Interactive Unification in the LKB

When do you want to use interactive unification?

Interactive unification is useful when you find that a sentence which should parse isn't parsing. Go look at the parse chart, and look for the lowest point in the tree where an edge you expect to see isn't there. For example, you have a determiner and a noun, but no edge spanning both of them, and you expect that the head-specifier rule should have built one.

The parser will have of course tried to build a phrase using those daughters and that rule (because it tries everything, as it's doing an exhaustive search). Because it didn't succeed, you don't find the edge in the chart. You can find out why exactly it didn't succeed using interactive unification. This will tell you where you might want to make changes in your grammar so that you can build that edge. Or you might decide that that edge shouldn't be built after all, and go look for another edge that should be there.

How do you do interactive unification?

• In the chart, click on each of the daughter edges, and select "feature structure" to get the feature structures for the edges. (Interactive unification can also be used to debug a unary rule. Here I'm going to talk about binary rules.)
• In the LKB top menu, choose View > Grammar Rule and pick the rule you think should have applied. This will give you a feature structure for the rule.
• In the feature structure for the rule, scroll down until you find the HEAD-DTR feature (it may be easier to shrink the values for SYNSEM, ARGS, and C-CONT on the way).
• Click on the type that is the value of HEAD-DTR (probably sign, but possibly something more specific), and choose "select" out of the pop-up menu.
• In the feature structure for the edge that should be the head daughter, click on the highest type (e.g., common-noun-lex or head-comp-phrase) and in the pop-up menu choose "unify".
• If the unification failed already, there will be a message in the LKB window telling you the first point at which the two feature structures didn't unify (there may be more).
• The rules tend to be fairly non-specific, so this first unification usually works. If it does, you'll get a new window called "unification result" which is exactly the information from the rule with the information of the head daughter unified in.
• In the unification result feature structure, scroll down (or shrink down) to the NON-HEAD-DTR feature. Click on the value of that feature and choose "select".
• In the feature structure for the edge that should be the non-head daughter, click on the highest type (e.g., det-lex or head-spec-phrase) and choose "unify".
• Go look at the LKB top window for a message indicating the first point at which the two feature structures didn't unify (there may be more). This information should be useful for debugging.

Some bugs that are hard to detect with interactive unification

The method described above makes use of the HEAD-DTR and NON-HEAD-DTR features. If you happen to have selected the wrong ordering supertype for your rule (head-final or head-initial), you might not be doing what you think you're doing. So, if the result of trying interactive unification is very surprising, consider doing it via the ARGS list instead, unifying the first daughter with the first thing on the ARGS list and the second with the second. (NB: ARGS is not ARG-S.)

Problems with the spelling change aspects of lexical rules can't be debugged in this way.

If you haven't in fact defined the rule (including an instance in rules.tdl) there won't be anything to try unifying with. But, you should discover the lack of the rule when you try to "view" it from the LKB menu.