Penalty rules.

The simplest disambiguation method consists of hand-written `penalty' rules which implement a variety of preferences. Each such penalty rule describes a partial parse tree. For a given parse tree, the system computes how often a sub-tree matches with a penalty rule, giving rise to the total penalty of that parse. The following lists characterizes some of the penalty rules:

• complementation is preferred over modification
• subject topicalization is preferred over object topicalization
• long distance dependencies are dis-preferred
• certain rules are dis-preferred (e.g. rules which coordinate categories without an explicit coordinator)
• certain lexical entries are dis-preferred (e.g. the preposition readings for the words aan, bij, in, naar, op, uit, voor, tussen are preferred over the adjectival, noun and/or verb readings).
• certain guesses for unknown words are preferred over others

As can be concluded from the preliminary results presented in table 3, it appears to be the case that about 60% of the disambiguation problem can be solved using this very simple technique.