User:Francis Tyers/Experiments

TODO

• Do LER in/out domain testing for the en-es setup with news commentary.
• Do BLEU in/out domain testing for the en-es setup with news commentary.
• mk-en: why is TLM LER/BLEU so much better ?
  • (partial) answer: 0-context rules (e.g. defaults) not applying properly. Fixed by running in series. This "solves" the LER issue.
  • (partial) answer: preposition selection is much better. We could try running with ling-default preps.
• Do pairwise bootstrap resampling for each of best baseline + best rules
  • (done) for parallel
• why do eu-es rules not improve over freq ?
  • (partial) answer: some rules do not apply because of tag wankery. See line #129774 in the test corpus. Need to define better how tags work. Perhaps only include tags where ambiguous ?
• why do breton numbers for monolingual rules not approach TLM ?
  • because of crispiness being too low.
• why when we add more data, do the results get worse ?
  • because of crispiness being too low.
• rerun the mk-en stuff with frac counts.
• run br-fr test with huge data.
• try decreasing the C with corpus size.

Corpus stats

Pair	Corpus	Lines	W. (src)	SL cov.	Extracted	Extracted (%)	L. (train)	L. (test)	L (dev)	Uniq. tokens >1 trad.	Avg. trad / word
br-fr	oab	57,305	702,328	94.47%	4,668	8.32	2,668	1,000	1,000	603	1.07
en-es	europarl	1,467,708	30,154,098	98.08%	312,162	22.18	310,162	1,000	1,000	2,082	1.08
eu-es	opendata.euskadi.net	765,115	10,190,079	91.70%	87,907	11.48	85,907	1,000	1,000	1,806	1.30
mk-en	setimes	190,493	4,259,338	92.17%	19,747	10.94	17,747	1,000	1,000	13,134	1.86
sh-mk	setimes

Evaluation corpus

Out of domain

Pair	Lines	Words (L1)	Words (L2)	Ambig. tokens	Ambig. types	Ambig token/type	% ambig	Av. trad/word
en-es	434	9,463	10,280	619	303	2.04	6.54%	-

In domain

Pair	Lines	Words (L1)	Words (L2)	Ambig. tokens	Ambig. types	Ambig token/type	% ambig	Av. trad/word
br-fr	1,000	13,854	13,878	1,163	372	3.13	8.39%	-
en-es	1,000	19,882	20,944	1,469	337	4.35	7.38%	-
eu-es	1,000	7,967	11,476	1,360	412	3.30	17.07%	-
mk-en	1,000	13,441	14,228	3,872	1,289	3.00	28.80%	-

•  % ambig = number of SL tokens with >1 translation

EAMT-style results

Out of domain

LER

Pair	freq	tlm	ling	alig	rules (c>1.5)	rules (c>2.0)	rules (c>2.5)	rules (c>3.0)	rules (c>3.5)	rules (c>4.0)
en-es	— [44.5, 52.0]	— [34.7, 41.9]	667 [24.7, 31.9]	630 [ 21.4 , 28.4 ]	2881 [20.2, 27.2]	2728 [20.2, 27.2]	1683 [20.7, 27.6]	1578 [20.7, 27.6]	1242 [20.7, 27.6]	1197 [20.7, 27.6]

BLEU

Pair	freq	tlm	ling	alig	rules (c>1.5)	rules (c>2.0)	rules (c>2.5)	rules (c>3.0)	rules (c>3.5)	rules (c>4.0)
en-es	[0.1885, 0.2133]	[0.1953, 0.2201]	[0.1832, 0.2067]	[0.1832, 0.2067]	[0.1831, 0.2067]	[0.1830, 0.2067]	[ [0.1828, 0.2063]	[0.1828, 0.2063]	[0.1828, 0.2063]

In domain

LER

$c$ is the "crispiness" ratio, the amount of times an alternative translation is seen in a given context compared to the default translation. So, a $c$ of 2.0 means that the translation appears twice as frequently as the default.

Pair	freq	tlm	ling	alig	rules (c>1.5)	rules (c>2.0)	rules (c>2.5)	rules (c>3.0)	rules (c>3.5)	rules (c>4.0)
br-fr	— [58.9, 64.8]	— [44.2, 50.5]	168 [54.8, 60.7]	115 [28.5, 34.1]	221 [27.8, 33.3]	213 [27.6, 33.0]	159 [26.3, 31.8]	150 [26.1, 31.6]	135 [27.2, 32.8]	135 [27.2, 32.8]
en-es	— [21.0, 25.3]	— [15.1, 18.9]	667 [20.7, 25.1	630 [7.2, 10.0]	2881 [5.9, 8.6]	2728 [6.0, 8.6]	1683 [5.7, 8.3]	1578 [5.7, 8.3]	1242 [6.0, 8.5]	1197 [5.9, 8.6]
eu-es	— [41.1, 46.6]	— [38.8, 44.2]	697 [47.8, 53.0]	598 [16.5, 20.8]	2253 [20.2, 24.7]	2088 [17.2, 21.7]	1382 [16.8, 21.0]	1266 [16.1, 20.4]	1022 [15.9, 20.2]	995 [16.0, 20.3]
mk-en	— [42.4, 46.3]	— [27.1, 30.8]	1385 [28.8, 32.6]	1079 [19.0, 22.2]	1684 [18.5, 21.5]	1635 [18.6, 21.6]	1323 [19.1, 22.2]	1271 [19.0, 22.0]	1198 [19.1, 22.1]	1079 [19.1, 22.1]

BLEU

Pair	freq	tlm	ling	alig	rules (c>1.5)	rules (c>2.0)	rules (c>2.5)	rules (c>3.0)	rules (c>3.5)	rules (c>4.0)
br-fr	— [0.1247, 0.1420]	— [0.1397, 0.1572]	168 [0.1325, 0.1503]	115 [0.1344, 0.1526]	221 [0.1367, 0.1551]	213 [0.1367, 0.1549]	159 [0.1374, 0.1554]	150 [0.1364, 0.1543]	135 [0.1352, 0.1535]	135 [0.1352, 0.1535]
en-es	— [0.2151, 0.2340]	— [0.2197, 0.2384]	667 [0.2148, 0.2337]	630 [0.2208, 0.2398]	2881 [0.2217, 0.2405]	2728 [0.2217, 0.2406]	1683 [0.2217, 0.2407]	1578 [0.2217, 0.2407]	1242 [0.2217, 0.2407]	1197 [0.2217, 0.2408]
eu-es	— [0.0873, 0.1038]	— [0.0921, 0.1093]	697 [0.0870, 0.1030]	598 [0.0972, 0.1149]	2253 [0.0965, 0.1142]	2088 [0.0971, 0.1147]	1382 [0.0971, 0.1148]	1266 [0.0971, 0.1148]	1022 [0.0973, 0.1150]	995 [0.0973, 0.1150]
mk-en	— [0.2300, 0.2511]	— [0.2976, 0.3230]	1385 [0.2337, 0.2563]	1079 [0.2829, 0.3064]	1684 [0.2838, 0.3071]	1635 [0.2834, 0.3067]	1323 [0.2825, 0.3058]	1271 [0.2827, 0.3059]	1198 [0.2827, 0.3059]	1079

Learning monolingually (winner-takes-all)

Setup:

• SL side of the training corpus
• All possibilities translated and scored
• Absolute winners taken
• Rules generated by counting ngrams in the same way as with the parallel corpus, only no alignment needed as it works like an annotated corpus.

Out of domain

LER

Pair	freq	tlm	ling	alig	rules (c>1.5)	rules (c>2.0)	rules (c>2.5)	rules (c>3.0)	rules (c>3.5)	rules (c>4.0)
en-es	[44.5, 52.0]	[34.7, 41.9]	[24.7, 31.9]		[30.2, 37.9]	[30.2, 37.9]	[29.2, 37.0]	[29.3, 36.8]	[29.0, 36.4]	[29.1, 36.5]

BLEU

Pair	freq	tlm	ling	alig	rules (c>1.5)	rules (c>2.0)	rules (c>2.5)	rules (c>3.0)	rules (c>3.5)	rules (c>4.0)
en-es	[0.1885, 0.2133]	[0.1953, 0.2201]	[0.1832, 0.2067]		[0.1806, 0.2042]	[0.1806, 0.2042]	[0.1808, 0.2043]	[0.1810, 0.2046]	[0.1809, 0.2045]	[0.1809, 0.2045]

In domain

LER

Pair	freq	tlm	ling	alig	rules (c>1.5)	rules (c>2.0)	rules (c>2.5)	rules (c>3.0)	rules (c>3.5)	rules (c>4.0)
br-fr	— [58.9, 64.8]	— [44.2, 50.5]	168 [54.8, 60.7]	115	261 [53.5, 59.2]	247 [52.1, 58.2]	172 [54.3, 60.2]	165 [52.7, 58.4]	138 [50.5, 56.3]	136 [50.6, 56.6]
en-es	— [21.0, 25.3]	— [15.1, 18.9]	667 [20.7, 25.1]	? ?	2595 [15.0, 19.0]	2436 [15.1, 19.1]	1520 [13.7, 17.6]	1402 [13.6, 17.3]	1065 [13.9, 17.7]	1024 [13.9, 17.8]
eu-es	— [41.1, 46.6]	— [38.8, 44.2]	? [47.8, 53.0]	?	2631 [40.9, 46.4]	2427 [40.9, 46.5]	1186 [40.7, 46.1]	1025 [40.7, 46.2]	685 [40.5, 45.9]	641 [40.5, 45.9]
mk-en	— [42.4, 46.3]	— [27.1, 30.8]	1385 [28.8, 32.6]	?	1698 [27.8, 31.5]	1662 [27.8, 31.4]	1321 [27.8, 31.4]	1285 [27.8, 31.4]	1186 [27.7, 31.4]	1180 [27.7, 31.4]

BLEU

Pair	freq	tlm	ling	alig	rules (c>1.5)	rules (c>2.0)	rules (c>2.5)	rules (c>3.0)	rules (c>3.5)	rules (c>4.0)
br-fr	— [0.1247, 0.1420]	— [0.1397, 0.1572]	168 [0.1325, 0.1503]	115	261 [0.1250, 0.1425]	247 [0.1252, 0.1429]	172 [0.1240, 0.1412]	165 [0.1243, 0.1416]	138 [0.1255, 0.1429]	136 [0.1255, 0.1429]
en-es	— [0.2151, 0.2340]	— [0.2197, 0.2384]	667 [0.2148, 0.2337]	?	2595 [0.2180, 0.2371]	2436 [0.2180, 0.2372]	1520 [0.2190, 0.2380]	1402 [0.2190, 0.2381]	1065 [0.2189, 0.2380]	1024 [0.2189, 0.2380]
eu-es	— [0.0873, 0.1038]	— [0.0921, 0.1093]	? [0.0870, 0.1030]	?	2631 [0.0875, 0.1040]	2427 [0.0878, 0.1042]	1186 [0.0878, 0.1043]	1025 [0.0878, 0.1043]	685 [0.0879, 0.1043]	641 [0.0879, 0.1043]
mk-en	— [0.2300, 0.2511]	— [0.2976, 0.3230]	1385 [0.2567, 0.2798]		1698 [0.2694, 0.2930]	1662 [0.2695, 0.2931]	1321 [0.2696, 0.2935]	1285 [0.2696, 0.2935]	1186 [0.2696, 0.2934]	1180 [0.2696, 0.2934]

Learning monolingually (fractional counts)

Setup:

• SL side of the training corpus
• All possibilities translated and scored
• Probabilities normalised into fractional counts (e.g. add them up to get a total, then divide each prob by the total).
  • log prob converted into normal prob using exp10()
• Rules generated by counting fractions from the translated file.

MaxEnt with alignments

Pair	alig	rule-best	ME (>5)	ME (>3)
br-fr	33.4	31.5	31.8	29.9
mk-en	19.9	19.8	18.9	17.8
eu-es	18.8	16.8	19.7	20.0
en-es	8.6	7.0	6.3	6.3

Notes

• Experiments in Domain Adaptation for Statistical Machine Translation
• Semi-supervised model adaptation for statistical machine translation
• Domain Adaptation for Statistical Machine Translation with Monolingual Resources

  "We found that the largest gain (25% relative) is achieved when in-domain data are available for the target language. A smaller performance improvement is still observed (5% relative) if source adaptation data are available. We also observed that the most important role is played by the LM adaptation, while the adaptation of the TM and RM gives consistent but small improvement."