|
Abstract:
Meng & Bader (1998, in press) have shown that case
conflicts are better cues for German garden-path reanalysis than
number mismatches. They argue that this sensitivity to cue-type
supports a diagnosis model of garden-path recovery (Fodor &
Inoue, 1994). They consider two varieties of diagnosis procedure:
(A) F&I's diagnosis system attaches the incompatible 'symptom'
word into the parse-tree and tries to resolve the resulting
grammatical conflicts. (B) A novel procedure outlined by M&B,
which we call TRIAGE, makes a prior evaluation of whether a
revision attempt is worthwhile. It estimates whether a current
problem is likely to have been due to an error during first-pass
processing. If so, type (A) diagnosis proceeds as usual. If not,
the problem is judged to be in the sentence itself, hence
incorrigible.
Triage could be very advantageous if it saves the parser from
wasting resources on futile reanalyses. And M&B show how triage
might account for the low rate of reanalysis for German
number-misagreement sentences (see below). But triage adds an extra
mechanism to the parser, with the power to engage in meta-level
reasoning about the parser's own functions. The question is: does
the HSPM have this extra triage capability? The following facts
indicate that it does not.
1) If a sentence is rejected at the triage stage, the meaning it
is assigned should reflect the incorrect first-pass structure, not
a post-revision structure. HOWEVER: comprehension questions
designed to distinguish these two interpretations, following
M&B's number-mismatch sentences, received 88% correct
answers.
2) Triage as explanation for reanalysis failure for
number-misagreement presupposes that first-pass number errors are
rare (unlike case errors). It is true that number is usually
morphologically unambiguous in German. HOWEVER: perceived
number-misagreement can result from mistakes about which NP is the
subject, and subject-assignment errors are NOT uncommon in
German.
3) Triage rejection rates should be constant for the same
symptom-type. HOWEVER: in some contexts number misagreement does
trigger successful reanalysis. E.g., the first-pass preference for
singular "sie" in (i) is readily shifted to plural by the plural
verb. This implies that reanalysis IS INITIATED for number
symptoms. Success or failure must arise DURING the attempted
repair.
4) If triage exists, it should apply to more than just number
symptoms. HOWEVER: reanalysis success is generally INSENSITIVE to
whether a current problem is attributable to a first-pass parsing
error. For example, low attachment errors (as in "While Mary was
mending the sock fell off her lap") are very common, and yet are
notoriously resistant to reanalysis; see papers in Fodor &
Ferreira (1998).
CONCLUSION: Though interesting and potentially efficient, triage
is not a feature of the HSPM. Success or failure of Type A
diagnosis, without triage, fits the data better. Success occurs
where the needed revision is evident -- in (i) and M&B's
case-conflict examples. Failure occurs where revision runs into
complications -- in M&B's number-mismatch examples, where an
additional ambiguity within the reanalysis misdirects the parser
(F&I, in press).
(i) Maria hat gesagt, dass sie oft schreien.
Maria has said that she/they often shout [PLURAL]
"Maria said that they often shout."
Fodor, J. D. & Ferreira, F. (eds.) [1998] Reanalysis in
Sentence Processing, Kluwer.
Fodor, J. D. & Inoue, A. [1994] The diagnosis and cure of
garden paths. J. Psycholing. Res. 23:407-434.
Fodor, J. D. & Inoue, A. [in press] Garden path diagnosis:
Agreement and case. J. Psycholing. Research.
Meng, M. & Bader, M. [1998] The role of case and number
features in syntactic ambiguity resolution. In M. de Vincenzi &
V. Lombardo (eds.) Cross-linguistic Perspectives on Language
Processing. Kluwer.
|
