Extracting data from Goethe Zertifikat B1 Wortliste pdf

Problem statement

I’m in the process of learning German1. And I think flashcards are the best thing ever.

Unfortunately the canonical German B1 vocabulary comes in the form of a gnarly PDF with four-column layout.

I would like that in a more sane form, so I can easily import it into my flashcard program.

This post details the journey to get there.

Spoiler alert: If you just want the resulting list, it’s linked in the Solution section. You’re welcome.

Detailed problem statement

For the rest of you, let me take you on a “diff driven development”2 journey to take a PDF whose pages look like this:

source example Example text from Goethe-Zertifikat_B1_Wortliste.pdf, page 16

and turn it into text:

Word Example
ab 1. Die Fahrt kostet ab Hamburg 200 Euro.
2. Ab nächster Woche bleibt unser Geschäft samstags geschlossen.
3. Mein Bruder besucht uns ab und zu.
abbiegen, biegt ab, bog ab, ist abgebogen An der nächsten Kreuzung müssen Sie links abbiegen.
die Abbildung, -en Auf der Abbildung sehen Sie, wie man das Gerät einschaltet.
abschreiben, schreibt ab, schrieb ab, hat abgeschrieben Er hat die Hausaufgaben von mir abgeschrieben.
das Abitur (D)
→ A, CH: Matura
Meine Tochter hat gerade Abitur gemacht.
ablehnen, lehnt ab, lehnte ab, hat abgelehnt Er hat mein Angebot, ihm zu helfen, abgelehnt.
abmachen, macht ab, machte ab, hat abgemacht Wir hatten doch abgemacht, dass du die Getränke besorgst.

The final form of the text doesn’t matter that much, HTML table is fine. As long as it can be imported into the Flashcards Deluxe program I’m using for flashcards.

Road to solution

First, I looked to see if someone else had done it. Unfortunately, I couldn’t find that.3

But, since I’m familiar with both pdftotext and tesseract (OSS OCR engine), I thought this must be easy. And normally, yes. Not so with this pdf:

$ pdftotext -f 16 -l 16 Goethe-Zertifikat_B1_Wortliste.pdf - | head -n 20


Alphabetischer Wortschatz

1. Die Fahrt kostet ab Hamburg
200 Euro.
2. Ab nächster Woche bleibt unser
Geschäft samstags geschlossen.
3. Mein Bruder besucht uns ab und zu.


abbiegen, biegt ab,
bog ab, ist abgebogen

An der nächsten Kreuzung müssen Sie
links abbiegen.

If you look closely, the ab follows its own example. Oh boy.

So, bright idea! I can cut up an individual page into columns, and then process each column!

Figuring out where the column boundaries are

In order to figure out the columns, I settled on exporting the pages to pdf, and then merging them one over the other:

pdftocairo -png -r 300 Goethe-Zertifikat_B1_Wortliste.pdf
IMG_SIZE=$(identify Goethe-Zertifikat_B1_Wortliste-*png | \
  awk '{print $3}' | sort -u)
convert -size $IMG_SIZE xc:white boundaries.png
# pages 16 - 102 are interesting (to me)
for i in $(seq -w 16 102); do
  convert boundaries.png G*$i*.png -compose darken -composite boundaries.png

By starting with all-white image and using -compose darken, I’m darkening down the resulting image – thus getting the boundaries clearly visible:

column boundaries (cut) column boundaries (click for full version)

A little bit of mucking around in gimp (with the Guide tool):

column cuts top (… snip …) column cuts bottom column cuts using Guides (click for full version)

gets me the correct coordinates (for 300 dpi):

coord from to
y 320 3260
column 1 140 540
column 2 540 1200
column 3 1300 1710
column 4 1710 2340

So I’ll just OCR the crap out of the individual columns

So I went ahead thinking that I might OCR the way out of it after all:

# Let's cut up page 16 ...
for i in 016; do
  convert G*$i.png -crop $[540-140]x$[3260-320]+140+320 $i-c1.png
  convert G*$i.png -crop $[1200-540]x$[3260-320]+540+320 $i-c2.png
  convert G*$i.png -crop $[1710-1300]x$[3260-320]+1300+320 $i-c3.png
  convert G*$i.png -crop $[2340-1710]x$[3260-320]+1710+320 $i-c4.png

But when it came to the actual OCRing:

# Column 1
$ tesseract -l deu 016-c1.png - | head -n 5
abbiegen, biegt ab,
bog ab, ist abgebogen

die Abbildung, -en

# Column 2
$ tesseract -l deu 016-c2.png - | head -n 5
. Die Fahrt kostet ab Hamburg
200 Euro.

. Ab nächster Woche bleibt unser
Geschäft samstags geschlossen.

# Column 3
$ tesseract -l deu 016-c3.png - | head -n 5
abschreiben, schreibt ab,
schrieb ab,
hat abgeschrieben

das Abitur (DJ)>A, CH:

# Column 4
$ tesseract -l deu 016-c4.png - | head -n 5
Er hat die Hausaufgaben von mir

Meine Tochter hat gerade Abitur

There were all kinds of problems:

  1. Column 1 is missing A and ab
  2. Column 2 is missing numbers
  3. Column 3 has (DJ)>A instead of (D)→A

I mean, see the full page 16 and the associated cuts (col 1, col 2, col 3, col 4) for comparison.

So, this sort of result won’t do.

Fine, I’ll pdftotext the individual columns

# Column 1
$ pdftotext -f 16 -l 16 -r 300 -x 140 -y 320 -W $[540-140] -H $[3260-320] \
  Goethe-Zertifikat_B1_Wortliste.pdf - | head -n 5

abbiegen, biegt ab,
bog ab, ist abgebogen

# Column 2
$ pdftotext -f 16 -l 16 -r 300 -x 540 -y 320 -W $[1200-540] -H $[3260-320] \
  Goethe-Zertifikat_B1_Wortliste.pdf - | head -n 5
1. Die Fahrt kostet ab Hamburg
200 Euro.
2. Ab nächster Woche bleibt unser
Geschäft samstags geschlossen.
3. Mein Bruder besucht uns ab und zu.

# Column 3
$ pdftotext -f 16 -l 16 -r 300 -x 1300 -y 320 -W $[1710-1300] -H $[3260-320] \
  Goethe-Zertifikat_B1_Wortliste.pdf - | head -n 5
abschreiben, schreibt ab,
schrieb ab,
hat abgeschrieben
das Abitur (D)→A, CH:

# Column 4
$ pdftotext -f 16 -l 16 -r 300 -x 1710 -y 320 -W $[2340-1710] -H $[3260-320] \
  Goethe-Zertifikat_B1_Wortliste.pdf - | head -n 5
Er hat die Hausaufgaben von mir

Meine Tochter hat gerade Abitur

And oh boy, that’s so much better! See the full dumps (col 1, col 2, col 3, col 4) if curious.

Only “one” problem – how does one split up what’s the definition and examples?

Fine, I’ll detect longer whitespace across c1+c2 and c3+c4

So, the idea is simple – the individual entries are split up whenever there’s long-ish whitespace in both definition and example columns.

How was I going to do that? No, fuck machine learning. XPM and Ruby, obviously.

So let’s convert the appropriate sections to XPM and extract the breaks:

# which page?

# columns 1 + 2
convert Goethe-Zertifikat_B1_Wortliste-$P.png \
  -crop $[1200-140]x$[3260-320]+140+320 $P-l.xpm
ruby detect-breaks.rb $P-l.xpm > $P-l.txt

# columns 3 + 4
convert Goethe-Zertifikat_B1_Wortliste-$P.png \
  -crop $[2340-1300]x$[3260-320]+1300+320 $P-r.xpm
ruby detect-breaks.rb $P-r.xpm > $P-r.txt

Let’s look at the magical detect-breaks.rb together, shall we?

#!/usr/bin/env ruby

require 'set'

# How many empty lines does there have to be?

# Is this pixels section of the file?
pixels = false

# What is the code for white?
white = nil

# What's the current Y coord of the pixel?
y = 0

# What's the state of our scan?
state = :trail

# Y coord of current start of empty lines
start = nil

# Start of the current rectangle
rect_start = 0

# Overrides
breaks = Hash.new { |h,k| h[k] = Set.new(); h[k] }
# spoiler alert, this will be important later:
#breaks['090-l'] = Set.new([486, 574, 715])

pfx = File.basename(ARGV.first, '.xpm')
File.readlines(ARGV.first).each do |l|
  # not in "pixels" section yet?
  if !pixels 
    if l =~ /"(\w+)\s+c\s+white"/
      white = $1
    if l =~ /^\/\*\s+pixels\s+\*\/$/
      pixels = true

  # we only want pixels here ...
  next unless pixels

  # skip trailing line
  break if pixels && l =~ /^};/

  # is the line empty?
  empty = l =~ /^"(#{white})+",?$/

  # is there an override?
  if breaks[pfx].include?(y)
    state = :overriden
    start = 0
  # teh state machine
  case state
  when :trail
    state = :look if !empty
  when :look
    if empty
      state = :found
      start = y
  when :found, :overriden
    if empty
      if y > start + THRESHOLD
	puts [rect_start, y].join(' ')
	rect_start = y
	state = :trail
      state = :look

  y += 1

# final summation...
puts [rect_start, y].join(' ') unless state == :trail

And the beautiful result for page 16 is:

$ cat 016-l.txt 
0 113
113 380
380 532
532 655
655 843
843 1485
1485 1632
1632 1755
1755 1892
1892 1998
1998 2121
2121 2274
2274 2411
2411 2539
2539 2686
2686 2824

$ cat 016-r.txt 
0 201
201 339
339 476
476 614
614 797
797 934
934 1026
1026 1164
1164 1301
1301 1392
1392 1529
1529 1709
1709 1846
1846 2035
2035 2172
2172 2354
2354 2488
2488 2676
2676 2809
2809 2940

Now, don’t tell me you can’t judge the correctness?! What are you, human?

So, yeah, I can’t either, let’s annotate the breaks, shall we:

# page

# take the png from pdftocairo... and annotate:
cp Goethe-Zertifikat_B1_Wortliste-$P.png $P-annot.png
cat $P-l.txt | ruby annotate.rb $P-annot.png 140 1200 320
cat $P-r.txt | ruby annotate.rb $P-annot.png 1300 2340 320

So together with annotate.rb (which just draws a bunch of rectangles in the specified places):

#!/usr/bin/env ruby

if ARGV.size != 4
  STDERR.puts "Usage: #{File.basename($0)} <f> <x0> <x1> <y>"
  exit 1

f = ARGV.first
x0, x1, y = *ARGV[1,3].map(&:to_i)

coords = []
STDIN.each do |ln|
  coords << ln.split(/\s+/,2).map(&:to_i)

cmd = ["convert", f, "-fill", "transparent", "-stroke", "red"]
for y0, y1 in coords
  cmd += ["-draw", "rectangle #{x0},#{y+y0} #{x1},#{y+y1}"]
cmd << f
system *cmd

we get something readable:

annotated rects (cut) annotated rectangles (click for full version)

I’m almost done… I’ll just extract it now

OK, we have rectangles, let’s extract the texts:


# Left side
ruby extract.rb "$F" $P $P-l.txt 140 540 1200 320 l

# Right side
ruby extract.rb "$F" $P $P-r.txt 1300 1710 2340 320 r

Obviously the interesting part is the extract.rb where the “magic” happens:

#!/usr/bin/env ruby

if ARGV.size != 8
  STDERR.puts "Usage: #{File.basename($0)} <pdf> <page> <yranges> <x0> <x1> <x2> <y> <col>"
  exit 1

pdf = ARGV.first
page = ARGV[1]
coords = []
File.readlines(ARGV[2]).each do |ln|
  coords << ln.split(/\s+/,2).map(&:to_i)
x0, x1, x2, y = ARGV[3,4].map(&:to_i)
col = ARGV[7]

outfile = "#{page}-#{col}.msh"

exit 0 if FileTest.file?(outfile)

out = []

coords.each_with_index do |(y0, y1), idx|
  i, l, r = nil
  i = "#{page}-#{col}-#{idx}.png"
  unless FileTest.file?(i)
    system(*["convert", "Goethe-Zertifikat_B1_Wortliste-#{page}.png",
	     "-crop", "#{x2-x0}x#{y1-y0}+#{x0}+#{y+y0}", "+repage", i])
  IO.popen(["pdftotext", "-f", page, "-l", page, "-r", 300,
	    "-x", x0, "-y", y+y0, "-W", x1-x0, "-H", y1-y0,
	    pdf, "-"].map(&:to_s), 'r') do |f|
    l = f.read.strip
  IO.popen(["pdftotext", "-f", page, "-l", page, "-r", 300,
	    "-x", x1, "-y", y+y0, "-W", x2-x1, "-H", y1-y0,
	    pdf, "-"].map(&:to_s), 'r') do |f|
    r = f.read.strip

  # Some fix-ups elided ...

  out << [i, l, r]

File.open(outfile, "w") { |f| Marshal.dump(out, f) }

The output are marshalled arrays containing [image, def, example]:

$ ruby -e 'require "pp"; pp Marshal.load(File.read("016-l.msh"))[0,5]'
[["016-l-0.png", "A", ""],
  "1. Die Fahrt kostet ab Hamburg\n" +
  "200 Euro.\n" +
  "2. Ab nächster Woche bleibt unser\n" +
  "Geschäft samstags geschlossen.\n" +
  "3. Mein Bruder besucht uns ab und zu."],
  "abbiegen, biegt ab,\n" + "bog ab, ist abgebogen",
  "An der nächsten Kreuzung müssen Sie\n" + "links abbiegen."],
  "die Abbildung, -en",
  "Auf der Abbildung sehen Sie, wie man\n" + "das Gerät einschaltet."],
  "das Abenteuer, -",
  "1. Unser Urlaub war ein tolles Abenteuer.\n" +
  "2. Die Kinder lieben Abenteuergeschichten."]]

The “only” issue now is to generate some useful output from it, yes?

As a foreshadowing – almost. Except for corner cases.

I’ll just generate useful output out of it now

The generation would be almost too easy, if it weren’t for a couple of pesky issues:

  1. The text extraction sometimes damaged lists – first spat out the bullets, then the text.
  2. Since the columns were narrow, there were newlines all over the place.
  3. Some examples with lists crossed column or page boundaries
  4. pdftotext considered dash (-) at the end of line as a word break, even when it made no sense.
  5. Some pages did not have sufficient whitespace between entries, thus they ended up merged together (hence breaks in detect-breaks.rb).

Long story short, I spent about the same time afterwards running cleanup jobs. Manually inserting section breaks between terms4. And then also things like:

# fix up newlines in examples
if e =~ /\A(\d+)\./
  start = $1.to_i
  # list
  e = e.split(/\d+\.\s*/)[1..-1].map { |x| x.strip.tr("\n", " ") }.
    inject([[], start]) { |(o,i),x| [o + ["#{i}. #{x}"], i+1] }.
  # sentence
  e = e.tr("\n", " ")

But I’ll skip these, as the post is getting too long, and most of it is just heavy context-dependent regexp use.

I’ll say this, though: This is where the diff driven development comes from.

Diff driven development

When I wrote the first version of generate.rb, I made it output both csv and html, because that’s what I wanted at the end.

Little did I know that checking the resulting output files to git makes for an excellent debugging tool.

You essentially hack on the generator, and periodically use ruby generate.rb 016; git diff 016.csv to see what’s what.

Prior commit, running the generate on the entire set of pages allows for rather fine testing whether the change was useful.

So, just like TDD uses tests, DDD uses diff. ;)

Final form of generate.rb & sample output

If you’re following along, you might be interested in the final form of generate.rb as well as in final output.

The script is rather unwieldy, and as such, you can find generate.rb in my goethe-b1-wortliste GitHub repo.

The resulting output for page 16: csv, html.


So maybe you only came here for the goods?

Here it is:

(Note: It was last updated 2024-02-17, to fix some issues with missing breaks and formattingdiff of the output)

Please heed the disclaimer:

All of it is extracted from Goethe-Zertifikat_B1_Wortliste.pdf (© 2016 Goethe-Institut und ÖSD) because their PDF was unusable for making flashcards.

It is highly likely you can use this for personal purposes, but I make no claim that I own the resulting data. In other words: if I were you, I wouldn’t go using this in any commercial capacity.

Closing words

Obviously I wouldn’t object to you dropping me an email telling me what you think of it. Or pointing out errors I might have missed5.

Updated 2023-12-11: Also, my friend Izi told me about pdftotext -layout which tries to conserve layout. I think it’s fantastic for general cases (Izi uses it to extract some credit card statement data), but in this case would introduce another problem – hyphenated words would remain so6.

  1. For the better part of a decade, that is.

  2. There! I coined it.

  3. Maybe I suck at searching? In which case, teach me to do better, pls?

  4. Easiest way: open the 016-annot.png in gimp, crop the image from the top red rectangle to the height where you want to place break, read off the image height (say, 118), stick it as breaks['016-l'] = Set.new([118]), re-run page generation, inspect the 016-annot.png if the change took.

  5. Because I did not comb over every single definition just yet.

  6. But, since I re-ran the extraction with -layout (for Science™), I also discovered some more cosmetic issues. Hooray for DDD! And thanks, Izi. ;)