load_data.md

加载数据

方案1

语种：

语种	语种全称	样本个数	数据来源
ar	arabic	10000	iwslt2017
bg	bulgarian	10000	xnli
bn	bengali	10000	open_subtitles
bs	bosnian	10000	open_subtitles
cs	czech	10000	ecb
da	danish	10000	open_subtitles
de	german	10000	ecb
el	modern greek	10000	ecb
en	english	10000	ecb
eo	esperanto	10000	tatoeba
es	spanish	10000	tatoeba
et	estonian	10000	emea
fi	finnish	10000	ecb
fo	faroese	10000	nordic_langid
fr	french	10000	iwslt2017
ga	irish	10000	multi_para_crawl
gl	galician	3096	tatoeba
gu	gujarati	-	-
hi	hindi	10000	open_subtitles
hi_en	hindi	7180	cmu_hinglish_dog
hr	croatian	10000	hrenwac_para
hu	hungarian	3801	europa_ecdc_tm; europa_eac_tm
hy	armenian	660	open_subtitles
id	indonesian	10000	id_panl_bppt
is	icelandic	2973	europa_ecdc_tm; europa_eac_tm
it	italian	10000	iwslt2017
ja	japanese	10000	iwslt2017
kk	kazakh	-	-
ko	korean	10000	iwslt2017
lt	lithuanian	10000	emea
lv	latvian	4595	europa_ecdc_tm; europa_eac_tm
mr	marathi	10000	tatoeba
mt	maltese	10000	multi_para_crawl
nl	dutch	10000	kde4
no	norwegian	10000	multi_para_crawl
pl	polish	10000	ecb
pt	portuguese	10000	tatoeba
ro	romanian	10000	kde4
ru	russian	10000	xnli
sk	slovak	10000	multi_para_crawl
sl	slovenian	4589	europa_ecdc_tm; europa_eac_tm
sw	swahili	10000	xnli
sv	swedish	10000	kde4
th	thai	10000	xnli
tl	tagalog	10000	multi_para_crawl
tn	serpeti	10000	autshumato
tr	turkish	10000	xnli
ts	dzonga	10000	autshumato
ur	urdu	10000	xnli
vi	vietnamese	10000	xnli
yo	yoruba	9970	menyo20k_mt
zh	chinese	10000	xnli
zu	zulu, south africa	10000	autshumato

备注：

问题1：
训练模型后发现它对短句子的识别能力很差。
当句子长度足够的时候，它能够识别准确，并给出接近1.0 的概率值。
但对于短句子，识别结果让人难以接受。例如把“你好”识别成 de 德语。

解答1：
（1）数据增强，将句子截短。
（2）我认为可考虑加大 dropout，狠狠的 dropout。

方案2

语种：

语种	语种全称	样本个数	数据来源
af	afrikaans	35214	spc
ar	arabic	100000	iwslt2017
bg	bulgarian	100000	xnli
bn	bengali	36064	open_subtitles
bs	bosnian	10212	open_subtitles
cs	czech	100000	emea
da	danish	100000	open_subtitles
de	german	100000	iwslt2017
el	modern greek	100000	emea
en	english	200000	iwslt2017
eo	esperanto	94101	tatoeba; open_subtitles
es	spanish	100000	xnli
et	estonian	100000	emea
fi	finnish	100000	ecb; kde4
fo	faroese	23807	nordic_langid
fr	french	100000	iwslt2017
ga	irish	100000	multi_para_crawl
gl	galician	3096	tatoeba
gu	gujarati	-	-
hi	hindi	100000	xnli
hi_en	hindi	7180	cmu_hinglish_dog
hr	croatian	95844	hrenwac_para
hu	hungarian	3801	europa_ecdc_tm; europa_eac_tm
hy	armenian	660	open_subtitles
id	indonesian	23940	id_panl_bppt
is	icelandic	100000	multi_para_crawl
it	italian	100000	iwslt2017
ja	japanese	100000	iwslt2017
kk	kazakh	-	-
ko	korean	100000	iwslt2017
lt	lithuanian	100000	emea
lv	latvian	100000	multi_para_crawl
mr	marathi	51807	tatoeba
mt	maltese	100000	multi_para_crawl
nl	dutch	100000	kde4
no	norwegian	100000	multi_para_crawl
pl	polish	100000	para_crawl_en_pl
pt	portuguese	100000	para_crawl_en_pt
ro	romanian	100000	iwslt2017
ru	russian	100000	xnli
sk	slovak	100000	multi_para_crawl
sl	slovenian	100000	para_crawl_en_sl
sw	swahili	100000	xnli
sv	swedish	100000	kde4
th	thai	100000	xnli
tl	tagalog	97241	multi_para_crawl
tn	serpeti	100000	autshumato
tr	turkish	100000	xnli
ts	dzonga	100000	autshumato
uk	ukrainian	88533	para_pat_en_uk
ur	urdu	100000	xnli
vi	vietnamese	100000	xnli
yo	yoruba	9970	menyo20k_mt
zh	chinese	200000	xnli
zu	zulu, south africa	26801	autshumato