Convuns in the EllaZ System
|[ Privacy, Policies, and Disclaimers ]|
First developed by EllaZ Systems, Conversational Units,
a/k/a Convuns are an important building block for natural language
interaction. Convuns are fuel for conversation, and are stored in a
knowledge base where they are classified as jokes, trivia, information, poems,
lyrics, limericks, fables, maxims, images, and so on.
"Do you know any good jokes about lawyers?" may result in "Would you like to hear something funny regarding lawyers, professional courtesy, and snakes?" A reply of "no, I've heard that one" may result in the offer of a different joke. After hearing the joke, you may be asked for your opinion, and the natural language response is saved in a database by Ella to eventually be used in an effort to be a better conversationalist by adapting to users.
Jokes, trivia, poems, and so on are not simply called up and delivered to the user. As the above example shows, Convuns are used in a manner which is calculated to involve the user in conversation, and also learn about the user's preferences.
Convuns can also be used by Ella to initiate conversation when she has no specific reply to a user. When a user says nothing, and Ella is in mode for automatically generating conversation, Convuns are one of the resources for this.
We improved the performance of Convun matching by changing from an attempt to parse "from, called, by, named, etc." to a more search-engine type approach, where matches in the author, title, text, etc. fields are scored and the best match is selected. Common "grammar words" such as pronouns, conjunctions, selected prepositions, and request words are stripped out first to improve the matching.
For subsequent similar requests, previous matches are held in a buffer, and matches to them are rejected. To overcome this feature, the user can type "again" or "repeat" as part of the input to see a prior request.
To add a sound dimension to Ella, some of the Convuns use a Flash® audio player. Voice talent Robby Garner provided 37 limericks, jokes, and poems for this feature. Example inputs to hear these audio Convuns:
Some of the tracks have lyrics (e.g. Plastic Hand) or descriptions (e.g. Autumnal Suite) that are shown as the music plays. Otherwise, a specified image is shown (from the Image Convun set). Also, when Ella asks if you want to hear the piece she found, a pre-selected image is shown.
Note: The features below have not been implemented. However, we are steadily progressing and no insurmountable technical obstacles have yet been discovered.
First, Ella can learn if a joke is any good. If nobody every likes it, she will progressively decide to use it less and less. This will allow Ella to phase out the bad content, and also phase out information that becomes stale. Sooner or later people will forget about the Jennifer Lopez dress.
Second, whether you like the joke, image, or whatever type of Convun tells Ella something about you. She will use various techniques to make a better offer in the future. Perhaps she will realize that you just don't like poetry and keep such suggestions to a minimum. She could compare your patterns of interests with others that have similar patterns of interest to help him do this.
Third, the patterns of interests she learns about you will be useful useful in her conversations with others. The number of visitors, the lengths of conversations, and the size of the convun database will collectively need to reach a critical mass in order for the system to be effective.
Fourth, Ella could make suggestions to her handlers about weaknesses in her convun database. If dozens of requests for Red Skelton jokes are going unfilled, Ella will be hungry for some rather specific data.
Finally, as the convun databases grow and Ella has larger amounts of data about people's interests, more analytical opportunities will present themselves. Hopefully this will be the basis for ever more sophisticated and powerful AI implementations and Ella's deeper understanding of the human condition. This system provides a potential test bed for advanced AI methods.
|Copyright © 2007 EllaZ Systems||All Rights Reserved|