[developers] morphology check in
Dan Flickinger
danf at csli.stanford.edu
Wed Aug 24 13:49:11 CEST 2005
Hi Ann -
I'll have a look at this later today, and send you a diagnosis and perhaps
a corrected grammar tomorrow (Thursday), once I'm back in Oslo.
Dan
>
> Dan - read this please!
>
> OK - I'd tested on the previously checked in version of the ERG which is fine
.
> I suspect the problem is that the current version of the ERG is buggy in the
> sense of having too many cyclically applicable rules for the rule filter to
> work with. As I explained, the rule filter computation exponential in this
> case - the bug in the last version of the code led to this not being obvious
> and cycles probably weren't being checked for properly. According to the val
ue
> of *cyclic-lr-list* when I aborted the computation, the following are cyclic:
>
> (#[rule PUNCT_COLON_RULE o] #[rule PUNCT_PERIOD_RULE o]
> #[rule PUNCT_COMMA_INFORMAL_RULE o] #[rule ATTR_ADJ_VERB_TR_PART ]
> #[rule PUNCT_LPAREN_RULE o] #[rule ATTR_ADJ_VERB_PART ]
> #[rule ENOUGHADD ] #[rule PUNCT_SQLEFT_RULE o] #[rule SAILR ]
> #[rule PUNCT_BANG_RULE o] #[rule PUNCT_SEMICOL_RULE o] #[rule ADVADD ]
> #[rule PUNCT_DQLEFT_RULE o] #[rule MONTHDET ]
> #[rule PUNCT_SQRIGHT_RULE o] #[rule PAST_VERB_INFL_RULE o]
> #[rule PUNCT_QMARK_RULE o] #[rule PUNCT_RPAREN_RULE o]
> #[rule WEEKDAYDET ] #[rule TAGLR ] #[rule PUNCT_DQRIGHT_RULE o]
> #[rule NON_THIRD_SG_FIN_VERB_INFL_RULE ] #[rule PUNCT_COMMA_RULE o]
> #[rule PUNCT_HYPHEN_RULE o])
>
> I haven't checked whether this is in fact true yet - I thought I should
> probably send this message quickly in case Dan is in response mode. There ma
y
> be something I can do if it's thought desirable for all these to apply to the
ir
> own output: my working assumption was that there are a relatively small numbe
r
> of rules where this is desirable, but I am willing to reconsider. We can may
be
> just compute the rule-filter for non-morphologically active lexical rules.
>
> I will check in a change so that the cyclic rules are listed as they are foun
d.
> Current state of the code is that one message is sent when all are found but
> that's obviously too late.
>
> Ann
>
>
>
>
>
>
>
More information about the developers
mailing list