Sep 27, 2021
Zooming in and out: lessons from dissertation writing and conversation design
tl;dr Zoom out! You’ll learn new things.
A dissertation is a long document. It’s the result of a lot of work, assembled into a single manuscript. It needs to be coherent at various levels of detail: the dissertation as a whole needs to tell a coherent story, but so does each chapter, each section within each chapter, and so on. It can be helpful to look at the dissertation at each level of detail, from 0 ft to 30000 ft. I think this is true of writing in general, including conversation design.
Zooming out in academic writing
In my 10 years in academic linguistics, not only did I study the structure of discourse as part of my research, I also taught college writing, linguistics, and benefited from the many resources at the University of Michigan supporting our work (the teaching center, the knowledge navigation center, the writing center, and much more). One habit that emerged from all this work was the value of looking at things from narrow to broad views.
At a practical level, writing papers (and eventually a dissertation) benefited from analyzing the document narrowly and broadly. From a narrow perspective, I may read a single paragraph at a time. From a broad perspective, I may print out all 20 pages of a draft and lay them out on the floor. The narrow perspective lets me analyze individual sentences and how they fit together. The broad perspective obscures individual sentences but can give a view of the document as a whole.
Much of what I’m discussing here can be captured by the pedagogical concept of a reverse outline, where people review an existing draft of a paper, summarize each paragraph, and then review those summaries. Doing this often reveals whether the sequencing of paragraphs as a whole makes sense. I have benefited from this so many times, discovering that one paragraph going into detail on a topic makes more sense as part of another section. To demonstrate, here are two views of the first chapter of my dissertation:
On the left is the standard page view we use while writing, with a single page of content visible on screen. On the right is a broader view, zoomed out such that individual words and sentences are no longer legible. This broader view reveals things you might not see in a narrow view. For example, it seems clear that the chapter starts with a lot of text, but later in the chapter there are diagrams and a table.
The degree of zoom is continuous. You could zoom out even more than the broad view above, do something in between the narrow and broad view, or go even narrower. Each level of zoom can focus your attention in new ways.
Zooming out in conversation design
Conversation design involves building conversation flows, dialogue logic, intent matching, writing copy, and more for the purposes of articulating the possible conversations one could have with an automated conversational agent (e.g. Siri, Alexa, a chatbot). In doing these tasks, the focus can be at various levels of detail.
Much like academic writing, the copywriting task is done at a fairly narrow level of detail. To write the words that a bot might say, you need to be focused on the words and sentences. But when it comes to designing and representing the structure of possible conversations you could have with the bot, the domain is much larger. You may start with a single turn of talk, and depending on the use case, that may be enough. But if the conversation is more complex (e.g. applying for insurance, triaging health symptoms, supporting complex conversational repair), you will need to assemble each of the turns into an overarching conversational structure.
A designer, like a programmer, may be expected to store that macro-structure in their own heads. That said, the tools that a conversation designer uses can make a difference. Does the conversation platform support zooming out to see the ENTIRE conversation structure at once? If yes, what information is accessible from this broad view? If no, what is lost by not having access to this broad view?
The narrow view may begin with a sample dialogue:
A: Hello! How can I help you?
B: More info about your company.
A: This company was founded in 2017. To read more, click here. Anything else?
B: That’s all. Thank you.
A: Come back anytime.
This dialogue can then be represented visually as a series of bot and user actions, including possible paths that were not included in the sample dialogue. An example from the book Conversations with Things:
These paths can then aggregate into a large-scale representation:
As the authors Rebecca Evanhoe and Diana Deibel discuss in their book, the conversation can be broken into components that themselves constitute the larger structure. For details, check out chapter 6 of their book. For our purposes, the point is that representations can be meaningful from a narrow view and a broad view, and the broad view may tell you something that you wouldn’t get from a narrow view. To demonstrate, what if zooming out showed the following instead:
This conversation still has a lot of content, but it is not as complex of a conversational structure.
How do some design tools approach this? Let’s take a look.
DialogFlow CX
DialogFlow CX, one of the Google’s conversation design tools, supports zooming. I imported one of their advanced pre-built agents (Financial Services), and get this:
It doesn’t all fit on screen. But I can zoom out!
While this pre-built agent may be “advanced”, it is not complex at this level of representation. This view does not include the structure inside each of these subcomponents, meaning you have to click on one to see what’s inside. But what if you could see them all at once?
One thing I see is that the last flow “Speak to an Agent” is much smaller than the others. I notice that some flows are more linear than others, e.g. the second one “Make a Payment” seems more linear than the third “Lost / Lock / Replace Card”. The colors of lines also seems to vary. If other types of information were being represented here, we may notice more differences.
Voiceflow
I created a new Voiceflow instance and started with a Chat type. After a short onboarding, I see the following:
There’s a lot happening here, and some of the components don’t fit on screen. But if we try to zoom out…
Look at that beautiful zooming! I haven’t pushed the tool with a huge conversation, but this is promising! Clicking on a step’s block brings up a menu where you can edit it. It appears that the home view includes the key information, but editing is done in a secondary window. As a result, the main view gives you a pretty solid overview. Voiceflow does support modularization with “flows”, where subsequences can be designed as single blocks. This makes reusing a whole block easier, which is awesome. It does make the flow’s internal structure invisible in the home view.
Voxable
Voxable is a newer conversation design platform. For now, it seems focused on the scripting side of things, but it does include a visualization option. And part of that visualization includes… zoom! From their “What’s New” page, it looks like this might have been added September 8, 2021. Happy to see that they think a zoom option was important enough to include!
OpenDialog
OpenDialog takes a somewhat different approach to conversation design and visualization. While the other tools prioritize a flow chart metaphor for the design representation, OpenDialog’s metaphor is more about conversations-to-be-had at various depths of conversational context. Instead of prioritizing a tree structure that you could step through linearly, it supports hierarchical embedding of contexts. You can still step through in a linear order (with transitions from one point to another), but the design view focuses on the components at one level of hierarchy at a time. For more on OpenDialog’s distinctive approach to conversation design, check out a recent article by Maaike Coppens, OpenDialog’s Chief Design Officer.
I started my OpenDialog instance with a template. The starting view is at the top level:
OpenDialog structures designs into a hierarchy of elements: scenarios, made up of conversations, made up of scenes, made up of turns, containing intents. At the top level, the scenario name is in the middle with embedded conversations in a surrounding circle. The contents of those surrounding conversations are not visible. If you click on one of the conversations, it launches the conversation view. You can keep clicking inside elements to launch those embedded views. Along the way, a context list is generated with clickable elements that take you to the corresponding part of the hierarchy.
To try to capture all levels at once, I assembled a set of screenshots into this image:
This image shows the top-level scenario, and one part of each level expanded. This view isn’t currently possible within the tool itself. I had to make it manually by clicking inside each one and taking a screenshot. And to think about the complete design, you need to remember what is happening at other levels while looking at one level. While currently zooming isn’t available, I have heard from folks at OpenDialog that more advanced model navigation is something they are working on, including zooming!
OpenDialog writes that they have a context-first way of defining conversations. Having spent some time with this tool and thinking through this approach, I realize it’s a different way of planning, visualizing, and designing. I see huge potential, but also feel the challenge of thinking things through differently.
Reflections
There are so many other tools! I would love to see how zooming is approached elsewhere. But even if a tool you use doesn’t have a zoom option, or doesn’t show everything in a zoom, one thing you could do is construct an enriched zoomed-out view manually outside of the tool (as I did above for Dialogflow and OpenDialog).
Also, I encourage you to try out these different tools for yourself (OpenDialog, Voiceflow, Voxable, Dialogflow)! It’s an interesting challenge confronting the constraints and affordances provided by each one.
Conclusion
Zoom out! You’ll learn new things.
Copyright 2021 Joseph Tyler All Rights Reserved
https://josephctylerwords.medium.com/zooming-in-and-out-lessons-from-dissertation-writing-and-conversation-design-d2d204deb16b
