Difference between revisions of "User:Techievena/GSoC 2018 Work Product Submission"
Jump to navigation
Jump to search
Techievena (talk | contribs) (Add work done) |
Techievena (talk | contribs) (Add work to be done) |
||
| Line 27: | Line 27: | ||
https://apertium.projectjj.com/gsoc2018/techievena.html |
https://apertium.projectjj.com/gsoc2018/techievena.html |
||
---- |
|||
=== Extend lttoolbox to have the power of HFST === |
=== Extend lttoolbox to have the power of HFST === |
||
| Line 246: | Line 247: | ||
==== Challenges ==== |
==== Challenges ==== |
||
==== Work to be done ==== |
==== Work to be done ==== |
||
Now that we can weight our morphological analysers, generators and bilingual dictionaries. |
|||
Here are some problems that can be solved: |
|||
* Having zero-context rules in your .lrx files. Now you can just put the weights directly in your bilingual dictionary |
|||
<pre> |
|||
$ echo "^estación<n><f><sg>$" | lt-proc -W -b testbidix.bin |
|||
^estación<n><f><sg>/season<n><W:1.000000><sg>/station<n><W:1.500000><sg>$ |
|||
$ echo "^estación<n><f><sg>$" | lt-proc -b testbidix.bin |
|||
^estación<n><f><sg>/season<n><sg>/station<n><sg>$ |
|||
</pre> |
|||
Analyses will be output according to lowest weight first. So you can mark your default translation as "1.0" and then all others as >1.0 ... because of how transfer works, it will always take the first, which will be the one with the lowest weight. |
|||
* Improving POS-tagging accuracy by ordering analyses by probability. This way if your CG doesn't mop up all the ambiguity, you will get the best remaining analysis. This works kind of like the unigram tagger, but because it can be in the analyser itself, it can be easier to control. |
|||
* Dealing with non-standard forms, instead of having to use LR/RL direction restrictions, you can just make non-standard forms have a high weight and ask for lt-proc to only generate the surface form with the lowest weight. |
|||
==== Thanksgiving ==== |
==== Thanksgiving ==== |
||
Revision as of 05:22, 9 August 2018
Apertium
Abinash Senapati
Hi I am Abinash, a final year undergraduate in the department of Electronics and Electrical Communication Engineering, Indian Institute of Technology, Kharagpur. I was a Google Summer of Code student for Apertium over the summer of 2018 and primarily worked on lttoolbox and apertium-core. My project involved extending the capability of performing morphographemics and adding lexical weights to the lttoolbox transducer in order to enable more complex translations with the transducer.
Project Title
Extend lttoolbox to have the power of HFST
GSoC Blog
https://techievena.github.io/categories/GSoC
Public Profiles
GitHub: Techievena
GitLab: Techievena
IRC nick: Techievena
Apertium wiki: Techievena
E-mail: abinashsena@gmail.com
Mentors
Francis Tyers and Tommi Pirinen
Patches
tarball: Download
zip: Download
Link to commits and repositories I have worked on
https://apertium.projectjj.com/gsoc2018/techievena.html
Extend lttoolbox to have the power of HFST
Work Done
CODING CHALLENGE: https://github.com/Techievena/lexc2dix
MORPHOGRAPHEMICS:
WEIGHTS:
- att_compiler: Support for weights to lttoolbox binary format
Make all the tweaks necessary to have a minimal implementation of weight based analyses in the att_compiler.
$ cat test.att 0 1 c c 4.567895 1 2 a a 0.989532 2 3 t t 2.796193 3 4 @0@ + -3.824564 4 5 @0@ n 1.824564 5 0.525487 4 5 @0@ v 2.845989 $ lt-comp lr test.att test.bin main@standard 6 6 $ lt-print test.bin 0 1 c c 4.567895 1 2 a a 0.989532 2 3 t t 2.796193 3 4 ε + -3.824564 4 5 ε n 1.824564 4 5 ε v 2.845989 5 0.525487
- lt-proc: Implement option to output n-best paths
Using the same option names as hfst-proc we add options in lt-proc to output n-best paths using the weight values.
$ echo "cats" | lt-proc test.bin ^cat/cat+n/cat+v$s $ echo "cats" | lt-proc -W test.bin ^cat/cat+n<W:6.353620>/cat+v<W:7.375045>$s $ echo "cats" | lt-proc -N 1 test.bin ^cat/cat+n$s $ echo "cats" | lt-proc -W -N 1 test.bin ^cat/cat+n<W:6.353620>$s
$ lt-proc -h lt-proc: process a stream with a letter transducer USAGE: lt-proc [ -a | -b | -c | -d | -e | -g | -n | -p | -s | -t | -v | -h -z -w ] [-W] [-N N] [-L N] [ -i icx_file ] [ -r rcx_file ] fst_file [input_file [output_file]] Options: -a, --analysis: morphological analysis (default behavior) -b, --bilingual: lexical transfer -c, --case-sensitive: use the literal case of the incoming characters -d, --debugged-gen morph. generation with all the stuff -e, --decompose-nouns: Try to decompound unknown words -g, --generation: morphological generation -i, --ignored-chars: specify file with characters to ignore -r, --restore-chars: specify file with characters to diacritic restoration -l, --tagged-gen: morphological generation keeping lexical forms -m, --tagged-nm-gen: same as -l but without unknown word marks -n, --non-marked-gen morph. generation without unknown word marks -o, --surf-bilingual: lexical transfer with surface forms -p, --post-generation: post-generation -s, --sao: SAO annotation system input processing -t, --transliteration: apply transliteration dictionary -v, --version: version -z, --null-flush: flush output on the null character -w, --dictionary-case: use dictionary case instead of surface case -C, --careful-case: use dictionary case if present, else surface -I, --no-default-ignore: skips loading the default ignore characters -W, --show-weights: Print final analysis weights (if any) -N, --analyses: Output no more than N analyses (if the transducer is weighted, the N best analyses) -L, --weight-classes: Output no more than N best weight classes (where analyses with equal weight constitute a class) -h, --help: show this help
- Allow weights on entries in lttoolbox XML
Modify the DTD and parser to allow weights on entries in lttoolbox XML.
$ cat test.dix
<?xml version="1.0" encoding="UTF-8"?>
<dictionary>
<alphabet>ABCDEFGHIJKLMNOPQRSTUVWXYZÀÁÂÄÅÆÇÈÉÊËÍÑÒÓÔÕÖØÙÚÜČĐŊŠŦŽabcdefghijklmnopqrstuvwxyzàáâäåæçèéêëíñòóôõöøùúüčđŋšŧž-</alphabet>
<sdefs>
<sdef n="n" c="Noun"/>
<sdef n="ma" c="Masculine (animate)"/>
<sdef n="mi" c="Masculine (inanimate)"/>
<sdef n="nt" c="Neuter"/>
<sdef n="f" c="Feminine"/>
<sdef n="sg" c="Singular"/>
<sdef n="du" c="Dual"/>
<sdef n="pl" c="Plural"/>
<sdef n="nom" c="Nominative"/>
<sdef n="gen" c="Genitive"/>
<sdef n="dat" c="Dative"/>
<sdef n="acc" c="Accusative"/>
<sdef n="ins" c="Instrumental"/>
<sdef n="loc" c="Locative"/>
<sdef n="voc" c="Vocative"/>
</sdefs>
<pardefs>
<pardef n="nan__n_ma">
<e w="1.56"><p><l></l><r><s n="n"/><s n="ma"/><s n="sg"/><s n="nom"/></r></p></e>
<e w="2.56"><p><l>a</l><r><s n="n"/><s n="ma"/><s n="sg"/><s n="gen"/></r></p></e>
<e w="3.56"><p><l>ej</l><r><s n="n"/><s n="ma"/><s n="sg"/><s n="dat"/></r></p></e>
<e w="4.56"><p><l>a</l><r><s n="n"/><s n="ma"/><s n="sg"/><s n="acc"/></r></p></e>
<e w="5.56"><p><l>om</l><r><s n="n"/><s n="ma"/><s n="sg"/><s n="ins"/></r></p></e>
<e w="6.56"><p><l>je</l><r><s n="n"/><s n="ma"/><s n="sg"/><s n="loc"/></r></p></e>
<e w="7.56"><p><l>o</l><r><s n="n"/><s n="ma"/><s n="sg"/><s n="voc"/></r></p></e>
<e w="8.56"><p><l>aj</l><r><s n="n"/><s n="ma"/><s n="du"/><s n="nom"/></r></p></e>
<e w="9.56"><p><l>ow</l><r><s n="n"/><s n="ma"/><s n="du"/><s n="gen"/></r></p></e>
<e w="10.56"><p><l>omaj</l><r><s n="n"/><s n="ma"/><s n="du"/><s n="dat"/></r></p></e>
<e w="11.56"><p><l>ow</l><r><s n="n"/><s n="ma"/><s n="du"/><s n="acc"/></r></p></e>
<e w="12.56"><p><l>omaj</l><r><s n="n"/><s n="ma"/><s n="du"/><s n="ins"/></r></p></e>
<e w="13.56"><p><l>omaj</l><r><s n="n"/><s n="ma"/><s n="du"/><s n="loc"/></r></p></e>
<e w="14.56"><p><l>aj</l><r><s n="n"/><s n="ma"/><s n="du"/><s n="voc"/></r></p></e>
<e w="15.56"><p><l>ojo</l><r><s n="n"/><s n="ma"/><s n="pl"/><s n="nom"/></r></p></e>
<e w="16.56"><p><l>ow</l><r><s n="n"/><s n="ma"/><s n="pl"/><s n="gen"/></r></p></e>
<e w="17.56"><p><l>am</l><r><s n="n"/><s n="ma"/><s n="pl"/><s n="dat"/></r></p></e>
<e w="18.56"><p><l>ow</l><r><s n="n"/><s n="ma"/><s n="pl"/><s n="acc"/></r></p></e>
<e w="19.56"><p><l>ami</l><r><s n="n"/><s n="ma"/><s n="pl"/><s n="ins"/></r></p></e>
<e w="20.56"><p><l>ach</l><r><s n="n"/><s n="ma"/><s n="pl"/><s n="loc"/></r></p></e>
<e w="21.56"><p><l>ojo</l><r><s n="n"/><s n="ma"/><s n="pl"/><s n="voc"/></r></p></e>
</pardef>
</pardefs>
<section id="main" type="standard">
<e lm="nan" w="22.56"><i>nan</i><par n="nan__n_ma"/></e>
</section>
</dictionary>
$ lt-comp lr test.dix test-mor.bin
main@standard 35 54
$ lt-print test-mor.bin
0 1 n n 0.000000
1 2 a a 0.000000
2 3 n n 22.560000
3 4 ε <n> 0.000000
3 5 a <n> 0.000000
3 6 e <n> 0.000000
3 7 o <n> 0.000000
3 8 j <n> 0.000000
4 9 ε <ma> 0.000000
5 10 ε <ma> 0.000000
5 11 j <ma> 0.000000
5 12 m <ma> 0.000000
5 13 c <ma> 0.000000
6 14 j <ma> 0.000000
7 15 ε <ma> 0.000000
7 16 j <ma> 0.000000
7 17 m <ma> 0.000000
7 18 w <ma> 0.000000
8 19 e <ma> 0.000000
9 20 ε <sg> 0.000000
10 21 ε <sg> 0.000000
11 22 ε <du> 0.000000
12 23 ε <pl> 0.000000
12 24 i <pl> 0.000000
13 25 h <pl> 0.000000
14 26 ε <sg> 0.000000
15 27 ε <sg> 0.000000
16 28 o <pl> 0.000000
17 29 ε <sg> 0.000000
17 30 a <du> 0.000000
18 31 ε <du> 0.000000
18 32 ε <pl> 0.000000
19 33 ε <sg> 0.000000
20 34 ε <nom> 1.560000
21 34 ε <gen> 2.560000
21 34 ε <acc> 4.560000
22 34 ε <nom> 8.560000
22 34 ε <voc> 14.560000
23 34 ε <dat> 17.560000
24 34 ε <ins> 19.560000
25 34 ε <loc> 20.560000
26 34 ε <dat> 3.560000
27 34 ε <voc> 7.560000
28 34 ε <nom> 15.560000
28 34 ε <voc> 21.560000
29 34 ε <ins> 5.560000
30 34 j <dat> 10.560000
30 34 j <ins> 12.560000
30 34 j <loc> 13.560000
31 34 ε <gen> 9.560000
31 34 ε <acc> 11.560000
32 34 ε <gen> 16.560000
32 34 ε <acc> 18.560000
33 34 ε <loc> 6.560000
34 0.000000
$ echo "nanow" | lt-proc -W test-mor.bin
^nanow/nan<n><ma><du><gen><W:32.120000>/nan<n><ma><du><acc><W:34.120000>/nan<n><ma><pl><gen><W:39.120000>/nan<n><ma><pl><acc><W:41.120000>$
- Other merged pull requests
- https://github.com/apertium/lttoolbox/pull/14 (Fix inconsistencies in the weighted branch)
- https://github.com/apertium/lttoolbox/pull/25 (Use default values in lttoolbox to prevent apertium-separable from failing)
Challenges
Work to be done
Now that we can weight our morphological analysers, generators and bilingual dictionaries. Here are some problems that can be solved:
- Having zero-context rules in your .lrx files. Now you can just put the weights directly in your bilingual dictionary
$ echo "^estación<n><f><sg>$" | lt-proc -W -b testbidix.bin ^estación<n><f><sg>/season<n><W:1.000000><sg>/station<n><W:1.500000><sg>$ $ echo "^estación<n><f><sg>$" | lt-proc -b testbidix.bin ^estación<n><f><sg>/season<n><sg>/station<n><sg>$
Analyses will be output according to lowest weight first. So you can mark your default translation as "1.0" and then all others as >1.0 ... because of how transfer works, it will always take the first, which will be the one with the lowest weight.
- Improving POS-tagging accuracy by ordering analyses by probability. This way if your CG doesn't mop up all the ambiguity, you will get the best remaining analysis. This works kind of like the unigram tagger, but because it can be in the analyser itself, it can be easier to control.
- Dealing with non-standard forms, instead of having to use LR/RL direction restrictions, you can just make non-standard forms have a high weight and ask for lt-proc to only generate the surface form with the lowest weight.
