Conversational AI roBots : Pragmas & Practices - 21 Lessons one could learn watching the wise Yoda Bots !
I have been thinking about (and working on) Conversational AI from multiple perspectives. Am co-authoring a book on AWS Connect/Lex titled “AI and NLP with AWS”, working on a Tutorial “How to train you Conversational AI to pay Attention”; and starting projects on Predictive Contextual IVR including Visual IVR (very preliminary stages, but one needs to think things through before forging ahead)
As a preparation, as Tom Hanks says “know the text & have a head full of ideas”, what best way to “know the text”, than a binge reading marathon - of books on Bots, Voice UI, Linguistics (Pragmatics, Discourse Analysis, and Conversational Implicature), Transformers, BERT, Stanford xcs224n(completed), cs224u,.. ? In short an accelerated course on Computer Theoretic and Linguistics !!!
I have the list of reference materials at the end of this blog — the wise at work !
This will probably be a three part series — This one, on lessons, 2nd blog on Conversational Implicature (way interesting and extremely relevant for Conversational Bots — I discovered Paul Grice, man !) and the concluding post to bring them all together … let us see.
Intuition
.
Conversational AI is a continuum — from a simple command interface with very limited vocabulary to the sentient “JARVIS-like” conversation …
.
.
On another level, the Computer Theoretic has (to some extent) taken over the NLP/NLU space, which was previously dominated by the Linguists …
.
.
But as David Ferrucci points out, we are far away from real understanding …
In short, we have a long way to go, and it needs Deep Learning, Knowledge Graphs and definitely Linguistics.
Leaving the abstract aside, let us now focus on the pragmatics …
21 Lessons one could learn watching the wise at work !
1 Think of interactions as Available, Smart and Autonomous
2 Focus on self-learning solutions that understand users and react to them in the same way as a human agent
3 When thinking about Bots, think Search, not a single Question-Answer session. This requires a little deep thought — the bot should want to understand what the user wants, participate in an iterative collaborative interaction to facilitate what the user is looking for
4 A bot (chatbot or voice or even the Visual IVR) is an experience. Which means, develop a Minimum Lovable Product — it shouldn’t be too ambitious and complex with lots of features; otoh, it shouldn’t be too narrow that it doesn’t serve a good purpose completely. Keep steps small & show value; fulfill a need
5 There is a misconception that bots do not have visual branding. The conversational UX, as a “transparent” user experience, still provides a good amount of visual aspects that impact the branding of your bot.
6 Incorporate as much small talk as possible — it is a big success. Questions like How are you? Who are you? Are you a bot? Are you married? Who do you work for? Who pays you? I’m not getting paid so well [wink emoji] What do you like to do/eat/drink? Do you have a boyfriend?Do you like me?Are you real ? What is the meaning of life ?
7 Monitor, adapt and improve — the utterance monitoring log is your friend. Use analytics extensively to ask questions like the user’s learning curve — How long does it take for the user to understand the scope of the bot?
8 Visual designers spend time on aesthetics, and like them, conversation designers spend a lot of time writing content and functionality that fits the scope and audience of your bot. We do not speak the way we write — so you can’t literally turn your web site into a bot
9 Build users trust and confidence slowly
10 Think about engagement trio — Attract-Engage-Delight. Be very crisp and clear about your bot’s purpose & core functionality
11 Bots are very iterative — you have to keep on improving as you learn as well as the users — Expose users to new functionalities, continuously; a welcome area with talks about new features is a good idea; or suggest new contextual features for original questions
12 Surprise your users with Acuity and Serendipity !
13 Manage the scope, complexity and feature creep very carefully — a confused user won’t return
14 Be careful about data overload — don’t show 100 flights — remember collaborative search ?
15 Dialog Flow Construction — Don’t reproduce your poor business processes
16 Humans subconsciously will map personality to your bot — so better to plan for one (… or more) personas than having accidentally attributed to one. While designing, ask persona questions like What’s your bot’s job description? Experience ? Education ? What it thinks ? How does it feel ? Who are your bot’s clients, target group? Professional or leisure users? Age range of your users? Gender ? Geographic spread ? What languages do they talk? Should your bot be multilingual? Do they use abbreviations or emojis?
17 Once you have defined a personality, it is important to keep it consistent across the experience. This gives the users the feeling that they are dealing with a cohesive service (or a persona), which in turn improves trust and engagement.
18 Build multiple personalities, if possible. Sometimes you don’t have the option e.g financial or healthcare bots need to be formal, relaxed and trustworthy. Frequently refresh the various common replies with new ones
19 Think of Bots as a gateway — an interface to services; you still need the backend to fulfill the services — queries et al
20 Model paths other than happy paths and fallback choreographies — fail gracefully. The idea about collaborative search is relevant here. The Happy path is the MVP, i.e. the basic structure around the perfect conversation, but that is rarely achieved !
21 Plan for learning from real users and improve your bot. Once your assistant is able to handle a few happy path stories, it is time to let it loose into the real world to steer the direction of the development !
22 Follow Grice’s maxims — Which, of course, is the topic of the next blog … on it’s way …
References — the wise Yodas:
- Pragmatics, by George Yule
- Discourse Analysis, by Brown and Yule
- Presumptive Meanings: The Theory of Generalized Conversational Implicature, by Stephen Levinson
- Back to the Future for Dialogue Research: A Position Paper, Philip R Cohen https://arxiv.org/abs/1812.01144
- https://blog.rasa.com/the-rasa-masterclass-handbook-episode-1/ The Rasa blogs talk about 5 Levels of Assistants
- Growth Through Transformation by Derek White at U.S.Bank’s Investor Day https://ir.usbank.com/events/event-details/us-bancorp-investor-day-2019
- Hands-On Chatbot Development with Alexa Skills and Amazon Lex, by Sam Williams
- Building Intelligent Chatbots with Amazon Lex & Amazon Polly, by Adam Larter https://www.youtube.com/watch?v=PLnRzHNmcao
- Designing Bots, by Amir Shevat
- Designing Voice User Interfaces, by Cathy Pearl
- Voice UI Systems, by Ann Thymé-Gobbel and Charles Jankowski
