I believe that the ability (or lack thereof) to perform semantic inferencing using encyclopedic, or ``world'' knowledge, will ultimately determine the upper bound on quality of MT. Humans use world knowledge to a great extent, greater than most people give them credit. Take the simple narrative:
The analysis (or generation) of the pronoun references in 1) are completely dependent on world knowledge. In general, a computer would find it very difficult to process something like 1), whereas a human would find it trivial.
[Goodman 75] was one of the earliest (and still one of the only) attempts to take semantic inferencing seriously. Embedded in a conceptual dependency (CD) framework, Goodman used discrimination nets to generate verbs based on the semantics of the clausal roles. For instance, ``eat'', ``drink'' or ``breathe'' would be chosen depending on whether the object was solid, liquid or gas. Naturally, this kind of inferencing was essential given the small number of semantic primitives in CD; however, its application to semantic theories with fine-grained ontologies still holds.
The problem of uniquely referring to an object that has many properties in the midst of a communication situation filled with many objects is addressed in EPICURE [Dale 91]. There, a ``discriminatory algorithm'' for picking out the smallest set of features needed to uniquely identify an object is given. This type of identification is needed when referring to objects for the first time in a discourse. Again, though, there is the problem of deciding which of a potentially large set of ``needed'' semantic properties can easily be inferred by the hearer. For instance, the agent of a ``read'' event hardly needs to be referred to as ``the human with the orange coat''.
Most of the literature on using focus constraints to generate appropriate anaphora concede that a mechanism to identify ``global'' focus is needed (see [Brennan et al 87]). The problem is that identifying ``global'' focus is dependent, to a large extent, on understanding the discourse. This, of course, entails the use of world knowledge. I would argue that the problem goes much further than that. As shown in 1), even local focus is extremely dependent on semantic constraints. Although the focus constraints tracked by methods such as ``centering'' are valuable, they eventually pale in comparison to the inferencing abilities humans use constantly. The effect of inferencing on focus is a largely unexplored area.