Current paradigms in computational linguistics:

• Rule-based systems: parsers, generators, translators that include grammars
• Template-based systems: slot-filler approach with pattern-matching for input and output
• Stochastic systems: machine-learning approach which optimizes input/output based on frequency in a corpus

Limitations to the approaches:

• Rule-based systems: limited fluency (idiomaticity); "whole grammar" problem (long development time, offset by reusability)
• Template-based systems: limited generality; "long list" problem (very large number of potentially relevant templates, limited reusability)
• Stochastic approach: systems hit performance threshold; corpus-dependence

The "ideal system" could use elements of all three approaches. The problem is knowing where which is most applicable.

Generality problem: all systems do better in limited application domains than broad ones, though some are more vulnerable than others.

Directionality problem: different issues present themselves first in understanding and generation. The ideal system is bidirectional but this is difficult to achieve.

World knowledge problem: in some applications a completely general (human-like) world knowledge may be required to go "all the way". Achievable?