Talk:Constraint-based lexical selection module
Rule formats
Note: Felipe doesn't like "skip".
- can't say I do either … sounds like a command rather than a constraint --unhammer 11:42, 18 November 2011 (UTC)
- also, in the <or>, should we read them as independent of each other? that's a bit confusing since otherwise they're all required and have a certain order --unhammer 11:42, 18 November 2011 (UTC)
The regular expression for the OR below is:
(nazi<adj>[0-9A-Za-z <>]*|totalitari<adj>[0-9A-Za-z <>]*|feixista<adj>[0-9A-Za-z <>]*|franquista<adj>[0-9A-Za-z <>]*|militar<adj>[0-9A-Za-z <>]*|fiscal<adj>[0-9A-Za-z <>]*)
- Francis Tyers 11:53, 18 November 2011 (UTC)
- 1
<rule>
<remove lemma="règim" tags="n.*">
<acception lemma="diet" tags="n.*"/>
</remove>
<or>
<skip lemma="nazi" tags="adj.*"/>
<skip lemma="totalitari" tags="adj.*"/>
<skip lemma="feixista" tags="adj.*"/>
<skip lemma="franquista" tags="adj.*"/>
<skip lemma="militar" tags="adj.*"/>
<skip lemma="fiscal" tags="adj.*"/>
</or>
</rule>
- 2
<rule>
<remove lemma="règim" tags="n.*">
<acception lemma="diet" tags="n.*"/>
</remove>
<or>
<pattern lemma="nazi" tags="adj.*"/>
<pattern lemma="totalitari" tags="adj.*"/>
<pattern lemma="feixista" tags="adj.*"/>
<pattern lemma="franquista" tags="adj.*"/>
<pattern lemma="militar" tags="adj.*"/>
<pattern lemma="fiscal" tags="adj.*"/>
</or>
</rule>
- 3
<rule>
<remove-from lemma="règim" tags="n.*">
<translation lemma="diet" tags="n.*"/>
</remove-from>
<pattern>
<or>
<pattern-item lemma="nazi" tags="adj.*"/>
<pattern-item lemma="totalitari" tags="adj.*"/>
<pattern-item lemma="feixista" tags="adj.*"/>
<pattern-item lemma="franquista" tags="adj.*"/>
<pattern-item lemma="militar" tags="adj.*"/>
<pattern-item lemma="fiscal" tags="adj.*"/>
</or>
</pattern>
</rule>
<rule c="la dona dels seus somnis">
<select-for lemma="dona" tags="n.*">
<translation lemma="wife" tags="n.*"/>
</select>
<pattern>
<pattern-item lemma="de" tags="pr.*"/>
<pattern-item lemma="*" tags="det.pos.*"/>
<pattern-item lemma="somni" tags="n.*"/>
</pattern>
</rule>
- 4
<rule>
<target lemma="règim" tags="n.*">
<remove lemma="diet" tags="n.*"/>
</target>
…
</rule>
<rule c="la dona dels seus somnis">
<target lemma="dona" tags="n.*">
<select lemma="wife" tags="n.*"/>
</target>
…
</rule>
Text
s ("estació" n) ("season" n) (1 "plujós")
s ("estació" n) ("season" n) (2 "plujós")
s ("estació" n) ("season" n) (1 "de") (3 "any")
s ("estació" n) ("station" n) (1 "de") (3 "Línia")
s ("prova" n) ("evidence" n) (1 "arqueològic")
s ("prova" n) ("test" n) (1 "estadístic")
s ("prova" n) ("event" n) (-3 "guanyador") (-2 "de")
s ("prova" n) ("testing" n) (-2 "tècnica") (-1 "de")
s ("joc" n) ("game" n) (1 "olímpic")
s ("joc" n) ("set" n) (1 "de") (2 "caràcter")
r ("pista" n) ("hint" n) (1 "més") (2 "llarg")
r ("pista" n) ("clue" n) (1 "més") (2 "llarg")
r ("motiu" n) ("motif" n) (-1 "aquest") (-2 "per")
s ("carn" n) ("flesh" n) (1 "i") (2 "os")
s ("sobre" pr) ("over" n) (-1 "victòria")
s ("dona" n) ("wife" n) (-1 "*" det pos)
s ("dona" n) ("wife" n) (-1 "el") (1 "de")
s ("dona" n) ("woman" n) (1 "de") (2 "*" det pos) (3 "somni")
r ("patró n) ("pattern" n) (1 "*" np ant)
Jacob's critique of the rule format
By 'skip' you actually means 'match'. And lemma selection should be called <select> :-)
So I'd write
<rule> <match lemma="el"/> <match lemma="dona" tags="n.*"> <select lemma="wife"/> </match> <match lemma="de"/> </rule>
and actually you would like to work on categories as well as lemmas.
Like, to prefer (human) beings and not things before "feel".
<rule> <match tags="n.*"> <select cat="beings"/> </match> <match lemma="feel"/> </rule>
--Jacob Nordfalk 12:22, 30 November 2011 (UTC)
- I like this a lot actually!
<rule>
<match lemma="estació"> <select lemma="season"/></match>
<match lemma="plujós" tags="adj.*"/>
</rule>
It seems much more consistent... Would you allow every match to have a rule, or each rule to have more than one operation ? e.g.
<rule>
<match lemma="juego"> <select lemma="set"/></match>
<match lemma="de"/>
<match lemma="etiqueta"> <select lemma="tag"/></match>
</rule>
- Francis Tyers 11:31, 1 December 2011 (UTC)
Idea de Sergio
<rule> <word source-lemma="pista" source-tags="n.*"> <remove-acception target-lemma="hint" target-tags="n.*"/> <remove-acception target-lemma="clue" target-tags="n.*"/> </word> <word source-lemma="més" source-tags="preadv.*"/> <word source-lemma="llarg" source-tags="adj.*"/> </rule> (10:44:10) Sergio Ortiz Rojas: mi objetivo es darle a la idea de jacob la idea de expresar la idea de bilingüe en la notación (10:48:06) Sergio Ortiz Rojas: yo creo que lo que prefieren jacob y felipe es mejor que lo original (10:48:19) Sergio Ortiz Rojas: la única crítica es que si lo dejáis así será demasiado "programático" (10:48:19) Sergio Ortiz Rojas: es decir (10:48:24) Sergio Ortiz Rojas: demasiado para gente que escribe programas (10:48:26) Sergio Ortiz Rojas: y poco declarativo
Old application strategy
The following is an inefficient implementation of the rule application process:
# s ("prova" n) ("event" n) (-3 "guanyador") (-2 "de")
#
# tipus = "select";
# centre = "^prova<n>.*"
# tl_patro = ["^event<n>.*"]
# sl_patro = {-3: "^guanyador<", -2: "^de<"}
CLASS Rule:
tipus = enum('select', 'remove')
centre = '';
tl_patro = [];
sl_patro = {};
rule_table = {}; # e.g. rule_table["estació"] = [rule1, rule2, rule3];
i = 0
DEFINE ApplyRule(rule, lu):
FOREACH target IN lu.tl:
SWITCH rule.tipus:
'select':
IF target NOT IN rule.tl_patro:
DELETE target
'remove':
IF target IN rule.tl_patro:
DELETE target
FOREACH pair(sl, tl) IN sentence:
FOREACH centre IN rule_table:
IF centre IN sl:
FOREACH rule IN rule_table[centre]:
matched = False
FOREACH context_item IN rule_table[centre][rule]:
IF context_item in sentence:
matched = True
ELSE:
matched = False
# If all of the context items have matched, and none of them have not matched
# if a rule matches break and continue to the pair.
IF matched == True:
sentence[i] = ApplyRule(rule_table[centre][rule], sentence[i])
break
i = i + 1
How Brian writes rules
In order to to come up with lexical rules, I use a wide variety of methods. Generally, I begin by thinking of common English words which can be used in several different contexts. Often, I look around the room for inspiration or write down ambiguous English words that I come across in my reading. Using this list of English words, I consult my English-Spanish dictionary (Collins Spanish Concise Dictionary, 6th Edition) and then check online to see if indeed, the word changes in Spanish based on the context. While checking on the internet, I visit several websites to ensure accuracy. One that I've found to be particularly reliable is www.spanishdict.com, which has an extensive English-Spanish translation dictionary, and covers most contexts of any given English word, along with many English idioms.
After I locate a word both online and in my English-Spanish dictionary, I compare the Spanish word each gives for a English word depending on the context. If the resulting word is the same, in both then I proceed to actually writing the rule. If not, I continue checking to try and come up with a consensus. If after repeated attempts I can't find the best word, I move and on and try a different word. If I am able to find with a consensus word, I move on to writing the rule.
Once I've found the word I want to use in a certain context, I begin writing the rule, paying particular attention to make sure it's a lexical selection rule rather than a multiword unit. Writing the rule is likely the easiest part of the process, as there is a template to follow. After I finish writing all of the rules, I validate first the file I'm writing the rules in, and if it validates I copy the rules into the apertium-en-es.en-es.lrx file. Before sending this file up to the SVN, I check to make sure it too validates.
Usage (old)
$ cat /tmp/test | python apertium-lex-rules.py rules.txt 2>/dev/null ^El<det><def><f><sg>/The<det><def><f><sg>$ ^estació<n><f><sg>/season<n><sg>$ ^més<preadv>/more<preadv>$ ^plujós<adj><f><sg>/rainy<adj><sint><f><sg>$ ^ser<vbser><pri><p3><sg>/be<vbser><pri><p3><sg>$ ^el<det><def><f><sg>/the<det><def><f><sg>$ ^tardor<n><f><sg>/autumn<n><sg>/fall<n><sg>$^,<cm>/,<cm>$ ^i<cnjcoo>/and<cnjcoo>$ ^el<det><def><f><sg>/the<det><def><f><sg>$ ^més<preadv>/more<preadv>$ ^sec<adj><f><sg>/dry<adj><sint><f><sg>$ ^el<det><def><m><sg>/the<det><def><m><sg>$ ^estiu<n><m><sg>/summer<n><sg>$ ^.<sent>/.<sent>$
- With rules
$ cat /tmp/test | python apertium-lex-rules.py rules.txt | apertium-vm -c ca-en.t1x.vmb | apertium-vm -c ca-en.t2x.vmb |\ apertium-vm -c ca-en.t3x.vmb | lt-proc -g ca-en.autogen.bin The rainiest season is the autumn, and the driest the summer.
- With bilingual dictionary defaults
$ cat /tmp/test | apertium-lex-defaults ca-en.autoldx.bin | apertium-vm -c ca-en.t1x.vmb | apertium-vm -c ca-en.t2x.vmb |\ apertium-vm -c ca-en.t3x.vmb | lt-proc -g ca-en.autogen.bin The rainiest station is the autumn, and the driest the summer.
Work plan (12th March 2012)
- Learn rules from a corpus where the best translation is selected by the TLM
- Try changing the parameters of the LM
- Remove stopwords
- Rank on lemmata
- Try neighbouring word permutations, and ranking on LM
- Remove stopwords
- Rank on lemmata
- Run tests on eu-es
| LER | WER | BLEU | |
|---|---|---|---|
| TLM (s5) | 248/603 41.1% | ||
| TLM (l5) | 211/603 34.9% | ||
| TLM (n5) | 271/603 44.9% | ||
| TLM (l5+flip2) | 220/603 36.4% | ||
| TLM (ptes-s5) | -/603 48.4% | ||
| TLM (fres-s5) | 280/603 46.4% |
Programs to make:
- Program that takes biltrans as output, generates all possible combos, translates, scores on language model takes the biltrans output with the highest LM score, and compares against a reference biltrans output.
