SALS-SIG Seminar

Contrastive-linguistic resources for multilingual generation: evidence from (machine) translation

Elke Teich

When: Tuesday 30th September, 1997

Time: 11:30am

Where: Room E6A357, Macquarie University

Abstract:

This talk deals with the question of how multilingual resources for use in the automatic generation of multiple languages can be motivated so that a better output quality can be produced than currently is the case with 'standard' multilingual technology, such as machine translation (MT). While there are clearly many reasons why machine translation has not made any significant progress in the last two decades or so, in this talk we focus on one source of problems, which is the types of information typically included in multilingual systems -- be it MT systems or multilingual (ML) generation systems, and the types of contrastive-linguistic relations that these allow to state. Using a sample machine translation, we compare the types of contrastive-linguistic relations that MT systems typically know about (most often encoded as transfer rules) with what translation theory maintains are the relevant contrastive-linguistic relation types (called translation procedures). We will argue

• that current ML approaches only partially implement the translation procedures described in translation theory,
• that their application (e.g., in the form of transfer rules in MT) remains undermotivated , and
• that therefore the quality of output is compromised.

In order to alleviate these deficiencies, other kinds of linguistically relevant information have to be included in multilingual systems. In the talk , we will try to identify these additional types of linguistically relevant information and sketch a multilinguality model that takes as a basis the concept of resource sharing as implemented for tactical generation in the Komet-Penman Multilingual (KPML) system. The linguistic data we present are from English and German.

Enquiries: sals@mri.mq.edu.au