机器学习做自动聊天机器人
There are too many chatbot vendors, platforms and approaches out there. Almost every one claims to have a unique AI enabled approach. A few work. Many don’t. With marketing amped up, it’s harder and harder to discern the good from the bad. On the other hand there are bold mandates from the senior management for comprehensive digital transformation. All this leaves the business leaders responsible for implementing chatbots at sea. Add to this limited budgets for trials, limited appetite for in-house teams, limited access to data, and you get a scenario with many false starts and bots that barely work.
聊天机器人供应商，平台和方法太多了。 几乎每个人都声称拥有独特的AI支持方法。 一些工作。 许多人没有。 随着市场营销的扩大，分辨优劣的难度越来越大。 另一方面，高级管理层有大胆的命令要求进行全面的数字化转型。 所有这些使业务负责人负责在海上实施聊天机器人。 再加上有限的试验预算，对内部团队的胃口有限，对数据的访问受限，您会遇到许多错误的开始和机器人几乎无法工作的情况。
The article talks about a systematic approach to enable conversational experiences for your organization. It just skims the surface, but message me if you would like to discuss more.
本文讨论了一种为您的组织提供对话体验的系统方法。 它只是表面，但如果您想讨论更多，请给我发消息。
As with most things AI, planning while keeping the end user at center is the most important part of any project. The first step is to decide what exactly to you need. Not want. Need. Every thing starts from here.
与大多数AI一样，在使最终用户居中的同时进行计划是任何项目中最重要的部分。 第一步是确定您到底需要什么。 不想。 需要。 一切都从这里开始。
As examples, here are a few well thought-through objectives I have seen in my career:
举例来说，以下是我在职业生涯中已经深思熟虑的一些目标：
To let an internal or external customer discover some specific information e.g. finding out vacation policy for pregnant employees, or finding the best number to call for cancelling international airline tickets.
To enable a customer to complete a task quickly without waiting for a human agent e.g. blocking a credit card, or creating a ticket to report a bug in software.
To engage a prospective lead and push them further down the marketing journey e.g. point of sale discounts for bundled products, or generating leads on Facebook.
To enhance the value of the organization’s brand e.g. check out Sephora chatbot — almost everything it does is barely a click away in any case.
To save costs of human call center agents and call center infrastructure. (To be honest, this the worst objective to start with. Let this be a byproduct of the other objectives.)
The end-goal must dictate everything. For example, if you are building an internal chatbot for HR policies for the company, you may think about a Slack app with to-the-point answers driven by recall-first natural language search (don’t worry if you don’t know what that means). It’s reasonable to assume that employees will click an “Escalate to Human” button if they need to. You will think about ways that HR experts can answer the escalated questions at their leisure and the app can learn from their answers.
最终目标必须决定一切。 例如，如果您正在为公司的人力资源策略构建内部聊天机器人，则可以考虑使用Slack应用程序，该应用程序具有由召回优先的自然语言搜索驱动的即时答案(如果您不知道，请不要担心那意味着什么)。 合理地假设，如果需要，员工将单击“升级为人类”按钮。 您将考虑人力资源专家可以在闲暇时回答升级的问题，应用程序可以从他们的答案中学习的方式。
Whereas if you are thinking about a conversational interface for blocking lost credit cards, it must be a bespoke widget on your website, mobile apps, and other touch points e.g. social media. If you already have a chat interface, this app must integrate with it, triggering at the right intent. It must be able to pop open some structured forms for critical information like SSN and credit card number to ensure security and to evoke trust. It must be truly conversational and empathize. You will rely less on AI and more on manually crafted responses. Of course, it must be high-precision, it must be real time, and it must know on its own when to escalate to a human.
而如果您正在考虑用于阻止信用卡丢失的对话界面，则该界面必须是您网站，移动应用和其他接触点(例如社交媒体)上的定制窗口小部件。 如果您已经有聊天界面，则此应用必须与之集成，并以正确的意图触发。 它必须能够弹出一些结构化的表格来存储诸如SSN和信用卡号之类的重要信息，以确保安全性并唤起信任。 它必须是真正的对话和同理心。 您将减少对AI的依赖，而将更多地依靠手工制作的响应。 当然，它必须是高精度的，必须是实时的，并且它必须自己知道何时升级为人类。
In short, think about the right experiences and interfaces rather than the right chatbot.
简而言之，请考虑正确的体验和界面，而不是正确的聊天机器人。
(2. Decide how much AI is necessary)
After you know what you want, you need to decide who to work with to make it happen. The most important factor for that is to understand how critical artificial intelligence is to your vision. This will dictate build vs buy and choice of vendors, among other things.
知道自己想要什么后，您需要确定与谁合作才能实现这一目标。 最重要的因素是了解人工智能对您的视觉有多重要。 除其他事项外，这将决定构建与购买以及供应商的选择。
To understand the role of AI, let’s quickly look at the anatomy of a conversational interface. Typical bots are designed around manually crafted conversation flows. The bots expect users to enter one of the pre-programmed sets of keywords e.g. “Hi! I lost my credit card.” This is called an intent. The bots are programmed to reply to such intents as well, e.g. “I am sorry to hear that! Don’t worry, I can help you deal with this hiccup. When did you lose the card?” Again, the bot expects another set of intents e.g. “yesterday”, “on Tuesday”, or “just now”, and depending on the customer’s answer the bot responds appropriately or invokes some other function e.g. actually blocking the card after verification.
要了解AI的作用，让我们快速看一下对话界面的结构。 典型的漫游器是围绕手工制作的对话流设计的。 机器人希望用户输入一组预先编程的关键字，例如“嗨！ 我丢失了信用卡。” 这称为意图。 机器人也被编程为回答这种意图，例如： “我很遗憾听到这个消息！ 不用担心，我可以帮助您解决这个问题。 您什么时候丢了卡？” 再次，机器人期望另一组意图，例如“昨天” ， “星期二”或“现在” ，并且机器人会根据客户的回答适当响应或调用其他功能，例如在验证后实际阻止卡。
Artificial Intelligence helps such a bot in three ways:
人工智能通过三种方式帮助这种机器人：
- It helps match user inputs with the right intent. For example, if the user says “My card was stolen,” it’s perhaps the same intent as “Hi! I lost my credit card.” If so, the app must know some semantic relationships like card := credit card and stolen = lost. AI driven natural language understanding (NLU), can be very helpful for this. 它有助于正确匹配用户输入。 例如，如果用户说“我的卡被盗”，则可能与“嗨！ 我丢失了信用卡。” 如果是这样，则应用程序必须知道一些语义关系，例如card：=信用卡和被盗=丢失。 人工智能驱动的自然语言理解(NLU)可能对此非常有帮助。
- It helps understand prominent intents and likely conversation flow based on history of users’ interaction with human agents. Most organizations have copious transcripts readily available. Analyzing them using AI, NLU or other data science techniques tells you which intents should be manually programmed.
- Let’s face it — there is no way your team can think of all possible intents and program them. AI helps deal with un-programmed intent. Before escalating to a human, AI driven natural language search (NLS) can go through all your information base and feature the right content. This is not trivial. At a client my team integrated NLS with a ticketing system to successfully deflect 64% of tickets. 让我们面对现实吧-您的团队无法想到所有可能的意图并对它们进行编程。 人工智能有助于处理未编程的意图。 在升级为人类之前，人工智能驱动的自然语言搜索(NLS)可以遍历您的所有信息库并提供正确的内容。 这不是小事。 在一个客户中，我的团队将NLS与票务系统集成在一起，以成功偏转64％的票证。
So now you need to look at your product vision, available data and its nature to plan what parts need AI, and what parts traditional software can handle. AI comes with huge costs, least of which is the dollars you pay. It comes with marvelous promise as well.
因此，现在您需要查看产品愿景，可用数据及其性质，以计划需要AI的部件以及传统软件可以处理的部件。 人工智能带来了巨大的成本 ，其中至少是您所付出的成本 。 它也带有奇妙的承诺。
Broadly, AI is very useful when there are too many intents and the language generally used by customer base is too varied.
广义上讲，当意图太多并且客户群通常使用的语言变化太多时，人工智能非常有用。
(3. Avoid the common pitfalls)
When someone thinks of chatbots, there are certain metaphors and use cases that come to mind. Unfortunately, at times these metaphors are completely the wrong ones. Many business leaders fall prey to the fallacy of focusing on the tool rather than the experience. Three mistakes are prominent.
当有人想到聊天机器人时，就会想到某些隐喻和用例。 不幸的是，有时这些隐喻完全是错误的。 许多企业领导者容易陷入专注于工具而不是经验的谬误。 三个错误是突出的。
First, we should not focus only on text. In reality humans respond much better to visuals. “A picture is worth a thousand words” is not just a cliche for conversational experiences. Similarly, a two minute video at times is much more powerful in explaining something than a series of bullet points. It’s better to think of the entire experience from the user point of view and build in a multi-media format. The anchor to this experience can surely be a text chat, but the delivery of information doesn’t have to be. This paper linked below talks about how NLS handles images (look for “Textual Representation of Images”). There are multiple other approaches.
首先，我们不应该只关注文本。 实际上，人类对视觉的React要好得多。 “一张照片值一千个单词”不只是对话经验的陈词滥调。 同样，有时，两分钟的录像比一系列要点更能说明问题。 最好从用户的角度考虑整个体验，并以多媒体格式构建。 当然，这种体验的锚点可以是文本聊天，但信息传递不一定必须如此。 下面链接的这篇文章讨论了NLS如何处理图像(查找“图像的文本表示形式”)。 还有多种其他方法。
Second, we should try not to anthropomorphize. The motivation to do so is powerful — a digital persona best captures the metaphor of an intelligence substituting for a human. However, the early trend toward anthropomorphization is often associated with errors, bugs, mis-translation, etc. IBM started this trend but ended up being ridiculed for over-promising and under-delivering. Many others followed the same path. While there are exceptions, users typically associate that mistrust with digital personas.
其次，我们应该尽量不要拟人化。 这样做的动机是强大的-数字角色最能捕捉代替人的智力的隐喻。 但是，拟人化的早期趋势通常与错误，错误，翻译错误等相关。IBM 开始了这种趋势，但由于过高的承诺和交付不足而被嘲笑。 许多其他人也走同样的道路。 尽管有例外，但用户通常会将这种不信任与数字角色相关联。
Related to that is the idea that bots should be completely transparent as to when is a user talking to a bot vs. a human. The expectations become completely different — users change their behavior as to what they type, they expect a lower level of accuracy and relevance of responses, and they are willing to click links or tap buttons when they know they are talking to a bot. Trust levels are also different for different interfaces. For example, while providing their credit card numbers, users trust a form, a bot and a human in that order.
与此相关的是，关于用户何时与机器人对话，机器人应该完全透明。 期望变得完全不同-用户改变了键入内容的行为，期望较低的准确性和相关性，并且当他们知道与机器人聊天时，他们愿意单击链接或点击按钮。 不同接口的信任级别也不同。 例如，在提供信用卡号的同时，用户按此顺序信任表单，机器人和人。
There are numerous such pitfalls that do not sound intuitive. The best way is to take an agile product management approach to conversational experiences with copious inputs from the customers.
有许多这样的陷阱听起来并不直观。 最好的方法是采用敏捷的产品管理方法，通过客户的大量输入来进行对话体验。
(4. Make business case for the roadmap, not the project)
Let’s say you have articulated your vision (or the Desirability), and figured out what is possible and how to avoid the pitfalls (or the Feasibility). Once all the design elements are thought through, socialized, tested with users to the extent possible and put in place, a new challenge emerges — that of Returns on Investment (or the Viability). Designing any product is an iterative process. For conversational experiences, the Viability part forces the biggest constraints.
假设您已经阐明了自己的愿景(或可取性)，并弄清了什么是可能的，以及如何避免陷阱(或可行性)。 一旦对所有设计元素进行了深思熟虑，进行了社交化并与用户进行了测试，然后就位，就出现了新的挑战-投资回报率(或生存能力)挑战。 设计任何产品都是一个反复的过程。 对于对话体验，生存能力是最大的限制。
Let’s understand the two aspects of this problem. First, conversational experiences are bespoke to each situation. It is nearly impossible to just get a product off the shelf and quickly implement it. Second, the solution must be accurate. In most cases users will lose trust in it otherwise (see the Virtuous Cycle of Trust image). The need for accuracy further adds to the customization and costs. There is a trade-off for the first project. If we go very broad, accuracy is hard to maintain. If we go very narrow, then the costs don’t justify the expenses.
让我们了解此问题的两个方面。 首先，根据每种情况定制对话经验。 仅仅将产品下架并快速实施几乎是不可能的。 其次，解决方案必须准确。 在大多数情况下，用户会对此失去信任(请参阅“信任的良性循环”图像)。 对准确性的需求进一步增加了定制和成本。 第一个项目需要权衡。 如果我们走得很广泛，那么就很难保持准确性。 如果我们缩小范围，那么成本就不能证明支出是合理的。
In my experience this problem can be solved only by a phased roadmap approach — Start small, but be very clear that the same solution can be easily extended to many other situations. It is much easier to justify the business case for the roadmap, rather than a project.
以我的经验，只能通过分阶段的路线图方法来解决此问题-从头开始，但非常清楚，同一解决方案可以轻松扩展到许多其他情况。 为路线图而不是项目辩护的商业案例要容易得多。
Similarly, it may be better to deploy the solution in a phased manner where possible. Start with situations that are more tolerant. For example, if you are planning a completely virtual agent for your customers, why not expose it first to your customer service agents for fine-tuning.
同样，在可能的情况下，以分阶段的方式部署解决方案可能会更好。 从更宽容的情况开始。 例如，如果您正在为客户计划一个完全虚拟的代理，为什么不先将其提供给客户服务代理进行微调。
The 2m video below talks more about choosing the right scale in planning AI projects. The same holds true for chatbots.
下面的2m视频更多地讨论了在规划AI项目时选择合适的规模。 聊天机器人也是如此。
(5. Understand that AI is different from other software)
Finally, this is beginning to look like any other business project. Still there are many nuances that separate a conversational interaction project from others. I have written a lot about different aspects in the past. For example — cost, alignment, build vs buy and ROI.
最后，这开始看起来像其他任何业务项目。 仍然有许多细微差别将对话交互项目与其他项目区分开。 我过去写过很多有关不同方面的文章。 例如， 成本 ， 一致性 ， 构建与购买以及投资回报率 。
Alignment with all stakeholders is of particular note. It does sound like yet another cliche, but I have found that for projects related to automation there is a threat of job-loss in the subtext, real or perceived. It is important for the conversational interaction team to build trust with its partners early on in the process. A good way to do that is to start with win-win projects rather than automation projects that may lead to reorganizations.
与所有利益相关者保持一致特别值得注意。 这听起来确实像是另一个陈词滥调，但是我发现对于与自动化相关的项目，潜在的或真实的潜台词都存在工作损失的威胁。 对于对话交互团队而言，在此过程中尽早与合作伙伴建立信任非常重要。 一个好的方法是从双赢项目开始，而不是可能导致重组的自动化项目。
As said earlier, this article only skims the surface. However, hopefully it provides with a framework to think through the harbinger of a company’s migration to the new age.
如前所述，本文仅略述其表面。 但是，希望它提供了一个框架来思考公司向新时代迁移的预兆。
翻译自: https://medium.com/swlh/a-business-leaders-guide-to-chatbots-22e0b4ac1103
机器学习做自动聊天机器人
