Every time I fire a linguist, the performance of the recognizer goes up. -- Fred Jelinek
Traditional NLP involves writing grammars, parsers, and semantic programs to understand language.
An alternative is to use statistical approaches, based on statistics from massive amounts of text, without attempting to understand the natural language. This approach has been surprisingly successful in several area:
The basic approach is Bayesian statistics: given the preceding language, what is the most probable interpretation of the current input?