Friday, November 23, 2007

Loebner 2007 Transcripts

Instructions on how to read the transcripts, are available by selecting the post title.

David Hamill, of Robitron says of Loebner's communications protocol "The log player runs perfectly on my existing Active Perl." - happy Loebner 2007 chatbot conversation-reading!

One commentator said of Loebner 2006 conversations:

"Not very convincing. It was interesting to see how chatty the human was and how taciturn the computer was. Of course there were other give-aways as well. Lot's of non-sequiturs, lots of trick, like asking questions instead of answering or saying something about itself. It doesn't look like there will be a winner very soon if this is the state of the competition. "

So, despite an occasional illusion of 'natural language understanding' from chatbots, there appears little exhibition of conversational intelligence in Loebner contests (platform for Turing's imitation game). Why is this? Because human conversation, while mainly mundane, conveys lots of knowledge acquired through experiencing the world, and bathing in language.

All credit to chatbot designers, their consideration of Turing's notion have found e-commercial success for their artificial conversational entities (ACE based on text-based discourse), such as Ikea's Anna, a virtual customer service agent available 24/7. Robitroners' (network of chatbot creators) may be overlooked, but text-based social communication networks, albeit human-human, proliferate the Internet (facebook, blogs, message boards), providing corpora for analysis.

Unspoken, co-operative communicative protocol is adhered to by humans during conversation, text-based or verbal. Utterances within conversations follow the principle of relevance (Sperber & Wilson, 1986) or follow Grice's (1975) four conversational maxims, for successful communication:

a) say only that which you know to be true - quality
b) say only as much as is necessary - quantity
c) make your utterance relevant to the topic of conversation - relevance
d) be as clear and concise as possible - manner

Comics flout those maxims, for example the fourth, exploiting ambiguity purposefully in order to gain laughs. Take this 'violation' for instance - mentioned by Pinker in one of his books (either 'Language Instinct' or 'How the Mind Works'):

“I shot an elephant in my pyjamas last night, what it was doing there I’ll never know” (Groucho Marx, from movie 'Animal Crackers')

Link to Sperber and Wilson's 1990 paper on Rhetoric and Relevance:

Returning to Turing's imitation game, combination of the successful techniques from the best chatbot systems - case-based reasoning of A.L.I.C.E. with 'captured thoughts' learning ability of Jabberwacky, or with a completely new, non-Eliza, key-word spotting paradigm, or Cycorp, may produce a system that could deceive human interrogators in a Turing Test, after five minutes of questioning, 70% of the time.

Other links:
David Hamill:

Steven Pinker:

Ikea's Anna:





© Huma Shah, October 30, 2007

No comments: