Chatbots – Notes on the Landscape Ahead


Wind up TeethThere has been much discussion around the conversational interface as the next runtime.  Currently, there are pockets of innovation scattered throughout the tech industry.  At Zeal we are currently focused on the “bot” extension of this runtime paradigm, specifically in developing our chatbot Ava.  We strive to create an efficient, and highly usable method of data collection via the chatbot paradigm.

This month I had an opportunity to facilitate a session at the Earley Knowledge Salon.  It was truly a pleasure to be part of such a great event.  Having a chance to discuss ideas with bot thought leaders and distinguished AI veterans was truly refreshing. If you haven’t heard about Earley Information Science, you can check them out here: http://www.earley.com.

The logistics of how the conversation took place is interesting.  I had prepared a list of interesting points about “conversational commerce” for discussion. As the conversation started naturally, it became abundantly apparent that the room wanted to discuss more fundamental issues around chatbots.  So, we went with that and had quite a discussion.

Summarizing some of the salient points of the discussion, and the structure of the conversation:

  • bot personality
    • development of personality based on prior human support / storyboards
    • gender of bots
    • cultural appropriation and bots
    • conceptual avatars
  • chatbot translation
    • the comp-sci problem
    • mechanical turk possibilities
  • context and space awareness
    • what is context
    • what is space
  • expectations of privacy and security
    • context based privacy and security
    • space based privacy and security
  • conceptual grounding around linguistics, grammar and meta-grammar
    • meta-grammar
    • space aware grammar
    • context aware grammar
    • code switching

Bot Personality:

We explored the idea of script-first development vs. personality-based on known interactions.  In the case of chatbots that don’t initiate a conversation to elicit a response, it was agreed that you have to start with a known set of interactions.  This starting point leads to better understanding and adaption around what will be asked of the bot.  Actual human support is a starting point.  Storyboarding is a technique for refining the interaction.  It was mentioned that even Pixar has been contracted to work on chatbot personality.

There is a proliferation of female bots.  While it seems natural to assign a gender to a chatbot, what does this mean for the enforcement of gender stereotypes?  What about human interaction?  It is easier for a human to converse with and avatar that has the attributes of life / humanity.  Gender is an attribute, but what are the ramifications of assigning a gender to a chatbot?

Cultural appropriation in bots is a slippery slope. Language has cultural elements, what are the tasteful and ethical extensions of cultural grammar.  When is it appropriate and when is it inappropriate? In the case of space aware and context aware bots coupled with meta-grammar and the ability to linguistically adapt to space and context machine learning gets risky.  When does a learning bot become a pantomime of the culture that they interact with?

Avatar was agreed to be the appropriate concept for a bot and the presence that they represent. In retrospect, this is interesting, because typically an avatar is defined as a projection of an independently self-aware entity. In the case of a chatbot … what is being projected?

Chatbot Translation

The white elephant.  It was universally agreed that this is an issue without resolution.  Computer Science has been struggling with this for years… “like a wave crashing against the rocks.” The group came to the conclusion that without being truly aware (as in conscious / self-aware) translation will continue to be a challenge.  Because interactions with chatbots are designed to emulate human interactions, auto translation (google translate etc.) are a disaster. This “problem” leads to great expense in multi-lingual chatbots.  The standard approach is to write multiple bots with language specialists. One way of democratizing the cost might be to contract the language expert as a final decision maker and farm out the grunt work to a mechanical turk.  However there is still the problem of linguistic structure, and how the language is parsed at a high level.  Even with a meta-grammar layer above the translation, this would be difficult and require complex mapping.

Context and Space Awareness

Context awareness in chatbots indicates they know who they are talking to.  This operates at a deeper level as they gain conversational state, and vary behavior based on context.

Space awareness in chatbots indicate that they know “where” they are talking.  Examples could include:

    • public / private conversation
    • channel
    • chatrooms
    • VR space
    • … and combinations thereof

Expectation of Privacy and Security

At some point… preferably early (for example, at on-boarding), is it incumbent of the bot to indicate whether or not the conversation is anonymous? What is being collected? There is an open ethical field here.  Not much has been defined in terms of right or wrong, legal or illegal. A case was sighted in which a chatbot was determined to have free speech as an extension of the corporation.As a collection mechanism what are the normal expectations of privacy? Security? We discussed a bit around assigning attributes to users, and how those might be used as filters later. The same concept could apply to space as well.

Conceptual Grounding

Throughout the conversation the group continually rooted itself around some ideas around grammar:

  • meta-grammar – grammar that operates at a level above grammar – defining concepts
  • space aware grammar – grammar that adapts to space.  Each space has a grammar unto itself
  • context aware grammar – grammar that adapts to context (target – who am I talking to)
  • code switching –  the practice of alternating between two or more languages or varieties of language in conversation.

We covered a lot of ground here.  Not exactly “conversational commerce” (well not at all). But, the audience seemed pretty excited about the discussion.  Technologists will spend years iterating on these very topics, and already have in other forums. My hope is that supplying some notes will help others with their thinking.

Leave a Reply

Your email address will not be published. Required fields are marked *