehensive issue involving the following three main issues:
1, responsegeneration (selection)
The dialog generation is the last step and is the part of the output. In a brief summary, there are four kinds of solutions:
Solution1 directly generates dialogs based on context. Recently, there are many papers in this area. Especially after the seq2seq+attention framework swept through many tasks of NLP, the benchmark generated by the dialog was refreshed by various models again and again. The problem generated by the dialogue is defined as the generation model based on a certain condition. The typical predictwords based on the context, which involves the problem of sentence generation, the evaluation problem will be a difficult problem.
Solution2 Of course, some paper does not define dialog generation as a language model problem, but a problem of nextutteranceselection, a multiple choice question, given a context, given a utterancecandidatelist, select one from the list as a response, of course The difficulty of the class problem is much smaller, and it is very easy to evaluate, but the data set is more time-consuming to prepare, and it is not easy to learn from the actual application.
Solution3rule-based or template-based, the final form of response is actually filled with a template, most of the things are given, only some specific values ​​need to be filled. This type of solution is well suited for task-orientedbots, but too many artificial features and templates make it difficult to port to other tasks.
Solution4query-based or example-based, response is from a database called a knowledge base, which contains a large number of rich examples, according to the user's query, find the closest example, the corresponding response is returned as output. This type of solution is very suitable for entertainment, funny bot, the core technology is to find more data to enrich the knowledge base to clean the knowledge base. But after all, respnose is taken from others, it may be very funny, but most of them will be wrong.
2, dialogstatetracking (DST)
Some papers call DST as believetrackers. This component is actually the core of the bot. Its role is to understand or capture userintention or goal. Only when you really know what the user needs, can you make the correct action or response. About this section, there will be DialogStateTrackingChallenge competition. In general, a range of states will be given. The context is used to predict which state the user belongs to. What kind of requirements are needed? It is necessary to check the weather or to check the train ticket.
Pen Electronic Cigarette,Pen Style Electronic Cigarette,Mesh Coil Disposable Electronic Cigarette,Disposable Electronic Cigarette
Jinhu Weibao Trading Co., Ltd , https://www.weibaoxd.com