EMU A European Multilingual Text Prediction Software

1
EMU A European Multilingual Text Prediction
Software
C. Beck, G. Seisenbacher, G. Edelmayer, W.L.
Zagler
Presenter
Research Group on Rehabilitation
Technology Institute integrated
study Technische Universität Wien Favoritenstraße
11/029, A-1040 Wien, Austria
2
Introduction

• Typing important for
• Written communication (letters, reports, etc.)
• Spoken communication (Text-to-speech synthesizer)
• Using Information Technology (E-mail, internet,
  etc.)
• Pre-EMU situation
• Many products for the English speaking market
• Products for other languages
• often derivate from English products
• not designed for highly inflecting languages

3
The Project FASTY

• Baseline of further Development
• Type EC co-funded RD project IST-2000-25420
• Duration January 2001 to March 2004
• Goal System for increasing the text generation
  rate of disabled persons
• Special pressure sensitive input devices
• Generic prediction engine
• Integrated Userinterface for all modules

4
The UI Features of the Final FASTY Prototype

• Different ways of positioning the text prediction
  window
• Fixed global shortcuts for the prediction
  selection
• Integrated Tool for Setting the Options
• Fixed sound and speech output systems with
  limited options

5
The Prediction Features of the Final FASTY
Prototype

• Uni- and Bigram-based statistical prediction
• Trigram-based Part-of-Speech statistical
  prediction
• Grammar-based prediction ranking
• Collocation based prediction
• Compound prediction

6
Some problems of the Final FASTY Prototype

• Not stable enough for a market product
• Not enough options to adapt the system to users
• General problems with the integration into the
  Operating System
• Heavy System Load
• Problems with the Dictionary Content
• Fixed Speech Output and Sound System
• Problems with the User Interface Design
• Many user ideas could not be implemented

7
EMU the Next Generation

• Using
• the knowledge of the FASTY Project
• the basic structure of the FASTY Project
• the prediction techniques of the FASTY Project
• the user comments from the FASTY Project tests
• New
• implementation from scratch
• detailed structure to cope with the encountered
  problems
• dictionaries
• features coming from additional discussions with
  users

8
General Structure of EMU

• Runtime
• Language Component
• User Interface
• Adjustment

9
The EMU User Interface

• Active part
• receives keystrokes
• initiates prediction procedure
• sends predicted word (parts)
• Runs on Windows 98, 2k, XP
• Available in English, German, Italian Swedish
  (working on French)

10
Detailed Structure of the EMU User Interface

• User Interface
• Core
• System Hook
• Drivers
• Text Sniffer
• Language Component
• Common Data Interface
• Adjustment Tool

11
The Language Component of EMU

• Platform independent
• Language independent
• Contains
• Contextbox
• Prediction Engine
• Predicticion Parameter
• Common Data Interface
• Language Data

12
Changes within the Language Component from the
FASTY Final Prototype

• New dictionaries available in 8 languages (Dutch,
  English, French, German, Italian, Spanish,
  Swedish, Turkish)
• No Grammar-based ranking (only 0,2 KSR gain)
• No Collocation based prediction (only 0,2 KSR
  gain)
• Introduced common data interface with new
  optimized file format for language data
• New abbreviation handling
• Drastically lowered the system resource needs

13
Prediction Performance of the FASTY Final
Prototype
14
Prediction Performance of EMU

• Little laboratory tests
• Much better user feedback
• Less wrong words (coming e.g. from news texts)
• Predictions are optimized for the every day
  language
• KSR some 5 points better than the FASTY Final
  prototype
• More text type independent

15
The Adjustment Tool of EMU

• Runs independently from the Runtime Program
• Uses the same Common Interfaces as the Runtime
  Program
• Very intuitive options with immediate feedback
• Supports automatic user related dictionary
  creation from existing text
• Supports import of existing MS-Word abbreviation
  dictionaries

16
Feedback example of the Adjustment Tool

• Example line shows combined result
• Color labels show single colors

17
Help example of the Adjustment Tool

• Structured Feature list
• Immediate help text
• Less important / expert settings only shown when
  needed

18
A European Multilingual Text Prediction Software
http//www.is.tuwien.ac.at/emu/index_en.html