Comenzar aquí la presentación

 
    OZ Characters
      
    Yuri A. Tijerino Ph.D.

    OZ Interactive Inc.
    525 Brannan Street, Suite 400
    San Francisco CA 94107
    Tel: 415-536-0500 Fax: 415-536-0536
    URL: http://www.oz.com/

    What are OZ Characters? 

    The OZ characters are synthetic agents that are used to enrich virtual worlds or web pages by providing information or entertaining the user. The user can talk to the agent through a chat interface, but in addition the agent can react to events in the environment. The agents are controlled by the OZ Intelligence Engine, which uses a combination of templates 

    and natural language processing to parse the user input and produce appropriate responses in the form of text and actions. This demonstration will be given in two stages. 

    OZ Intelligence Engine 

    The OZ Intelligence engine is the AI engine used to control the agent. The engine is template-based and it uses a natural language parser to process the user input. The template matcher selects the appropriate template matches from the knowledge base and then presents the best answer by selecting from the candidates. The answers can be given in the form of text, speech or by some type of action such as opening a web site or playing a video. 

    The structure of the Intelligence Engine
     

    The main parts of the Intelligence engine are: 

    • A lexicon, organised hierarchically by the meaning of the words. The basic lexicon (used by all our agents) now contains more than 6000 words. The words are marked with the word group and also have a mood parameter (i.e. what feelings does this word evoke in the agent). Words belonging to the same sub-tree in the lexicon are identified with a placeholder.
    • A natural language parser, which parses the user input, analyses the sentence structure and identifies the words.
    • Templates can used to match the user input. Each template contains one or more answers that become candidates in the output selection if the template matches the user input. The templates can contain words or placeholders from the lexicon. Template matches can be at different levels of abstraction, i.e. the match might be very tight (several matches in the sentence), very loose (a word in the input is contained in a placeholder) or somewhere in between.
    • Answers can belong to one or more templates. If a template is matched, one of its answers is selected as a reply. An answer contains:
      • the verbal reply (to send to the user) and/or
      • the action to perform (avatar action, event triggering etc)
      • criteria for when to use the answer
      

    The templates are typically used to match sentences typed in by the user. They can also be used to react to some event in the environment, such as a proximity sensor, the presence of an object or what song is being played. 

    The template and answer selection is controlled by the flow of the conversation, template and answer priorities, how good the match is, the mood of the agent and the context of the conversation. 

    Authoring agents with the OZ Character Creator 

    The Character Creator is an authoring tool developed specifically to create OZ characters. It allows an author to design the personality of the character based on its intended purpose in virtual life. Once the character has been drafted, the author creates the templates and responses needed to model it. A template language is used to create the templates needed to match the user input. The author maintains a hierarchically organised list of words that the agent should recognise. The author can create several replies for a template and specify the criteria for when the answer is appropriate 

    Designing a Character 

    The knowledge of a character can be divided into 3 categories: 

    • General knowledge. This is something everybody knows and covers information such as a dog is an animal or the sky is blue.
    • Personal knowledge. This is information everybody knows about themselves and their surroundings, but the information varies from person to person. It includes knowledge about your name, the name of your parents and so on.
    • Area of expertise. The character has in-depth knowledge on this area.
    The personality of the character dictates how it handles these three categories. Specialised knowledge is usually the easiest to create, as it tends to be based on facts; this is the information we want the character to tell the user. 
    The author must create a convincing background story for the character in order to cover the personal knowledge information. The personality of the character dictates how much it is willing to tell about its background or even how much people will ask about it. 

    General knowledge is the hardest category to handle. The author must decide what should or should not be included and how the character should react when it does not know the answer. 

    Once the personality of the character has been designed, the author can select the strategies it should use in the conversation. The main concerns should be 

    • How to deal with ignorance
      • Does the character admit ignorance?
      • Does it agree/disagree with the user when it does not know what the user is talking about?
    • Who should control the conversation?
      • Does the character change the subject?
      • Does it continue with its thread, although the user tries to change the subject?
      • Does it make controversial statements to get the user into a conversation?
    • Can the character deal with direct questions?
    • Does the character use humor?
    • What is the character's attitude? Is it:
      • Helpful or arrogant
      • Jolly or sad
      • Polite or rude
    • How should the character introduce its area of expertise?
      • Dropping hints
      • Direct questions
    Once the strategies have been selected, the responses of the character should match the model to keep the character consistent and believable. 

    More information can be accessed at http://www.oz.com. 

    Screen Shots 

    The following series of screen shots demonstrate some of the functionality and capabilites of both OZ agents and the Character Creator, the agent authoring tool. 

    Figure 1. Dione is an agent OZ created for Ericsson World, a 3D environment that showcases some of Ericsson’s products. The agent can interact with Ericsson’s clients, answer their questions about specific products, take them to specific places in the environment and show video clips or 3D simulations of the products. The angel in this screen shot represents the agent, Dione, and the lady represents a user called Honey fly. 
     

    Figure 2. Her e the Dione has taken Honey fly to an specific area of the 3D environment of Ericsson World. 
     

    Figure 3. Theresa is an agent with an atitude that is particularly knowledgable of Greek literature. She is fun agent to have conversations with. She lives with other buch of agents in the OZONE, a 3D environment that can be downloaded and played with at our web site (http://www.oz.com/). 

      

     
     

    Figure 4. Zon is also a very mature agent with a philosophy background. His is also play of fun to talk with. He lives with Theresa in the OZONE. 
     
      

    Figure 5. The Character Creator allows the interactive and intuitive creation of new agents. It ships with a basic agent that has over 6,000 word lexicon and several hundred templates. The user can create new lexicons and new templates to represent general or specific knowledge. It is easy to extend and integrate several knowledge bases with this tool. 
     
     

     
    Figure 6.
    principal    indice
    Responses can be associated with mood parameters. This helps the agent choose specific responses from a template for a particular template matching to user input. This is very useful because templates can be associated with more than one response to avoid repetition of the same answers.
     

    For more information visit our web site at http://www.oz.com/. 
     

 
 
principal    indice