When we talk about human-computer dialogue system
Man-machine system reality
Development of man-machine dialogue
Four main functions of human-computer dialogue system
Five main modules common bot
Enter the pretreatment
Speech recognition technologies include feature extraction, pattern matching model training techniques and guidelines aspect 3
Natural Language Understanding
Chatterbot natural language understanding system includes a user intention recognition function, the user emotion recognition, refer to digestion, recovery is omitted, returns an acknowledgment and rejection judgment ¨ art cited. And syntax analysis, semantic analysis, keyword extraction, the similarity calculation natural language processing techniques are indispensable.
Dialogue Management
Enclosed
Closed means that there is a clear performance targets and service objects, only deal with specific issues in the field, to the theme of the dialogue was limited.
Open
Open is no clear objectives and targets, the theme of dialogue involving a wide range, the amount of knowledge required is enormous.
Answer generator
The answer involves a whole generation of content selection, text planning, synthetic statement, referring to representatives of generation, surface realization phase.
Generating a search formula including technical and formula
Search query
Refers to the search query matches the search to find the best answer for user input statements give a reply in the dialog library.
Generative
Formula refers to the use of certain techniques (such as deep learning techniques) automatically generate a new reply.
Output processing
Chat robots are built
Artificial template-based chat robot
The sentence input by the user, found in the template library template matching question and answer generating a response in accordance with a corresponding template, return to the user, such as ALICE, Chat Scfipt like.
Based search bot
Library requires dialogue is very high, and needs to be large enough, but it has the advantage of high-quality answer, more natural expression.
Based on formula bot depth study
According to the sentence entered by the user, using a model-by-word or generate answers verbatim, then the answer will reply to the user. Most of the techniques employed Encoder.D ecoder model, i.e., a decoded coding model
In this way the idea is simple, scalable, able to better understand the context, but it is difficult to model training, often there are some grammatical errors reply.
Comparison of three methods of constructing
Engineering practice topics of research and analysis of similar products
我的工程实践选题的大致方向是基于深度学习在一定程度上实现多轮对话，关于人机对话系统的一些软件产品及其特点上文中已在表格中呈现。作为商业产品的人机对话系统往往采用基于人工模板和检索式模型，如Cortana小娜，度秘，Siri等均为封闭领域的个人助理产品，依附于成熟的平台而存在。而开放领域的生成式人机对话系统由于其训练困难，且容易出现各种回复错误而很难产生商业价值，多数为尝试性质的“玩具产品”，为大家所熟知的微软小冰是此类产品中能够尝试进行商业落地的产品，通过其生成式模型所带来的创造性回答及功能丰富的插件收获了不少粉丝，但她相对于商业化产品更倾向于是着眼未来的研究性尝试。
我是大一的时候知道有这样一个微信公众号，是被同学推荐关注的。但当时觉得对话质量真的很难吸引我，而且我也不是一个喜欢养小猫小狗的人，更何况是一个虚拟的存在呢，所以也很少去看。而当我对相关的实现方法有了一点微末了解之后，才发觉想要实现这样的效果需要投入多少的人力物力，且其效果已经达到了可望而不可及的地步，顿时心生敬意。
该领域尚处于发展的初期，我们所说的成熟也只是相对于早年投入测试的阶段，但距离成熟的应用还有很长的路要走。但微软小冰从14年公测至今，其成长已经带给我们太多的惊喜，开发团队也以很高的频率进行着新的有趣的尝试，或许再过十年二十年，开放域的聊天机器人真的会以稳定的姿态融入我们的日常生活中。
