Capturing the Implicit – an iterative approach to enculturing artificial agents

Capturing the Implicit - an iterative approach to enculturing artificial agents, Peter Wallis & Bruce Edmonds, CASA@IVA, Edinburgh, August 2013. slide 1

Capturing the Implicitan iterative approach to enculturing artificial agents

Peter Wallis and Bruce EdmondsCentre for Policy Modelling

Manchester Metropolitan University


The Issue: Producing Social Actors

• Computational devices and software increasingly intrude into the human social sphere (e.g. conversational help systems, satnavs, robot vacuum cleaners, telephone answering systems)

• Most (if not all) of these are, at best, socially inept and, at worst, downright rude

• To produce socially competent humans takes (roughly) 20 years of effort by society

• This paper looks at how one might produce a socially competent artificial agent


The Centrality of Social Intelligence

• We are a fundamentally social species• We achieve most of society’s products collectively• The intelligence that organises these

achievements are both cognitive (between brain cells in our brain) and social (between individuals)

• Indeed the “Social Intelligence Hypothesis” suggests the crucial evolutionary advantage our brains give us is in terms of social abilities

• Our “individual intelligence” is a side effect of our “social intelligence” (not the other way around)


The prevailing social “autism” of artificial devices

• Given the centrality of social intelligence and interaction to human life…

• …it is very odd that our artifacts are designed with individual intelligence as central and any social attributes as an afterthought (if they have any)

Some reasons for this might be:• The prevailing understanding of computers etc. as

tools to be engineered for a specific purpose• The gulf between AI/cognitive science and the

social sciences• The sheer difficulty of engineering social abilities


The Importance of “Folk Theory” in communication and co-adaption

• The classic model of language involves a set of signs with fixed meaning

• Instead, signs are largely interpreted in context and part of the continuum from action to communication

• We humans model each other in order to figure out what a conversational partner really means

• Using a common understanding of how other people’s minds work: intention etc. (“folk theory”)

• … and we work hard at it


Eggins & Slade (1997) on sequential relevance

A: What's that floating in the wine?

B: There aren't any other solutions.

"You will try very hard to find a way of interpreting B's turn as somehow an answer to A's question, even though there is no obvious link between them, apart from their appearance in sequence. Perhaps you will have decided that B took a common solution to a resistant wine cork and poked it through into the bottle, and it was floating in the wine. Whatever explanation you came up with, it is unlikely that you looked at the example and simply said `it doesn't make sense', so strong is the implication that adjacent turns relate to each other.”


The Difficulty of Accessing Social Knowledge

• Most social knowledge (how to productively interact with other humans) is implicit – that is, the people wielding that knowledge do so without “thinking about it” (in the sense of conscious explicit thought)

• Thus people have difficulty explaining how they do social things (indeed they are often unaware of these skills but simply adapt/use them)

• Social knowledge is extensive, messy and highly context-dependent

• High levels of social skills, with respect to any particular social, are only acquired by humans by living within that culture for many years: adapting, making mistakes, observing others, interacting…


Kinds of Utterance

Paul Seedhouse (2004) summarizing the outcomes of Conversation Analysis (with added numbering):

“a conversational partner’s utterance will go

1. seen but unnoticed

2. noticed and accounted for, or

3. risk sanction”

• Things that are accountable for and polite, have a role and do not necessarily consciously stand out

• Anything that is not easily accountable for sticks out and would normally trigger further dialogue to understand them


Aim of our approach

• To implement a feasible approach – not aiming for perfect social abilities, but at least to make artificial entities less socially inept

• To start with an existing entity that exists within the social sphere, and make some improvements

• To produce an approach that mimics (in a small way) some of the enculturation that humans undergo, but leveraged by technology to elicit the most social knowledge possible (which will be limited)

• As a consequence understand more about making and deploying social devices


Principles of the Approach

• A iterative approach – stepwise improvement of social knowledge and abilities (as opposed to a one-off analyse-specify-design method)

• Each interation based on records/logs of device in situ, interacting with humans in current state

• Relying on the shared “folk theory” of cognition concerning what normally people do and why

• Also that what goes wrong in interactions is much easier to spot that what goes right

• Facilitating a human (in a natural, non-CA-expert mode) to annotate logs with social knowledge

• Implementing a repair strategy to fix these faults


Diagram of a Standard NL Approach


An Ethnomethods Approach


Leveraging Common Knowledge of Folk Theory


A Summary of Our Approach


Towards facilitating social annotation using a “social” workbench

The workbench prompts the “folk theorist” user to notice bits that seem wrong with the conversation and then using a series of pre-scripted prompts tries to capture what the user thinks is wrong in a slightly constrained (i.e. somewhat standardised) way.


Machine conversation (circa ‘97)

Machine: G'day mate.

Human: G'day. How's it going?

Machine: My mother once said "How is it going is a mystery." She was a bit nuts.

Human: Uh-huh.

Machine: Are you still learning to speak, my friend?

Human: NO, no. But your anecdote took me a bit by surprise.

How was your morning?

Machine: Oh well, that's a pity!

Human: I'm sorry; what's a pity?

Machine: Apologies accepted.

Human: I feel compelled to insult you, and we've only been talking a minute.

Machine: We are bloody well talking, aren't we?

Human: Oh baby. We sure are. So here are some keywords for you:shit mother baseball.

...


(Lots of) Further Issues

• Work still very much in infancy• At the moment it takes an expert engineer to

interpret the knowledge elicited by the system in terms of repair to the original conversational scripts

• The best form for the database of social knowledge about the interactions needs to be established

• How such an iterative elicitation approach can work alongside traditional scripting NL modelling approaches and models is unclear


The Future?


Thanks!

Peter Wallis

http://cfpm.org/~peter

Bruce Edmonds

http://bruce.edmonds.name

Centre for Policy Modelling

http://cfpm.org

Capturing the Implicit – an iterative approach to enculturing artificial agents

Education

Transcript of Capturing the Implicit – an iterative approach to enculturing artificial agents