Chatbots provide for natural language computer conversations, popularized with personal voice assistants like Siri and Alexa. Suitable architecture frameworks for building chatbots are reframed into the “hospitality” of a development platform. This more human sounding hospitable architecture recognizes the unique embedded “uncertainty” of these chatbots, their complex natural language understanding (NLU) components, speech-to-text challenges, mapping of colloquial terms to popular user terms, and even understanding sarcasm. The supporting frameworks exhibit unique and complex attributes, making it hard to pick a platform.

The paper analyzes and helps choose between three popular chatbot platforms: (1) Watson Assistant, (2) Dialogflow, and (3) Lex. Four traditional architecture quality attributes (QAs) are contrasted: (a) modifiability, (b) security and privacy, (c) interoperability, and (d) reliability. A scoring approach is used to help evaluate how “hospitable” they are to those various QAs desired of the different end products. The QAs are further divided into architectural tactics, such as deferred binding to support modifiability (a); keeping chatbot conversations private (b); supporting multiple data formats to support interoperability (c); and restricting chatbots from responding with low confidence (or failures) for reliability (d).

While these are traditional architecture evaluation approaches, the paper shows specific aspects that are more applicable to chatbot architectures. An example of selling fruit through an online interactive natural language interface is described. A feature card is used to define features and collect various stakeholder opinions. These include the desired feature (such as ability to externalize response generation), platform, status, criteria, decision, and reason.

The paper’s significant contribution, and most applicable to other types systems, is in the process to develop and use a hospitality index. Mixing in QAs, tactics, features, and candidate platforms, an index is calculated. Relative weights for desired features is also mixed in. As with the example, the three frameworks are analyzed, resulting in a hospitality index for tactics and another for QAs.

While acknowledging other systematic architecture evaluation approaches, such as the architecture tradeoff analysis method (ATAM) and the software architecture analysis method (SAAM), the authors say their approach can convert the analysis into critical design decisions. By linking between platform features and QA impacts, the result is basically a more executable architecture evaluation approach.