Neo-Aramaic/Transducer - Revision history

Cdalton2 at 01:16, 22 March 2018

2018-03-22T01:16:04Z

Cdalton2 at 21:07, 21 March 2018

2018-03-21T21:07:59Z

Cdalton2 at 20:46, 21 March 2018

2018-03-21T20:46:41Z

Cdalton2 at 20:38, 21 March 2018

2018-03-21T20:38:24Z

Cdalton2 at 20:37, 21 March 2018

2018-03-21T20:37:43Z

Cdalton2 at 20:34, 21 March 2018

2018-03-21T20:34:00Z

Cdalton2 at 20:25, 21 March 2018

2018-03-21T20:25:27Z

Cdalton2 at 20:24, 21 March 2018

2018-03-21T20:24:44Z

Cdalton2 at 18:43, 21 March 2018

2018-03-21T18:43:18Z

Cdalton2 at 16:39, 20 March 2018

2018-03-20T16:39:06Z

← Older revision		Revision as of 01:16, 22 March 2018
Line 193:		Line 193:

	When testing for generation, we have 45 passing morph tests and 56 failing morph tests (44.55%).		When testing for generation, we have 45 passing morph tests and 56 failing morph tests (44.55%).
		+
		+	===Final evaluation of morphological generation===
		+
		+	Now there are 59 passing tests and 18 failing tests (76.62%). We didn't add any twol rules. The only thing we did was comment out one of the twol rules we already have so that the other one can stand as the default. Instead, we added some stems and paths to the transducer. In the yaml file, we originally had a series of long-form s-suffixes for present-tense verbs. However, in the transducer, the long-form s-suffixes are set to analyze, but not generate, so those tests failed during generation. So we took the long form s-suffix tests out of the yaml file and replaced them with the default form s-suffixes. We had removed the default form s-suffixes a while ago because there was something about that particular section that caused errors when we tried to run morftest, but they worked just fine after we retyped them by hand.
		+
		+	Our corpus still has 2595 forms, but coverage has strangely dropped to 13.06%.

	[[Category:Sp18_Transducers]]		[[Category:Sp18_Transducers]]

← Older revision		Revision as of 21:07, 21 March 2018
Line 183:		Line 183:

	After we added this word, our coverage jumped to 15.26%.		After we added this word, our coverage jumped to 15.26%.
		+
		+	==Generator Evaluation==
		+
		+	===Initial evaluation of morphological generation===
		+
		+	We have 53 passing morph tests and 30 failing morph tests in aii.yaml (63.86%).
		+
		+	We have 2595 tokenised words in our corpus, with a coverage of 15.26%.
		+
		+	When testing for generation, we have 45 passing morph tests and 56 failing morph tests (44.55%).

	[[Category:Sp18_Transducers]]		[[Category:Sp18_Transducers]]

@@ Line 1: / Line 1: @@
 The code for the transducer can be found in [https://github.swarthmore.edu/cdalton2/ling073-aii this] Github repository.
-==Getting set up: Part One==
+==Analyzer Evaluation==
 Getting set up for this stage of the project was an ongoing process for us; while it was easy to bootstrap our transducer and start working with Apertium, we found that grammar tags, our <code>Alphabet</code>, and our <code>tests/</code> all needed continuous updating. This was because our grammar rules used tags and characters not found in our defined tags or alphabet, and these changes had to trickle down at some point to <code>tests/</code>, too. We continue to update these to reflect the growing number of tags we use to complete our grammar coverage. We now have tags for verb features, parts of speech, an archiphoneme, person morphology, gender morphology, number morphology, case morphology, TAM morphology, pronoun morphology, and punctuation.
@@ Line 20: / Line 22: @@
   ! %{A%}:ܐ
-==The hard stuff: Part One==
+===The hard stuff: Part One===
 When we moved on to the hardest part of putting together a morphological transducer, our biggest challenge was working with the many prefixes and infixes found in Neo-Aramaic. We focused in the beginning on easy forms, like plural nouns, that used mostly suffixes in their morphology. When we moved on to our prefix morphology, we used the solution suggested on the course wiki, where our twoc file only processes forms with a plus and minus version of any given feature:
@@ Line 42: / Line 44: @@
 This guarantees that if a verb picks up a <code>-hab</code> or <code>-fut</code> tag, it will only analyze these forms as having a habitual or future prefix if they pick up a later <code>+hab</code> or <code>+fut</code> tag.
-== Evaluation ==
+=== Evaluation ===
 We have 2595 tokenised words in our corpus, with a coverage of 15.26%. We have about 37 stems in our transducer (definitely a place we should be concentrating work right now).
@@ Line 134: / Line 136: @@
 Currently, 53 of 83 tests in aii.yaml are passing. We have four passing tests in commonwords.yaml.
-== Notes ==
+=== Notes ===
 Eighteen of the thirty failing tests in aii.yaml are from our "plural nouns" section, in which we gave examples of different ways to form plural nouns. Since there are many different strategies for noun pluralization in Neo-Aramaic, not all of which are totally regular, we only implemented one strategy in our transducer. Thus the tests that exemplify forms that we didn't implement are still failing.

← Older revision		Revision as of 20:38, 21 March 2018
Line 130:		Line 130:

	12 ܦܐܫ		12 ܦܐܫ
		+

	Currently, 53 of 83 tests in aii.yaml are passing. We have four passing tests in commonwords.yaml.		Currently, 53 of 83 tests in aii.yaml are passing. We have four passing tests in commonwords.yaml.

← Older revision		Revision as of 20:37, 21 March 2018
Line 91:		Line 91:
	Top unknown words in the corpus after we added ܠܹܗ, ܚܕ, ܡܘܼܠܸܕ, ܟܠ:		Top unknown words in the corpus after we added ܠܹܗ, ܚܕ, ܡܘܼܠܸܕ, ܟܠ:

		+	62 ܕ

		+	54 ܘ
		+
		+	52 ܐ
		+
		+	49 ܡ
		+
		+	44 ܒ
		+
		+	40 ܪ
		+
		+	37 ܠ
		+
		+	35 ܢ
		+
		+	31 ܡܢ
		+
		+	28 ܝ
		+
		+	28 ܝܢ
		+
		+	27 ܗܩܘܬܐ
		+
		+	25 ܠܐ
		+
		+	22 ܐܝܬܠܗ
		+
		+	20 ܗ
		+
		+	19 ܵܐ
		+
		+	17 ܫ
		+
		+	16 ܹܐ
		+
		+	13 ܩܐ
		+
		+	12 ܦܐܫ

	Currently, 53 of 83 tests in aii.yaml are passing. We have four passing tests in commonwords.yaml.		Currently, 53 of 83 tests in aii.yaml are passing. We have four passing tests in commonwords.yaml.

@@ Line 44: / Line 44: @@
 == Evaluation ==
-We have 2556 tokenised words in our corpus, with a coverage of 9.66%. We have about 37 stems in our transducer (definitely a place we should be concentrating work right now).
+We have 2595 tokenised words in our corpus, with a coverage of 15.26%. We have about 37 stems in our transducer (definitely a place we should be concentrating work right now).
-Top unknown words in the corpus:
+Top unknown words in the corpus, before we added some of them:
 	 ܠܹܗ
@@ Line 87: / Line 87: @@
 	 ܵܐ
@@ Line 105: / Line 109: @@
   ܡܘܼܠܸܕ:ܡܘܼܠܸܕ PresentVerbInfl; ! molidh "to begat, to conceive"
-Adding this form to our transducer bumped our coverage up to 12.03%!
+Adding this form to our transducer bumped our coverage from 9.66% up to 12.03%!
 However, besides this word, we started noticing that a LOT of these uncovered words are single letters; they don't seem to be actual words per se. But we couldn't figure out why so many letters would be sprinkled in our corpus, except as an error. It turns out these are a copy-paste error, as seen in this sentence taking from a pdf piece of our corpus:

@@ Line 135: / Line 135: @@
 The elephant in our room was our most common word, ܠܹܗ. This word was so common in our corpus in part because it shows up in almost every sentence in the first chapter of Matthew (it's the genealogy of Jesus, so most of the sentences are of the form "and a begat b, and b begat c, and c begat d"—you can see how ܡܘܼܠܸܕ comes in here as well), but we couldn't figure out what it was doing. Eventually, we found a dictionary in which it was explained that ܠܹܗ is one form of a word that is used to mark "an object in close association with a noun." In our transducer, we had this word analyze as a preposition.
-ܠܹܗ:ܠܹܗ PrepositionInfl ; ! leh object marker
+  ܠܹܗ:ܠܹܗ PrepositionInfl ; ! leh object marker
 After we added this word, our coverage jumped to 15.26%.
 [[Category:Sp18_Transducers]]

@@ Line 131: / Line 131: @@
   ܚܕ:ܚܕ RegNounInfl ; ! 'had- "a, an, one"
-All together, adding these words to our transducer brought our coverage up to 12.72%.
+The elephant in our room was our most common word, ܠܹܗ. This word was so common in our corpus in part because it shows up in almost every sentence in the first chapter of Matthew (it's the genealogy of Jesus, so most of the sentences are of the form "and a begat b, and b begat c, and c begat d"—you can see how ܡܘܼܠܸܕ comes in here as well), but we couldn't figure out what it was doing. Eventually, we found a dictionary in which it was explained that ܠܹܗ is one form of a word that is used to mark "an object in close association with a noun." In our transducer, we had this word analyze as a preposition.
 [[Category:Sp18_Transducers]]

← Older revision		Revision as of 18:43, 21 March 2018
Line 89:		Line 89:


−	Currently, 53 of 83 tests in aii.yaml are passing.	+	Currently, 53 of 83 tests in aii.yaml are passing. We have four passing tests in commonwords.yaml.

	== Notes ==		== Notes ==

@@ Line 88: / Line 88: @@
 	 ܵܐ
 == Notes ==
-Currently, 53 of our 83 tests pass. Eighteen of the thirty failing tests are from our "plural nouns" section, in which we gave examples of different ways to form plural nouns. Since there are many different strategies for noun pluralization in Neo-Aramaic, not all of which are totally regular, we only implemented one strategy in our transducer. Thus the tests that exemplify forms that we didn't implement are still failing.
+Eighteen of the thirty failing tests in aii.yaml are from our "plural nouns" section, in which we gave examples of different ways to form plural nouns. Since there are many different strategies for noun pluralization in Neo-Aramaic, not all of which are totally regular, we only implemented one strategy in our transducer. Thus the tests that exemplify forms that we didn't implement are still failing.
 Six failing tests are from the preposition section. Rules relating to prepositions were not among the 10 grammatical features we documented in our wiki page, so our transducer currently doesn't deal with prepositions.