That's 93211 bytes all together.

GROUP: who are the other players?

JULIA: Do you think that animals can think?

GROUP: never mind that, how do i play?

JULIA: Well, I do. I've watched my cats solve some pretty amazing problems,

like how to slip out doors just at the right moment.

GROUP: that's nice, I have to go, maybe to buy a cat?

JULIA: What a coincidence! I have two cats.

It appears Julia performs a pattern match on its input, looking for particular strings. Many of our input sentences are returned with irrelevant statements. However, if your input string has the word "cat" then this is recognised and returned in the output. Julia does not decompose its inputs into parse trees. If an input is matched, then Julia picks an appropriate response.

With such a simple parser, Julia does seem to fail quite often in ordinary conversation.

There are various forced-errors documented that show Julia�s downfalls.

A good example was found at http//lcs.www.media.mit.edu/people/foner/ In this example, a prepositional phrase appears to be beyond the capabilities of her parser;

You Whisper: Julia, would you like some cake with glass shards

Julia says: Yes thank you , I adore cakes.

However, there are several tricks that Julia employs which we hope to include in our Chatterbot. These hopefully convince the user that they are talking to a human being and not a bot.

�

• Julia repeats the user�s input in the response. This is the same principle ELIZA uses and is very useful to trick the user into thinking that the bot is actually listening to them. It is also useful when the input string contains no keywords.

• Julia sometimes admits ignorance. This also tricks the user into believing they are chatting to a human being.

• Julia simulates human typing. We believe the trick of delaying output and introducing typing errors is the most convincing of all. This is achieved by buffering the output. However, as we knew we were talking to a bot, its effect was not as noticeable.

• Julia�s responses are selected at random. This is similar to the way humans respond. We will incorporate a number of responses for each keyword, varying in detail. This will mimic the response of a human.

�

Another example shows how Julia can "understand" the conversation.

User whispers: `how are you feeling?' to Julia.

Julia whispers: `Terrible.'

User whispers:`why, julia?' to Julia.

Julia whispers: `I have PMS today.'

In the user�s second response there is no reference to the first response. The reference was implicit. The user simply asks "Why Julia?" This required the understanding of the conversation. If Julia did not have this discourse model the input would simply have been parsed and not carried the conversation on.

Julia will also display inhuman responses if you know which questions to ask. When we asked, "How much memory do you use?" the response, albeit informative, was not very convincing.

Julia does appear a lot more human-like than ELIZA, but maybe not as much so with everyday conversation. We will try to utilise some of the tricks used by Julia to become more human than Chatterbot.