Intelligent Interactive Technology
The Technology:If you are interested in a business case for Tiddler, please view this summary. 1. OverviewTiddler draws on technology from both the Technology of Electronic Documents (TED) group and the Intelligent Interactive Technology (IIT) group which as expertise in Natural Language Generation. The IIT technology allows the contents of documents to be planned such that it is coherent. The content planning stage uses Norfolk technology from TED, which is able to extract sections of text, and graphics from structured documents, such as certain websites. The content planner merges together this information into a single document. IIT technology is the used again to plan how to present the information on various devices such as small screen, web or paper. 2. User CustomisationThe user model includes the profile data entered by the user and the discourse history, which indicates what Tiddler has already told the user. The profile is used to find information that is relevant to the user's particular trip, filtering irrelevant answers out if it conflicts with the user's information need. The discourse history is used so that Tiddler does not bore the user with information previously mentioned in a past travel guide. For example, when travelling to a specific destination, the travel guide includes general information about the region. Should the user change to another destination in the same region, Tiddler will indicate that the general information for the region is the same to save the user time. Of course the information is still available should the user desire to read it again. 3. Document Re-useThe Norfolk technology is a language for accessing data sources. While it can retrieve information from sources such as SQL databases, it is specially useful in navigating through and manipulating document tree structures, such as well structured web pages. 4. Content and Presentation planningThe content planner uses a library of discourse plans, which indicate how a discourse goal can be achieved. This planner is based on one used in the Isolde project. The discourse plans were designed based on a corpus analysis and represent the prototypical structure of a travel guide. In our application, we studied a variety of travel guides, including travel books, travel leaflets, and on-line guides. The resulting overall structure of the guide is one where, after a general introduction, there is usually, depending on the user model, a need to provide information about accommodation, restaurants, special events when available and activities. All the information provided is tailored, based on the user model. At the end of the planning process, an intermediate tree structure called a discourse tree is produced. It represents the content of the document to be generated and contains explicitly represented coherence relations between various text spans, and the intermediate discourse goals. The particular theory of discourse structure used to represent coherency is Rhetorical Structure Theory (RST). By using such a planner, only relevant content is selected and assembled for the user, and, importantly, a coherent presentation is produced. It is the discourse tree that is re-arranged for presentation on the medium selected by the user. The presentation planner makes inferences based on the coherence relationships in the discourse tree to make the presentation suit the medium. For example, on a small screen device, the relationship might indicate that some text is not of primary importance, and can be placed in a hyperlink to save space. By re-using the discourse tree across all media, the content is organised the same way, and thus navigation is consistent across the different devices. 5. Natural Language GenerationTiddler uses text planning technology, a well established method for generating text. Since most of the text is extracted from other documents, there is little need for complex sentence generation technology. Thus, it uses templates to generate sentences regarding the meta content of the document. However, by employing a Natural Language Generation architecture, domain specific resources are modularised and are easily maintainable without modifying the main planning engine. |
|
Tiddler:
To
toplast updated
July 11, 2005 10:52 AM
Andrew.Lampert@csiro.au