Hurdles to wide-spread pandorabot deployment
How do you create a good pandorabot? People willingly spend time with good pandorabots. The quality of your pandorabot depends almost entirely on how long people spend time with it. And this depends on its ability to interact with others while offering mesmerizing and spellbinding content.
Let’s understand how content creation works by considering how young children learn language. The analogy we are developing only goes so far, yet helps to develop an understanding of the limits of the content creations process.
We imagine a young child engaged in learning language. Unlike in the real world, we imagine a parent who has the capacity and strength to diligently observe each of the child’s interactions with the external world (including the interactions with the parent). The parent tirelessly examines interactions and teaches the child appropriate responses.
The content creation process involves a pandorabot author (also known as a botmaster) examining a conversational log of interactions with clients looking for inappropriate responses to client inquiries, and authoring new responses.
The original Alice pandorabot emerged out of just such a process over a 4 year period. Dr. Wallace examined conversational logs, making corrections to inappropriate responses. The number of interactions Alice encountered with clients was very important in this process.
Please take a moment to notice how important these interactions really were to the process. Developing Alice was only possible through a process involving many willing clients. Real interactions with actual clients are required to build up the content.
Falling back on our analogy for a moment: A child only exposed to parental interactions typically lacks the experience necessary to engage in conversations with others. Conversations with others is vital to the process of authoring content creation.
Most people start with one of the copies of the Alice pandorabot content, and go on to author variants of the Alice responses.
We’ve got experimental evidence suggesting that a high-quality pandorabot requires authoring about 10,000 responses. You need about one week or 170 hours assuming on average you can author one response a minute and don’t need sleep. You can complete the task in three weeks working 8 hours a day.
Crafting response depends on obtaining a large corpus of high-quality input sentences, and Pandorabots Data Services provides high-quality input sentences for a variety of specialized domains. Contact email@example.com for more information.
We’ve identified these hurdles preventing wide-spread deployment of pandorabots.
Rapid authoring of high-quality pandorabot content
High-quality authors are available. How should they be compensated? Tools supporting rapid content creation and modification are already available at Pandorabots.
High-quality text to speech will improve emotional communication.
High-quality speech recognition removes the requirement of interacting with pandorabots through text messages (and the use of a keyboard).
High-quality game characters are already available on fast processors. We can expect characters to become available on a variety of other personal devices in the future.
Studies demonstrate that much human interaction occurs outside of the confines of simple conversational content. People use gestures, emotional communication, a variety of non-conscious body-based responses, voice tone, voice inflection, eye-contact, etc. We are working on adding these capabilities to pandorabots now.
Natural Language Understanding
Our open-source Artificial Intelligence Markup Language - AIML - provides one of the important pieces for robust Natural Language Understanding. Authors create language patterns that uniquely identify classes of equivalent statements - as in “The sky is blue”, “There are no clouds in the sky”, “The sun is shining brightly”, etc. Authors create responses to these classes of equivalent statements. Each author brings a their own unique perspective to the world and thus provide responses that best fit a unique category of people.
Semantic Analysis - Understanding World Categories and Facts
We make extensive use of semantic information, adhering to the World Wide Web Consortium specifications, to build up world knowledge bases.