Co-chairs: Michael Carl, IAI, Saarbrücken; Andy Way, Computer Applications, Dublin City University

In recent years, corpora of multilingual translated texts have become widely available for a number of languages. Notwithstanding the seminal paper by Nagao (84), it is primarily since the early 90's that such bilingual texts have been exploited in the area of Machine Translation (MT).

The two main paradigmatic approaches which have been proposed are Statistics-based Machine Translation (SBMT) and Example-Based Machine Translation (EBMT). A related variant of EBMT that we ignore here, despite being widely used in the localisation area, is that of Translation Memories (TM). No new translations are created afresh from previously existing examples in the system database: rather, the closest translation matches are proposed to the user for post-editing into the correct translation.

While translation memory systems are used in restricted domains, SBMT systems require training on huge, good quality bilingual corpora. As a consequence TMs can hardly be applied as a general purpose solution to MT and SBMT as yet cannot produce complex translations to the desired quality, even if such translations are given to the system in the training phase. EBMT seeks to exploit and integrate a number of knowledge resources, such as linguistics and statistics, and symbolic and numerical techniques, for integration into one framework. In this way, rule-based morphological, syntactic and/or semantic information is combined with knowledge extracted from bilingual texts which is then re-used in the translation process.

However, it is unclear how one might combine the different knowledge resources and techniques in an optimal way. In EBMT, therefore, the question is asked: what can be learned from a bilingual corpus and what needs to be manually provided? Furthermore, we remain uncertain as to how far the EBMT methodology can be pushed with respect to translation quality and/or translation purpose. Finally, one wonders what the implications and consequences are for size and quality of the reference translations, (computational) complexity of the system, sizeability and transportability, if such an approach is taken.

Given this background, we propose to organize a workshop in order to shed some light on these open questions, among others. We are seeking contributions which go beyond the purely statistical and/or rule-based approaches to MT. We welcome visionary and technical descriptions, reports of empirical research as well as feasibility studies and system demonstrations. We would welcome contributions on any of the following topics and sub-headings:

• (semi-)automatic preparation of existing bi/multilingual corpora for EBMT
  • extraction of bi/multilingual texts from the web
  • preparation of treebanks for EBMT
  • bi/multilingual alignment/bracketing/parsing
  • inference of bi/multilingual grammar and transfer rules
• description of `pure' EBMT systems
  • knowledge resources used
  • representation of numeric and symbolic knowledge
• descriptions of `hybrid' systems integrating EBMT with rule-based or other methodologies
• evaluation of EBMT results and/or comparison with other MT systems
• considerations on domain-(in)dependence of EBMT systems
• computational and/or system complexity of EBMT systems

Programme

The workshop organizers have just produced the final programme:

10:00 - 10:15 Opening Remarks

10:15 - 10:45 Davide Turcato and Fred Popowich, What is Example-Based Machine Translation?

10:45 - 11:45 Harold Somers, EBMT seen as Case-based Reasoning

11:45 - 12:15 Coffee Break

12:15 - 12:45 Andy Way, Translating with Examples

12:45 - 13:15 Reinhard Schäler, Beyond Translation Memories

13:15 -- 14:45 Lunch

14:45 - 15:45 Kevin McTait, Linguistic Knowledge and Complexity in an EBMT System Based on Translation Patterns

15:45 - 16:15 Ralf D. Brown, Transfer-Rule Induction for Example-Based Translation

16:15 - 16:45 Coffee Break

16:45 - 17:15 Michael Carl, Inducing Translation Grammars from Bracketed Alignments

17:15 - 17:45 Arul Menezes and Stephen D. Richardson, A best-first alignment algorithm for automatic extraction of transfer mappings from bilingual corpora

17:45 - 18:00 Closing Remarks

Papers

The papers which were included in the conference proceedings are now online.

Publication

There will be a common publication format for all workshops in line with the main conference proceedings. These will be the usual book format for conference proceedings as well as CD-ROM. However, it is anticipated that the proceedings will be published more widely, and relevant publishers have been approached in this regard.

Important Dates

Details of registration procedures, including registration fees, have just been announced. The attendance fee for our workshop is Euro 50.

Organizing Committee

• Sivaji Bandyopadhyay, India
• Ralf Brown, USA
• Michael Carl, Germany
• Ilyas Cicekli, Turkey
• Brona Collins, Belgium
• Oliver Streiter, Taiwan
• Stephan Vogel, Germany
• Andy Way, Ireland