Grammar Engineering Frequently Asked Questions

When I try to generate, the LKB says "probable runaway rule". How do I debug this?

As with the parallel error in parsing, this happens when the generator chart gets too big. Because the generator is working from a bag of lexical items corresponding to the semantic relations it was given, but no particular order of those lexical items, it has a larger search space than the parser does. Thus, it is possible that you could hit this error without actually having a runaway rule. However, in the most common case, it is recursion that is causing the problem.

One thing that can cause recursion in generation without also causing it in parsing is semantically empty lexical items (e.g., complementizers). Unless you write filter rules for such lexical items, the generator has to keep positing them, as many as might be needed to complete a parse that covers all the semantics. Since the generator is also doing an exhaustive search, this means that the empty lexical items can keep getting thrown into the chart. If, for example, a phrase built out of a complementizer and it complement is an acceptable complement for the complementizer, this gets you a "runaway rule" situation.

To debug this, try generating, get the error, and then chose "Generate > Show gen chart" from the LKB top menu. (Be aware that it might take a while to display the chart, since it will be big.) As with runaway rules in parsing, the chart should help you see what's spinning.

If (due to lack of memory), the LKB won't display the big generator chart in the GUI, you can see it in emacs. Go to the common-lisp buffer and type:

(pprint *gen-chart*)

Because the generator's search space is larger than the parser's, in some cases it's possible that you could run out of chart edges without finding all possible realizations, without there actually being any illicit recursion involved. You can test this by increasing the number of edges. In the emacs buffer common-lisp, do:

(setf *maximum-number-of-edges* 8000)

(This doubles the default setting, and might mean that you end up using virtual memory in trying to generate.) Try generating again. If you're successful in generating with the increased edge limit, then you don't have illicit recursion. There might still be something that's underconstrained in your grammar, which is unnecessarily enlarging the generator's search space. It's possible that adding filter rules for semantically empty lexical items might help.

Scenarios that can lead to runaway rules in generation

  1. A lexical rule (inflecting or not) is written so that the mother is compatible with the daughter and contributes no relations (its C-CONT.RELS is empty). The generator isn't guided by surface forms, so it will happily keep adding that affix forever. Note that in this case, you actually get the *"Probably circular lexical rule" error.
  2. A unary phrase structure rule is written so that the mother is compatible with the daughter and the rule contributes no relations (its C-CONT.RELS is empty). In this case, you are likely to spin on parsing as well, though it's possible that you could have a string that parses without spinning but can't generate because the surface string prevents the parser from exploring the spinning rule as a possibility while the generator is able to explore that possibility.
  3. You have a semantically empty lexical entry (e.g., a complementizer) that can take phrases built out of itself as complements (so that you'd get a CP with another CP as the non-head daughter.) Filter rules can help in this case, but you should also fix the grammar error. Presumably, your grammar is also overgenerating (e.g., parsing strings like Kim thinks that that that that that Sandy slept).
  4. You have a semantically empty lexical entry (e.g., a complementizer) (even one that won't cause spinning on its own) and a phrase structure or lexical rule that causes a valence list (SPR, SUBJ, COMPS) to be underspecified for length. This would happen if the value of any of those features wasn't constrained on the mother. In this case, the generator will happily keep positing another instance of the semantically empty element and attaching it via a head-valence phrase of the appropriate type. Here again, filter rules would keep the generator from spinning, but you also need to fix the grammar bug, as once again your grammar is likely overgenerating.

Related topics


Back to FAQs page

Back to main course page

-- EmilyBender - 02 Nov 2004, 19 May 2005

Topic revision: r2 - 2005-05-20 - 04:42:00 - EmilyBender
 

This site is powered by the TWiki collaboration platformCopyright & by the contributing authors. All material on this collaboration platform is the property of the contributing authors.
Ideas, requests, problems regarding TWiki? Send feedback
Privacy Statement Terms & Conditions