Difference between revisions of "Judeo-Tat/Transducer"

From LING073
Jump to: navigation, search
(→‎Generator Evaluation)
(→‎Second evaluation of morphological generation)
Line 3: Line 3:
 
The transducer has 43 stems as of 2/14/17.
 
The transducer has 43 stems as of 2/14/17.
 
==Generator Evaluation==
 
==Generator Evaluation==
−
===Second evaluation of morphological generation===
+
===Final evaluation of morphological generation===
 
Between the initial evaluation and the second evaluation, I added another 24 cases.
 
Between the initial evaluation and the second evaluation, I added another 24 cases.
  
Line 13: Line 13:
 
There is a 21.90% coverage of the corpus.  
 
There is a 21.90% coverage of the corpus.  
  
−
Two sets of rules were added to help deal with prefixes.
+
4 rules were added to the twoc file to help deal with prefixes.
−
 
 
  
 
===Initial evaluation of morphological generation===
 
===Initial evaluation of morphological generation===

Revision as of 03:48, 22 February 2017

The code for a Morphological Transducer in Judeo-Tat can be found here.

The transducer has 43 stems as of 2/14/17.

Generator Evaluation

Final evaluation of morphological generation

Between the initial evaluation and the second evaluation, I added another 24 cases.

In jdt.yaml, 87 out of 91 tests pass.

In commonwords.yaml, 6 out of 20 tests pass.


There is a 21.90% coverage of the corpus.

4 rules were added to the twoc file to help deal with prefixes.

Initial evaluation of morphological generation

In commonwords.yaml, 6 out of 20 tests pass.

In jdt.yaml, 57 out of 66 tests pass.

There is a 21.66% coverage of the corpus.

Generation test

In commonwords.yaml, 6 out of 20 tests pass.

In jdt.yaml, 57 out of 77 tests pass.

Analyser Evaluation

Evaluation

Coverage

The corpus has a current coverage of 21.66%.

Before adding specific words for the corpus, there was a coverage for 854 tokenised words of 6.09%.

  • «by adding "эри<post> ↔ эри" to the transducer, coverage went from 6.09% to 8.90%»
  • «by adding "э<art> ↔ э" to the transducer, coverage went from 8.90% to 13.70%»
  • «by adding "эз<abl> ↔ эз" to the transducer, coverage went from 13.70% to 15.57%»
  • «by adding "мэрд<n> ↔ мэрд (man)" to the transducer, coverage went from 15.57% to 18.03%»
  • «by adding "хирс<n> ↔ хирс (bear)" to the transducer, coverage went from 18.03% to 20.61%»
  • «by adding "мугу<vblex> ↔ мугу (says)" to the transducer, coverage went from 20.61% to 21.66%»

Current list of most common unknown words

  • кор (9)
  • е (9)
  • зугьун (9)
  • пирэ (8)
  • ве (6)
  • гэндуьмэ (6)
  • ми (5)
  • Би (5)
  • чугундурэ (5)
  • в (5)
  • Хьэйфи (5)
  • Пирэ (5)
  • гьер (4)
  • корим (4)
  • сохим (4)
  • гуфдире (4)
  • вэсал (4)
  • нэ (4)
  • тамбургъорэ (4)
  • бахш (3)

Test Coverage

In commonwords.yaml, 6 out of 20 tests pass.

In jdt.yaml, 57 out of 66 tests pass.

Notes

There are currently 3 sets of tests that are not working for a total of 9 failed tests.

Compound words

I couldn't get compounds to work. It was difficult using the unfamiliar syntax to do anything more complicated than suffixes.

Prefixes

Prefixes are also difficult due to the unfamiliar syntax.

Multiple words

The final failed tests were those that involved multiple words. I was unable to create a transducer than successfully compiled.