Vlax Romani (rmy) subword embeddings

Vocab size vocab model 25 dim 50 dim 100 dim 200 dim 300 dim
1000 vocab model txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix
3000 vocab model txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix
5000 vocab model txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix 		txt | bin
bokeh | umap | matrix

View Vlax Romani subword embeddings in TensorFlow Projector

2-D UMAP plots

Embedding matrix plots

Training corpus sample, encoded with different BPE vocabulary sizes

Vocab sizermywiki sample
original kana puqelpe: kaske qhave? irinelpe: oleske qhave, olake qhave, akalenge qhave, olenge qhave, okolenge qhave ty (dativ). akate muthavela so ko puqipa
albumul de dubut al lui kygo a fost lansat pe data de 00 mai 0000. au fost incluse și alte colaborări cu tom odell,
lesko anav avel andar béal feirste, ''mui le farsetosko'' (irlandikanes), sostar o foro kerdo pashal o than kai i len farset chhorel pes ando atlantik
1000 ▁kana ▁pu q elpe : ▁k as ke ▁qh ave ? ▁i rin elpe : ▁ol eske ▁qh ave , ▁ol ake ▁qh ave , ▁ak al enge ▁qh ave , ▁ol enge ▁qh ave , ▁o k ol enge ▁qh ave ▁ty ▁( d at iv ). ▁ak ate ▁m ut h av ela ▁so ▁ko ▁pu q ipa
▁al b um ul ▁de ▁du b ut ▁al ▁l u i ▁k y go ▁a ▁f ost ▁l ans at ▁pe ▁d ata ▁de ▁00 ▁mai ▁0000 . ▁a u ▁f ost ▁in cl us e ▁ ș i ▁al te ▁c ola b or ă ri ▁c u ▁t om ▁od ell ,
▁l esko ▁an av ▁avel ▁andar ▁b é al ▁f e ir s te , ▁'' m u i ▁le ▁f ar se to sko '' ▁( ir land ikanes ), ▁so st ar ▁o ▁foro ▁ker do ▁pash al ▁o ▁than ▁kai ▁i ▁len ▁f ar s et ▁ch h or el ▁pes ▁ando ▁at l ant ik
3000 ▁kana ▁puq elpe : ▁k aske ▁qhave ? ▁i rinelpe : ▁ol eske ▁qhave , ▁olake ▁qhave , ▁ak al enge ▁qhave , ▁ol enge ▁qhave , ▁o kol enge ▁qhave ▁ty ▁( d ativ ). ▁akate ▁m uth av ela ▁so ▁ko ▁puqipa
▁albumul ▁de ▁du but ▁al ▁lui ▁kygo ▁a ▁fost ▁lansat ▁pe ▁data ▁de ▁00 ▁mai ▁0000. ▁a u ▁fost ▁in cl us e ▁și ▁al te ▁c ola bor ări ▁cu ▁tom ▁od ell ,
▁lesko ▁anav ▁avel ▁andar ▁b é al ▁f e ir s te , ▁'' m ui ▁le ▁f ar se to sko '' ▁( ir land ikanes ), ▁so star ▁o ▁foro ▁kerdo ▁pashal ▁o ▁than ▁kai ▁i ▁len ▁f ars et ▁chh or el ▁pes ▁ando ▁atlant ik
5000 ▁kana ▁puqelpe : ▁kaske ▁qhave ? ▁irinelpe : ▁ol eske ▁qhave , ▁olake ▁qhave , ▁ak alenge ▁qhave , ▁olenge ▁qhave , ▁o kol enge ▁qhave ▁ty ▁( d ativ ). ▁akate ▁muthav ela ▁so ▁ko ▁puqipa
▁albumul ▁de ▁dubut ▁al ▁lui ▁kygo ▁a ▁fost ▁lansat ▁pe ▁data ▁de ▁00 ▁mai ▁0000. ▁au ▁fost ▁in cl use ▁și ▁al te ▁colabor ări ▁cu ▁tom ▁od ell ,
▁lesko ▁anav ▁avel ▁andar ▁b é al ▁f e ir ste , ▁'' m ui ▁le ▁f ar se tosko '' ▁( ir land ikanes ), ▁sostar ▁o ▁foro ▁kerdo ▁pashal ▁o ▁than ▁kai ▁i ▁len ▁fars et ▁chh or el ▁pes ▁ando ▁atlant ik