bboborjan/comic-translator-accents

Accent tables for OCR output

Per language, the most frequent accented spelling of each unaccented word (only when that spelling has at least 70% of the uses), one word per line. Used by comic-translator's free in-browser path to put back accents that OCR drops: fold a word (lowercase, strip combining marks, ł→l, đ→d, ø→o, ß→ss, æ→ae, œ→oe, ı→i, ħ→h, ŀ→l), look it up, keep the casing.

Built from the top 500k words of FrequencyWords (OpenSubtitles 2018) by Hermit Dave, CC BY-SA 4.0. These tables are CC BY-SA 4.0 too.

Score: word tokens weighted by corpus frequency (in-sample). "Before" = words right with every accent stripped, "after" = right after restoring from the table, recall = accented tokens restored right. On real comic OCR (Hungarian, PaddleOCR): accents kept 32% to 78%.

lang	accented tokens	before	after	recall	entries	MB gzip
hu	37.5%	62.5%	96.6%	91.5%	299118	0.90
cs	43.0%	57.0%	96.8%	93.5%	226046	0.67
sk	38.3%	61.7%	97.3%	95.0%	223385	0.68
pl	34.3%	65.7%	96.9%	92.8%	156876	0.46
de	6.8%	93.2%	99.1%	88.0%	76136	0.29
fr	11.8%	88.2%	96.1%	69.6%	69577	0.23
es	13.1%	86.9%	95.2%	67.5%	69723	0.22
pt	15.8%	84.2%	96.5%	80.1%	52255	0.18
pt-BR	16.7%	83.3%	96.8%	82.2%	46579	0.16
ro	24.9%	75.1%	91.1%	78.2%	68082	0.22
hr	16.0%	84.0%	98.8%	96.5%	92156	0.25
bs	14.6%	85.4%	97.3%	95.5%	88195	0.24
sr	15.7%	84.3%	98.7%	96.1%	95020	0.26
sl	13.5%	86.5%	99.2%	95.0%	86890	0.24
tr	38.8%	61.2%	96.4%	91.6%	249260	0.67
vi	82.1%	17.9%	67.1%	61.2%	4351	0.02
lt	33.0%	67.0%	94.8%	86.7%	124386	0.38
lv	41.1%	58.9%	93.1%	84.4%	75555	0.23
et	15.4%	84.6%	99.4%	96.6%	109087	0.34
fi	25.7%	74.3%	99.2%	97.3%	103240	0.31
is	50.6%	49.4%	99.0%	98.6%	149441	0.54
ca	13.2%	86.8%	97.3%	85.0%	11034	0.04
gl	15.0%	85.0%	96.8%	82.4%	20343	0.07
da	11.6%	88.4%	99.4%	95.5%	101246	0.38
sv	24.0%	76.0%	97.6%	91.8%	139730	0.53
nb	11.1%	88.9%	99.1%	95.7%	69460	0.26
sq	35.3%	64.7%	93.5%	92.1%	49121	0.15
eo	12.1%	87.9%	99.4%	96.6%	6359	0.02

Skipped (gain under 2 points): it, nl, af, eu.