Raising the Interlingual Ceiling with Multilingual Text Generation

In a typical interlingual machine translation (MT) system, the tasks of text planning and content selection are not explicitly performed. Rather, they are assumed to be implicit in the interlingual representation derived in the source language analysis phase. This simplifies the task of target language generation, but greatly restricts the flexibility of the resulting text. Recent MT systems have pushed the level of the interlingua to higher, more semantic/pragmatic levels of representation, and have correspondingly included some text planning capabilities in their generation phases, but the process remains closely bound by the structure and content of the source text. This places a “ceiling” on the extent of the capability of interlingual systems. In this paper we argue that in the general case, the structure and content of “parallel” texts need not be the same. We then propose that multilingual text generation (MLG) has the potential to address this phenomenon by allowing a more flexible variation of text structure and of content selection in a manner unrestricted by a “source” text. This work is partially supported by the Engineering and Physical Sciences Research Council (EPSRC) Grant J19221, the Commission of the European Union Grant LRE-62009, and by BC/DAAD ARC Project 293.