The example I gave is pretty extreme, but it's a good example of why natural language parsing is a difficult problem. The ability to strain interpretations from sentences isn't just limited to nonsensical sentences, it works for perfectly good sentences too. You can seriously narrow down the number of possible parses for a sentence by actually having and enforcing a grammar, but even then you often get thousands of legal parses for a typical long sentence, and there are few hard-and-fast rules for selecting the correct one - basically, you need world knowledge to be sure, and that problem is AI-complete. The problem is sufficiently severe that in some cases humans can't do it if they lack sufficient knowledge of the subject matter - an issue we came up against when we got a trained linguist (who wasn't a chemist) to make parse trees for sentences from chemistry papers.