Grammar Engineering Frequently Asked Questions
I've tried to use interactive unification to find out why an edge can't be built, and it can be built interactively, but it's still not in the chart. What's going on?
With interactive unification, you can mimic interactively what the parser does when it tries to parse a sentence. Therefore, edges that the parser can't find one way it shouldn't be able to build the other. However, sometimes you use interactive unification to try to find out why an edge can't be built, only to find that seemingly it can. There are probably many ways this can come about. Here are seven I've found so far:
- The edge you thought was missing actually wasn't. Double check the chart.
- You took the two daughters and unify each separately with the grammar rule. If you do this, they will almost always both unify. The proper way to apply interactive unification is to unify one daughter with the grammar rule, and then unify the other not with the rule but with the unification result.
- Edges you are trying to use as daughters appear to be contiguous in the chart, but actually are not. This can happen because if the LKB can't construct an edge for a lexical entry, it won't put anything for that entry in the chart. It will, however, record the fact that there is something there, and not let constituents bridge the material it can't account for. On the possible causes of this situation, see this FAQ .
- The problem is actually with the direction of headedness (i.e., order of daughters) in the rule. For example, say your language has VO word order, but you've made your head-complement rule inherit from
head-final. If you try to unify the verb with the
HEAD-DTR of the head-complement rule and the NP with the
NON-HEAD-DTR, it will work just fine. However, when the LKB tries to parse the sentence, it can't apply the rule because it's expecting the NP (
NON-HEAD-DTR) as the first (i.e., left-hand) daughter, per the constraints on
head-final. (The LKB uses the feature ARGS to determine the order of the daughters. The types
head-initial link the values of
NON-HEAD-DTR to their respective positions on
- A rule you are expecting to appear in the parse hasn't been instantiated in
rules.tdl. (In this case, you wouldn't expect to see the rule in the View > Grammar Rule or View > Lexical Rule menus, but you could still have accessed its type definition through View > Expanded Type, for example.)
- If the rule that is not firing is a lexical rule, it's possible that you've said it's an inflecting lexical rule (i.e., one that adds an affix) but defined it (in
lrules.tdl) so that it doesn't in fact add one.
- The word in question is misspelled or the lexical rules are buggy such that the morphophonological rules produce only analyses of the lexical item which are spurious according to the morphosyntactic rules. For example, say you have a stem 'CV' and a lexical rule which adds a -V suffix, but isn't actually morphosyntactically compatible with the stem 'CV'. In parsing a string which includes the 'word' 'CVV', the LKB will 'strip' the -V suffix (per the morphophonological rule) and find the 'CV' stem. The LKB will accordingly create an edge in the chart for this lexical item, labeled with the corresponding form from the input string ('CVV'). However, the LKB has remembered that this edge needs to be put back through the -V rule in order to be legitimate. If the CV stem is not compatible morphosyntactically with the
DTR of the -V rule, it won't spawn any further edges in the chart, even if some other rules (e.g., a phrase structure rule) could happily take it as a daughter.
- The rule applies to the lexical edge as it appears in the chart, but that lexical edge represents some spelling changes, and the lexical rules associated with those spelling changes failed to apply.
- The rule is one of a chain of lexical rules, and some following (spelling changing) lexical rule fails to apply. Since the LKB won't put edges derived by lexical rule in the chart unless the whole word can be rebuilt to the given surface form, edges for successful but intermediate lexical rule applications might not appear. In this case, the problem is actually further along the chain of lex rules, so the next step is to do more interactive unification, using the result from the interactive unification as input to the next expected lex rule.
Back to FAQs page
Back to main course page
- 02 Nov 2004
Topic revision: r2 - 2009-05-07 - 00:04:13 - ebender