Difference between revisions of "Guarani and Warlpiri/Structural transfer"

From LING073
Jump to: navigation, search
(→‎Pre-Evaluation)
(→‎Post-Evaluation)
Line 41: Line 41:
 
*Word error rate (WER): 115.52 %
 
*Word error rate (WER): 115.52 %
 
*Position-independent word error rate (PER): 115.52 %
 
*Position-independent word error rate (PER): 115.52 %
 +
 +
===gug → wbp ===
 +
Results when running on gug.sentences.txt:
 +
*Word error rate (WER): 100.00 %
 +
*Position-independent word error rate (PER): 100.00 %
  
 
[[Category:Sp17_StructuralTransfer]]
 
[[Category:Sp17_StructuralTransfer]]

Revision as of 02:53, 7 April 2017

Pre-Evaluation

gug → wbp

Results when running on gug.sentences.txt:

  • Word error rate (WER): 113.79 %
  • Position-independent word error rate (PER): 113.79 %

wbp → gug

Results when running on gug.tests.txt:

  • Word error rate (WER): 83.33 %
  • Position-independent word error rate (PER): 83.33 %

Results when running on gug.sentences.txt:

  • Word error rate (WER): 100.0 %
  • Position-independent word error rate (PER): 100.0 %

(I think the WER and PER got worse because nouns are no longer translating correctly after adding the <subj> tag to all nouns and pronouns).

Structural Transfer Rule

gug → wbp

  • Sample sentence: mitã okaru → kurdu ka ngarni
  • tagger: ^mitã<n><sg><subj>$ ^karu<v><tv><ar><pres><p3><sp>$^.<sent>$
  • biltrans: ^mitã<n><sg><subj>/kurdu<n><sg><subj>$ ^karu<v><tv><ar><pres><p3><sp>/ngarni<v><tv><ar><pres><p3><sp>$^.<sent>/.<sent>$
  • chunker:
    • apertium-transfer: Rule 2 mitã<n><sg><subj>/kurdu<n><sg><subj>
    • apertium-transfer: Rule 3 karu<v><tv><ar><pres><p3><sp>/ngarni<v><tv><ar><pres><p3><sp>
    • apertium-transfer: Rule 1 .<sent>/.<sent> ^n<SN><sg><subj>{^kurdu<n><2>$}$ ^vaux<Saux><SubjTBD>{^ka<vaux><pres><2>$}$ ^v<SV>{^ngarni<v><tv><pres>$}$^sent<SENT>{^.<sent>$}$
  • postchunk: ^kurdu<n><sg>$ ^ka<vaux><pres><subj3sg>$ ^ngarni<v><tv><pres>$^.<sent>$
  • gug-wbp: #kurdu #ka #ngarni# (there is a lingering issue in my generator)

wbp → gug

  • add ar, air, chen tags to dix
  • do reverse of previous steps


Post-Evaluation

gug → wbp

Results when running on gug.sentences.txt:

  • Word error rate (WER): 115.52 %
  • Position-independent word error rate (PER): 115.52 %

gug → wbp

Results when running on gug.sentences.txt:

  • Word error rate (WER): 100.00 %
  • Position-independent word error rate (PER): 100.00 %