Properties of wit ai vs API ai

Chatbots - Chatbot Platforms

There are many chatbot platforms on which it is possible as a company to create your own chatbot. The most frequently used examples with a chatbot connection include Facebook, WeChat, Slack, Telegram, Skype and Kik (Kemp 2018, p. 60). So-called APIs are given to give companies the opportunity to offer services in the form of chatbots on such platforms. APIs are programming interfaces that allow external program code, such as a chatbot, to be integrated into existing software, for example a messenger platform. One should keep in mind that not all companies have the expertise to build their own bot and integrate it into a messenger platform such as Facebook (Gentsch 2017, p. 88).

Concrete examples

Facebook

Facebook is one of the most widely used social media platforms and has the highest number of monthly chatbots with over 300,000 (Kraus 2018). Facebook launched the messenger platform with chatbots on April 12, 2016 during the F8 conference. This also includes the use of the Messenger Send / Receive API. This is necessary to connect the chatbots to Facebook Messenger. Several functions are provided by this API. As can be seen in Figure 11, the chatbot can, for example, send text messages and pictures, receive text messages from the user or give the user the choice between different response options. These can also contain pictures. All messages are visually displayed to the user in the form of speech bubbles. The user must then select one of the answer options so that the conversation can continue. This can prevent incorrect entries that the chatbot would not understand. A welcome text can also be set and there are third-party plugins for the use of other websites or user names. With the help of websites such as Wit.ai or Dialogflow.com, it is also possible to create complex bots that are able to recognize natural languages ​​and continuously improve this recognition (Marcus 2016).

Figure 11: Facebook Messenger sample chatbot
Source: La (2016)

Through direct interaction with people, the chatbots can, among other things, provide the user with automated subscription content, such as the weather report, news, information about the traffic situation or sports results. However, the user can also receive certain documents or information, such as receipts or dispatch notifications. Some chatbots are even able to actively advise the user. According to Facebook, people come first in their philosophy. This is why the user has the option at any time to block communication with a chatbot that is perceived as unusable (Holotescu 2016, p. 91). To prevent spam messages from being sent, Facebook is introducing important rules. The chatbot must not offend anyone, it must follow the instructions of the person and protect its own existence, as long as this protection does not conflict with the first two rules. In addition, every chatbot must be approved by Facebook before it can be published (Constine 2016).

In order to implement such a chatbot, Facebook provides several aids. The first point of contact is the introductory documentation for the messenger platform, where information on various components can be found. The messenger platform offers a bundle of APIs, web plugins and a full web view, which enables good integration. In addition, Facebook offers best practices for its chatbots to make it easier for developers to get started (Facebook 2018g). Many important functions are explained in the documentation. For example, it is possible to make payments using a Facebook chatbot. This function is currently only supported in the USA in cooperation with the online payment service PayPal and is currently still in the beta phase (Facebook 2018i).

Facebook offers developers the opportunity to reach more customers through advertising, sponsored messages to customers or a customer chat plug-in on their own company website. Chatbots are also displayed in the “Discover” part of the Facebook Messenger. With the help of so-called messenger codes, which work like QR codes, but are visually structured like a robot, developers or companies can share their chatbots (Facebook 2018c).

Another function is the integration of web views in Facebook Messenger. Since some features of the company's own website are difficult to implement using speech bubbles, such as selecting products to buy, places to book or appointments to reserve, Facebook has created a way to integrate the web views of its own company website into Facebook Messenger. This opens many doors for companies (Facebook 2018h).

Figure 12: Facebook web views in Messenger
Source: Facebook (2018h)

For some processes in the messenger it is necessary to establish the identity of the user and to authenticate him. With the help of various APIs, it is possible, for example, to find and contact current customers of the company website using their phone number and name in Facebook Messenger. It is also possible to use APIs to perform web-based login processes in order to authenticate the identity of the user and to link himself to his messenger account. This would theoretically enable the user to perform account functions in Messenger that are normally only available on the company website (Facebook 2018d).

Customer analyzes are possible on the basis of customer demographic data. The developers have the option of integrating third-party analysis software into the messenger chatbot and continuing to use it. In addition, users can also rate chatbots and thus signal to developers or other customers whether and how mature the chatbot is (Facebook 2018f).

A feature that is currently still in the closed beta phase is the use of augmented reality (AR). These can be added to a chatbot to target users in new and interactive ways. It is possible to switch to the camera app in a conversation, to use AR effects to try on suggested products, such as clothes or to position furniture in the room (Facebook 2018b).

With the help of APIs for Natural Language Processing (NLP), the chatbots can automatically recognize and understand important information from the user's natural language messages. Due to the many different language styles and dialects, NLP represents the biggest hurdle for a chatbot. Different entities such as greeting, date, telephone numbers or location can be recognized in around 20 different languages. At this point, it is now possible to expand the NLP features by linking the chatbot to other platforms that use the machine learning algorithms to interpret and assign user text inputs (Facebook 2018e).

WeChat

Even if Facebook is the most widely used social media platform internationally, there are countries in which other platforms dominate. In China, for example, platforms such as Facebook, Whatsapp, Youtube and Twitter are banned or can be used with severe restrictions (Zeit.de 2017). As a result, another messenger called WeChat has established itself. WeChat is the most widely used social media and messenger platform in China. In the second quarter of 2018, WeChat had over 1.082 billion active users (Tencent 2018). WeChat is a messenger service from the Internet company Tencent for smartphones that, similar to WhatsApp, can be used to send messages and make voice and video calls. However, WeChat offers its users many more options, as the platform has been providing chatbots for several years. In Messenger, users can, for example, book taxis, buy tickets, order food, transfer money or share documents with other users. In China it is common to use voice input instead of text input because of the time savings (Chan 2015, p. 2). Figure 13 shows conversations with different chatbots in WeChat Messenger. The user asks, among other things, for a specific restaurant and a specific route.

Figure 13: WeChat - conversations with different chatbots
Source: (Graziani 2016)

WeChat distinguishes between two types of chatbots that a company can set up: subscription accounts or service accounts. Subscription accounts are used to publish content and share it with your users or subscribers. The subscribers often represent potential or existing customers. They usually receive automatic updates on new information or articles once a day. These subscriptions are good for users who want to stay up to date. Companies, in turn, can, for example, analyze how many users have seen an article and can thus draw conclusions about which content is liked by which target group (Jerry 2017). Service accounts, on the other hand, are a good way to handle customer service via WeChat and to get in touch with customers. Fashion brands, airlines, restaurants, hotels, and e-commerce stores often use chatbots to support customer service. For example, with the help of such a chatbot, the user can obtain information about current promotions or events as well as the location of the nearest restaurants. They can also process more complex inquiries via the chatbot. At this point, however, the developers admit that a real employee often has to intervene in order to be able to answer more complex queries. With the help of AI, this could change in the next few years (Jerry 2017).

The benefit that results from the analysis options of WeChat for companies is not insignificant. WeChat offers companies several options. On the one hand, a follower analysis is possible, in which it is analyzed how many new followers have been added and how many have dropped out over a period of 30 days. This analysis also takes into account properties of the followe such as gender, language, geographic distribution, operating system used and the device used. Another possibility is to analyze the reach of broadcasts. This method shows how many users have seen a broadcast, how many of them have read the broadcast, and how many users have shared this broadcast. In addition, it can be shown from where the reader got access to the broadcast. Sources can include direct broadcasts from the public WeChat account, dialogues with the public WeChat account or dialogues with friends on WeChat. Another possibility is the analysis of menu usage, in which figures such as the number of clicks, the number of WeChat users who clicked the menu or the average number of clicks per WeChat user are determined. Companies can also analyze incoming messages and measure average times for certain activities in milliseconds (Liu 2018, pp. 81-84).

WeChat also offers companies the option of creating a WeChat website that can be better integrated into the WeChat app and accessed from there. It should offer companies an ideal mobile internet place for company presentations and support mobile marketing (Liu 2018, p. 183).

However, some functions are only available for the Chinese version of WeChat, which now requires registration with the real name and thus makes anonymity on the Internet more difficult (Liao 2017). A chatbot can be written using phyton, for example. (Nosotti 2017). WeChat also offers other functionalities. WeChat enables users to discover people or friends nearby. You can send virtual messages in a bottle that other users around the world can collect and read. This function is intended to support companies in marketing their chatbots (Liu 2018, pp. 46-48). WeChat also offers its users a payment system called WeChat Pay, which allows them to make cashless payments at the checkout using a QR code generated on their mobile phone. To do this, your own bank account must be linked to the WeChat account. You can also use the WeChat app to pay for items when shopping online on your mobile phone (WeChat 2018b).

Tencent promises its users to protect users' personal information and not to sell user data to third parties without permission. In 2016, WeChat received the ISO / IEC 27001: 2013 certificate, an international standard for IT security. The standard includes, among other things, requirements for the establishment, implementation, maintenance and continuous improvement of a documented information security management system. It also contains requirements for assessing and handling information security risks. However, data security is still a major market entry barrier for WeChat in Europe, since Tecent, the largest Chinese Internet company in the West, is under general suspicion of espionage for the Chinese government (Liu 2018, p. 11). The communicative exchange via WeChat is also subject to a certain degree of censorship. Users must agree, among other things, not to criticize the Chinese government in WeChat. Furthermore, WeChat does not currently support end-to-end encryption (Appvisory 2018).

Integration of machine learning

In order to improve the recognition of natural language, various platforms that specialize in ML processes for chatbots can be integrated into the chatbot. The reason for this is that speech inputs can vary greatly and unsupervised methods of the ML can cluster and assign these inputs without having to be defined beforehand. The best-known platforms include Wit.ai, Dialogflow, Luis and IBM Watson. All platforms are able to connect to other chatbot-enabled social media platforms such as Twitter, Slack or Telegram (Jose 2017).

Wit.ai is a software-as-a-service platform that was bought by Facebook. Wit.ai is a platform without a visual development environment that works with entities and intents, among other things. Intents stand for the intention that is read from the user input (Figueroa 2017). Entities describe details and can, for example, be certain types of products, such as a certain type of pizza. You can fall back on predefined entities or define new ones. In order to recognize these entities and intents, Wit.ai uses certain ML learning methods, which are not described in more detail in the documentation (Wit.ai 2018).

Dialogflow, in turn, is a platform that has been owned by Google since 2016, but also provides connections for other social media platforms. In contrast to Wit.ai, Dialogflow contains many ready-made models that can be used without having to write complex codes yourself (Figueroa 2017). The platform is used by companies such as Domino’s, ticketmaster or KLM Royal Durch Airlines. Dialogflow uses entities, parameters, speech-to-text, text-to-speech together with machine learning algorithms that continuously train the chatbot in the background (Jose 2017). Dialogflow also uses Google's resources and infrastructure. The chatbot learns both from the training phases defined by the developer and from the language models integrated in Dialogflow. Based on this, an algorithm is built that assigns the user's expression to a specific intention. With Dialogflow, you can choose between two different machine learning modes. There is a pure machine learning mode without predefined rules and a hybrid mode. In hybrid mode, the entries are first compared with a database and if there is no match, ML is applied (Dialogflow 2018a).

Another option is the Luis platform, which is provided by Microsoft. According to Microsoft, Luis should be able to support chatbots with cloud-based ML models without the developers themselves needing previous knowledge of ML. It is offered as part of the Microsoft Azure Cognitive Language package (Sirosh 2018).

IBM Watson is one of the AI ​​platform that can understand all forms of data, interact with people and learn from these conversations. IBM Watson is deeply involved in deep learning and offers a deep learning service within Watson Studio that enables neural networks to be visually designed and your training runs to be scaled. Various deep learning frameworks such as Tensorflow, Keras, PyTorch or Caffe are supported. In order to train the neural networks in parallel, powerful graphics cards from NVIDIA are used (IBM 2018).

PyTorch and Caffe are open source deep learning frameworks from Facebook that are designed to bring together research and production in the field of AI. The Open Neural Network Exchange format, which was developed by Microsoft, Amazon and Facebook, is also supported in this framework. Automatic translations on Facebook, for example, work with such frameworks (Bohn 2018). For the purpose of such research, Facebook introduced Facebook Artificial Intelligence Research, which deals with the research of theories, algorithms, applications, infrastructures in the field of deep learning and speech recognition (Facebook 2018a).So there is a very goal-oriented, open source and cross-company research to be observed, which only has rapid progress in the field of AI as its goal.

WeChat also offers the possibility to implement an AI. WeChat uses machine learning algorithms for natural language processing, image and speech recognition and translations (WeChat 2018a). However, since the documentation is in Chinese, a more detailed description is not possible at this point.