http://artistdetective.com/arckon

You may have seen my conversational AI around in the Loebner Prize in past years. I have recently set up an online interface so you can test its capabilities.

The AI can learn, make inferences, remember context, interpret shorthand and pronoun references, detect basic sarcastic statements, compliments, insults, unethical requests, and of course you can try out the common Loebner Prize questions and Winograd Schemas. It’s more built for technical prowess than casual chatting, but the online interactions help me improve the conversational subroutine.

I’m still looking into more reliable hosting arrangements, so for the time being, Arckon is only regularly online on Mondays, irregularly on other days.

Nice looking bot Don. Must admit it’s a lot closer in approach to what I’ve been wanting and occasionally trying to do than most others I’ve seen. I think doing proper grammatical parsing and marrying that with a semantic triples base if the most likely way to get to a step change in what a chatbot can do. I like the way Arckon starts taking control of the conversation and starts asking for the related triples (cat name, cat personality), we’ve been using a similar entity prototype model for one of our bots to collect related information once a user introduces a new entity. Surprised that given “I have a black cat” it didn’t manage to generate the “cat hasColour black” triple though.

Nice work, will read and play a bit more.

David

Thanks, David. The initial questions are a bit of a default list like people use in small talk, but there is method behind all of it. Asking for the name allows for efficiently storing fact triples centrally under that name, rather than dispersing it under a multitude of descriptives “my cat”, “my pet”, “the feline” etc. Not much different for humans, I reckon.

Surprised that given “I have a black cat” it didn’t manage to generate the “cat hasColour black” triple though.

You have good reason to be surprised because it was designed to extract just that. The cause is that the online version redirects knowledge to a makeshift temporary memory, whereas open questions are directed to search the main database. As a result there is still some misalignment in what knowledge the program finds for a straight question (“is my cat black?”) vs an open question (“what colour is my cat?”). I’ll try and tie those ends together.

What are you using for your triples base? We started with the internal CS one but now use Apache Fuseki.

Oh, I made up my own. The database is just a structure of subdirectories with a text file per fact triple, which is pretty slow but flexible. As far as I’ve seen, the OWL ontology is the best closest thing.