M-PIRO Project Annual Report 2001
The M-PIRO project, which started in February 2000, is concerned with the development of language engineering technology for Personalised Information Objects, concentrating on multilingual information delivery in virtual and other museum settings, where M-PIRO will deliver highly personalised descriptions of museum exhibits. M-PIRO concentrates on the following key technologies:
multilingual generation from a single source
adaptive user modelling for personalised information presentation
improved speech synthesis through the use of domain constraints
generation of spoken and written messages from the same source
authoring of the single source through the use of symbolic authoring techniques
System specification and design
As last year's report noted, we have based the M-PIRO text generation system on the Ilex system, which described objects from a virtual jewellery collection. We are now working to enhance that design by providing more flexible interfaces with systems for user modelling, collections databases, and delivery platforms. Following on from the SOLE project, we are also producing a combination of text generation with speech synthesis based on the Festival system.
Summary of 2001 Activities
The project achieved a number of significant milestones in 2001, including the first versions of the M-PIRO authoring tool. The tool allows easy manipulation of the M-PIRO database contents and provides other facilities, for example for updating the lexicon and entering canned stories decribing the exhibits. In other areas, we made further significant advances in speech synthesis in Italian and Greek and continued our work on the production of Exprimo, the text generation component.
The first design of the demonstrator reused the Ilex system to describe items from a virtual collection of fifteen ancient Greek artefacts. The reimplementation in Java, called Exprimo, was done for various reasons:
We created the first Italian female voice for the Festival speech synthesiser and experimented with new voices for the MBROLA synthesiser. The lexicon of Italian was extended, and we created a new M-PIRO lexicon for Greek and English.
The design and the implementation of the DEMOSTHeNES Speech Composer for Greek was completed. DEMOSTHeNES is a TTS framework which accommodates previous work on text analysis and prosody generation. We moved to stochastic models for prosody generation and developed a text corpus for museum domains which was extracted from M-PIRO's exhibition descriptions. This forms the basis for the creation of a phonetically balanced speech corpus which supports the study of Unit Selection. Our evaluation of Greek speech quality showed that trained prosody models significantly enhance the speech quality of diphone-based TTS compared to rule-based models.
In addition to the English resources, we now have large-scale generation resources for Greek and Italian including the enhanced lexicons mentioned above as well as extensive grammatical information for both languages.
We completed the development of the first version of the user modeling subsystem, which supports long-term personal models and stereotypes. This consists of a personalization server and code which translates between M-PIRO's user modeling API and the HTTP-based protocol that the server uses as its native language. An instance of the personalisation server is running on an NCSR machine to which partners have access. Forms showing the types of information in the domain database were constructed and used to elicit responses from experts at FHW. The data involved stereotypical values regarding the interest, importance and assimilation rate of various kinds of information for three user types (child aged 12-13, non-expert adult aged 30-35, and archaeologist). The resulting stereotypes were loaded into the personalization server. FHW also created a mock-up of the part of the interface where the users will enter the system and the profile is determined.
The development of the first versions of the authoring tool were completed. The tool provides mechanisms for manipulating the database schema and contents, as well as facilities for updating the domain-dependent lexicon, specifing micro-planning expressions, and entering canned stories. Some modifications in the original design of the tool were made, mostly to address suggestions made by the reviewers at the first project review. Deliverable D3.2 describes the functionality of the first version of the tool and plans for future work on authoring, including evaluation plans and exploring ways to import information from existing museum databases. The authoring tool is being tested by experts at FHW in order to improve its functionality, providing on-going evaluation. An extended collection has also been constituted.
Future Work
In the final year of the project we will concentrate on the architecture for a collection based on virtual-reality delivery and on evaluation of the prototypes by curators from FHW, museum clients, and the advisory board, using the guidelines in the evaluation and assessment plan. Preparatory work for the VR application (technical specifications, requirements for adaptation and integration, as well as the VR collection and related information) was started by FHW, NCSR and UoA in October 2001. The final version of the authoring tool will also be completed and evaluated.
Further Information
For more information, please see the project web page: http://www.ltg.ed.ac.uk/mpiro.