Some critics have questioned the fundamental presumption of translation reusability as it is currently implemented, arguing that translation memories are really most effective only with documents that change very little over time and have significant sentence repetition. Webb (1999) studied eleven kinds of translation and determined that only four of the eleven types, legal, scientific, technical and commercial, would benefit greatly from translation memory. As soon as variability between documents and document versions is introduced as a factor, the utility of translation memory declines because sentence repetition declines.

Another severe limitation of translation memory is a dependence on the sentence as the primary linguistic translation unit. This dependence has several implications. Macklovitch and Russell (2002) have been critical of the inability of current systems to exploit reusable text at the subsentential level, which we might call microreuse. The authors claim that the great bulk of reusable material consists of elements at the subsentence level: phrases, collocations and multiword terms. Computer-assisted translation systems with finer linguistic granularity could exploit linguistic resources contained within translated sentences. Conversely, Macklovitch and Russell (2000: 137) also criticize the artificial segmentation of the text into discrete, semiautonomous units, and the subsequent loss of access to suprasentential relations, arguing that the "very notion of a document is lost. Not only are the segmented units in a new text extracted from their context and submitted to the database in isolation, but the contents of the database are also stored as isolated sentences, with no indication of their place in the original document." This is an extremely important point. The adaptations that translators and localizers make to documents during translation are not confined to the sentence, but often cross sentence and paragraph boundaries. Linguistic elements considered during translation decision-making - or that should be considered - are also almost certainly not confined to the immediate linguistic microcontext of the sentence.

Clearly, translation memory systems do not preclude access to surrounding sentences or the ability to read a paragraph or a document as part of decision-making, but there is a clear predisposition, even channeling, of translator behavior to sentence-level processing (Dragsted 2002; Webb 2000). An unintended side effect of the focus on the sentence in translation memory systems might be an undesirable feedback effect on the translation process, leading translators to translate, perhaps unknowingly, more in the microcontext than they normally would.

Translation reuse and language reuse are not synonymous. In translation reuse only certain portions of translated texts are reused. Language reuse, at least from the perspective of computational linguistics, could also include discovering and reusing both subsentential and suprasentential textual elements. A broader application of language reuse would also remove the restriction that all reusable (or usable) elements be derived from the relatively limited corpus of translated texts. It is clear that the full potential of language reuse has not been exploited by the language industry.

责任编辑：admin