Difference between revisions of "Lingala and Kikuyu/Structural Transfer"
From LING073
(Created page with "This page details transfer rules for translation between Lingala and Kikuyu ===Pre-Evaluation== ===Implementation=== ===Post-Evalatuion===") |
(→kik → lin evaluation) |
||
| (11 intermediate revisions by the same user not shown) | |||
| Line 1: | Line 1: | ||
This page details transfer rules for translation between Lingala and Kikuyu | This page details transfer rules for translation between Lingala and Kikuyu | ||
| − | + | ==Pre-Evaluation== | |
| + | Below are the evaluation statistics for both the mini test corpus and the full parallel corpus: | ||
| + | ===lin → kik evaluation=== | ||
| + | Evaluation of mini corpus: | ||
| + | * WER: 87.50% | ||
| + | * PER: 87.50% | ||
| + | * Percentage unknown words: 0.0% | ||
| + | * Percentage known words: 100% | ||
| − | + | Evaluation of full corpus: | |
| + | * WER: 99.06% | ||
| + | * PER: 94.34% (unknown-word marks removed) | ||
| + | * Percentage unknown words: 66.67% | ||
| + | * Percentage known words: 33.33% | ||
| − | ===Post-Evalatuion=== | + | ===kik → lin evaluation=== |
| + | Evaluation of mini corpus: | ||
| + | * WER: 80.00% | ||
| + | * PER: 80.00% | ||
| + | * Percentage of unknown words: 0.0% | ||
| + | * Percents of known words: 100% | ||
| + | |||
| + | Evaluation of full corpus: | ||
| + | * WER: 98.13% | ||
| + | * PER: 90.65% (unknown-word marks remove) | ||
| + | * Percentage of unknown words: 70.75% | ||
| + | * Percentage of known words: 29.25% | ||
| + | |||
| + | ==Implementation== | ||
| + | ===lin → kik=== | ||
| + | We implemented one rule in the Lingala to Kikuyu translation direction to deal with basic verb transfer. It works by mapping Kikuyu tenses to Lingala tenses and adding a default aspect and noun class. The mapping of the Lingala negative particle te to a negative tag on the verb in the same sentence is done by the Lingala transducer using constraint grammar to add a negative tag to the Lingala verb whenever te appears in the same sentence. The sentence used in our second contrastive grammar point ({{transferTest|lin|kik|ezalí malámu te|ti wega}}) can be used to illustrate this rule: | ||
| + | *Tagger: <code>^kozala<v><pres><p3><sp><nn><neg>$ ^malámu<adj>$ ^te<neg>$^.<sent>$</code> | ||
| + | *Biltrans: <code>^kozala<v><pres><p3><sp><nn><neg>/rĩ<cop><pres><p3><sp><nn><neg>$ ^malámu<adj>/ega<adj>$ ^te<neg>/$^.<sent>/.<sent>$</code> | ||
| + | *Chunker: <code>^verb<verb>{^rĩ<cop><pres><neg><p3><cl15><sg>$}$ ^default<default>{^ega<adj>$}$ ^sent<SENT>{^.<sent>$}$</code> | ||
| + | *Interchunk: <code>^verb<verb>{^rĩ<cop><pres><neg><p3><cl15><sg>$}$ ^default<default>{^ega<adj>$}$ ^sent<SENT>{^.<sent>$}$</code> | ||
| + | *Postchunk: <code>^rĩ<cop><pres><neg><p3><cl15><sg>$ ^ega<adj>$ ^.<sent>$</code> | ||
| + | *Complete translation: <code>ti wega </code> | ||
| + | |||
| + | ===kik → lin=== | ||
| + | We implemented two rules in the Kikuyu to Lingala translation direction, one of which fixes a number of contrasts between the two languages. The first and simplest dealt with our [[Lingala and Kikuyu/Contrastive Grammar|first contrastive grammar point]], which states that a locative tag on a noun in Kikuyu is realized as a preposition preceding the noun in the Lingala. The sentence used in our contrastive grammar can be used to illustrate the rule: {{transferTest|kik|lin|kĩambĩrĩriainĩ|na ebandeli}}. The output is as follows: | ||
| + | *Tagger: <code>^ambĩrĩria<n><cl7><sg><loc>$^.<sent>$</code> | ||
| + | *Biltrans: <code>^ambĩrĩria<n><cl7><sg><loc>/ebandeli<n><cl7><sg><loc>$^.<sent>/.<sent>$</code> | ||
| + | *Chunker: <code>^noun_locative<loc>{^na<pr>$ ^ebandeli<n><cl7><sg>$}$^sent<SENT>{^.<sent>$}$</code> | ||
| + | *Interchunk: <code>^noun_locative<loc>{^na<pr>$ ^ebandeli<n><cl7><sg>$}$^sent<SENT>{^.<sent>$}$</code> | ||
| + | *Postchunk: <code>^na<pr>$ ^ebandeli<n><cl7><sg>$^.<sent>$</code> | ||
| + | *Complete translation: <code>na ebandeli</code> | ||
| + | |||
| + | The other rule we implemented was a basic verb transfer rule. It works by mapping Kikuyu tenses to Lingala tenses, removing extraneous Kikuyu tags not used in Lingala, and adding the negative particle "te" at the end of a sentence for Kikuyu verbs that are tagged as negative. It also adds a default animacy tag to the Lingala verb. The sentence used in our second contrastive grammar point can be used to illustrate the rule: {{transferTest|kik|lin|ti wega|ezalí malámu te}}. The output is as follows: | ||
| + | *Tagger: <code>^rĩ<cop><pres><neg><p3><cl5><sg>$ ^ega<adj>$^.<sent>$</code> | ||
| + | *Biltrans: <code>^rĩ<cop><pres><neg><p3><cl5><sg>/kozala<v><pres><neg><p3><cl5><sg>$ ^ega<adj>/malámu<adj>$^.<sent>/.<sent>$</code> | ||
| + | *Chunker: <code>^verb<verb>{^kozala<v><pres><p3><sp><nn>$}$ ^default<default>{^malámu<adj>$}$^negative<neg>{ ^te<neg>$}$^sent<SENT>{^.<sent>$}$</code> | ||
| + | *Interchunk: <code>^verb<verb>{^kozala<v><pres><p3><sp><nn>$}$ ^default<default>{^malámu<adj>$}$^negative<neg>{ ^te<neg>$}$^sent<SENT>{^.<sent>$}$</code> | ||
| + | *Postchunk: <code>^kozala<v><pres><p3><sp><nn>$ ^malámu<adj>$ ^te<neg>$^.<sent>$</code> | ||
| + | *Complete translation: <code>ezalí malámu te</code> | ||
| + | |||
| + | ==Post-Evalatuion== | ||
| + | ===lin → kik evaluation=== | ||
| + | N.b. Due to an Apertium bug, the Kikuyu generator cannot generate certain forms that it can analyze. The structural transfer rules that have been implemented do produce valid analyses (see previous section), but the evaluation is limited since these can't be generated due to the bug. | ||
| + | Evaluation of mini corpus: | ||
| + | * WER: 75.00% (technically being calculated as 100%, but this is, for some reason, due to space that is somehow being adding before a period when the negative Lingala particle te is realized as nothing. The 75% figure represents the removal of this space.) | ||
| + | * PER: 75.00% | ||
| + | * Percentage known words: 100.00% | ||
| + | * Percentage unknown words: 0.00% | ||
| + | |||
| + | Evaluation of full corpus: | ||
| + | * WER: 99.06% | ||
| + | * PER: 94.34% (unknown-word marks removed) | ||
| + | * Percentage unknown words: 66.67% | ||
| + | * Percentage known words: 33.33% | ||
| + | |||
| + | ===kik → lin evaluation=== | ||
| + | Evaluation of mini corpus: | ||
| + | * WER: 30.00% | ||
| + | * PER: 30.00% | ||
| + | * Percentage of unknown words: 0.00% | ||
| + | * Percents of known words: 100.00% | ||
| + | |||
| + | Evaluation of full corpus: | ||
| + | * WER: 96.26% | ||
| + | * PER: 86.92% (unknown-word marks remove) | ||
| + | * Percentage of unknown words: 74.77% (Not really sure why this went up. Added some periods to the corpus to make sure the verb negative transfer rule worked properly, but this seems to affect the word count somehow.) | ||
| + | * Percentage of known words: 25.23% | ||
| + | |||
| + | [[Category:Sp17_StructuralTransfer]] | ||
| + | [[Category:Lingala]] | ||
| + | [[Category:Kikuyu]] | ||
Latest revision as of 18:40, 6 April 2017
This page details transfer rules for translation between Lingala and Kikuyu
Contents
Pre-Evaluation
Below are the evaluation statistics for both the mini test corpus and the full parallel corpus:
lin → kik evaluation
Evaluation of mini corpus:
- WER: 87.50%
- PER: 87.50%
- Percentage unknown words: 0.0%
- Percentage known words: 100%
Evaluation of full corpus:
- WER: 99.06%
- PER: 94.34% (unknown-word marks removed)
- Percentage unknown words: 66.67%
- Percentage known words: 33.33%
kik → lin evaluation
Evaluation of mini corpus:
- WER: 80.00%
- PER: 80.00%
- Percentage of unknown words: 0.0%
- Percents of known words: 100%
Evaluation of full corpus:
- WER: 98.13%
- PER: 90.65% (unknown-word marks remove)
- Percentage of unknown words: 70.75%
- Percentage of known words: 29.25%
Implementation
lin → kik
We implemented one rule in the Lingala to Kikuyu translation direction to deal with basic verb transfer. It works by mapping Kikuyu tenses to Lingala tenses and adding a default aspect and noun class. The mapping of the Lingala negative particle te to a negative tag on the verb in the same sentence is done by the Lingala transducer using constraint grammar to add a negative tag to the Lingala verb whenever te appears in the same sentence. The sentence used in our second contrastive grammar point ((lin) ezalí malámu te → (kik) ti wega) can be used to illustrate this rule:
- Tagger:
^kozala<v><pres><p3><sp><nn><neg>$ ^malámu<adj>$ ^te<neg>$^.<sent>$ - Biltrans:
^kozala<v><pres><p3><sp><nn><neg>/rĩ<cop><pres><p3><sp><nn><neg>$ ^malámu<adj>/ega<adj>$ ^te<neg>/$^.<sent>/.<sent>$ - Chunker:
^verb<verb>{^rĩ<cop><pres><neg><p3><cl15><sg>$}$ ^default<default>{^ega<adj>$}$ ^sent<SENT>{^.<sent>$}$ - Interchunk:
^verb<verb>{^rĩ<cop><pres><neg><p3><cl15><sg>$}$ ^default<default>{^ega<adj>$}$ ^sent<SENT>{^.<sent>$}$ - Postchunk:
^rĩ<cop><pres><neg><p3><cl15><sg>$ ^ega<adj>$ ^.<sent>$ - Complete translation:
ti wega
kik → lin
We implemented two rules in the Kikuyu to Lingala translation direction, one of which fixes a number of contrasts between the two languages. The first and simplest dealt with our first contrastive grammar point, which states that a locative tag on a noun in Kikuyu is realized as a preposition preceding the noun in the Lingala. The sentence used in our contrastive grammar can be used to illustrate the rule: (kik) kĩambĩrĩriainĩ → (lin) na ebandeli. The output is as follows:
- Tagger:
^ambĩrĩria<n><cl7><sg><loc>$^.<sent>$ - Biltrans:
^ambĩrĩria<n><cl7><sg><loc>/ebandeli<n><cl7><sg><loc>$^.<sent>/.<sent>$ - Chunker:
^noun_locative<loc>{^na<pr>$ ^ebandeli<n><cl7><sg>$}$^sent<SENT>{^.<sent>$}$ - Interchunk:
^noun_locative<loc>{^na<pr>$ ^ebandeli<n><cl7><sg>$}$^sent<SENT>{^.<sent>$}$ - Postchunk:
^na<pr>$ ^ebandeli<n><cl7><sg>$^.<sent>$ - Complete translation:
na ebandeli
The other rule we implemented was a basic verb transfer rule. It works by mapping Kikuyu tenses to Lingala tenses, removing extraneous Kikuyu tags not used in Lingala, and adding the negative particle "te" at the end of a sentence for Kikuyu verbs that are tagged as negative. It also adds a default animacy tag to the Lingala verb. The sentence used in our second contrastive grammar point can be used to illustrate the rule: (kik) ti wega → (lin) ezalí malámu te. The output is as follows:
- Tagger:
^rĩ<cop><pres><neg><p3><cl5><sg>$ ^ega<adj>$^.<sent>$ - Biltrans:
^rĩ<cop><pres><neg><p3><cl5><sg>/kozala<v><pres><neg><p3><cl5><sg>$ ^ega<adj>/malámu<adj>$^.<sent>/.<sent>$ - Chunker:
^verb<verb>{^kozala<v><pres><p3><sp><nn>$}$ ^default<default>{^malámu<adj>$}$^negative<neg>{ ^te<neg>$}$^sent<SENT>{^.<sent>$}$ - Interchunk:
^verb<verb>{^kozala<v><pres><p3><sp><nn>$}$ ^default<default>{^malámu<adj>$}$^negative<neg>{ ^te<neg>$}$^sent<SENT>{^.<sent>$}$ - Postchunk:
^kozala<v><pres><p3><sp><nn>$ ^malámu<adj>$ ^te<neg>$^.<sent>$ - Complete translation:
ezalí malámu te
Post-Evalatuion
lin → kik evaluation
N.b. Due to an Apertium bug, the Kikuyu generator cannot generate certain forms that it can analyze. The structural transfer rules that have been implemented do produce valid analyses (see previous section), but the evaluation is limited since these can't be generated due to the bug. Evaluation of mini corpus:
- WER: 75.00% (technically being calculated as 100%, but this is, for some reason, due to space that is somehow being adding before a period when the negative Lingala particle te is realized as nothing. The 75% figure represents the removal of this space.)
- PER: 75.00%
- Percentage known words: 100.00%
- Percentage unknown words: 0.00%
Evaluation of full corpus:
- WER: 99.06%
- PER: 94.34% (unknown-word marks removed)
- Percentage unknown words: 66.67%
- Percentage known words: 33.33%
kik → lin evaluation
Evaluation of mini corpus:
- WER: 30.00%
- PER: 30.00%
- Percentage of unknown words: 0.00%
- Percents of known words: 100.00%
Evaluation of full corpus:
- WER: 96.26%
- PER: 86.92% (unknown-word marks remove)
- Percentage of unknown words: 74.77% (Not really sure why this went up. Added some periods to the corpus to make sure the verb negative transfer rule worked properly, but this seems to affect the word count somehow.)
- Percentage of known words: 25.23%
