13:10-14:10

Børre, Tommi, Linda


*evaluation corpus should come from boundcorpus
* how big is it?
* how many compound errors do we find there?



which corpus should we use for what:

1. training
2. validation 15%
3. evaluation 15%

admin, facta and science in freecorpus are the biggest
news is the biggest in boundcorpus
in total 6,8 giga
5 are boundcorpus

2.8M        freecorpus/orig/sme/Len
1.1G        freecorpus/orig/sme/admin - training
8.5M        freecorpus/orig/sme/bible - training
4.5M        freecorpus/orig/sme/blogs
437M        freecorpus/orig/sme/facta - training
5.0M        freecorpus/orig/sme/ficti - eval
480K        freecorpus/orig/sme/grammar-realword
 14M        freecorpus/orig/sme/laws - eval
420K        freecorpus/orig/sme/news - eval
 40K        freecorpus/orig/sme/odda_mahppa
275M        freecorpus/orig/sme/science - eval
440K        freecorpus/orig/sme/speccorp
2.1M        freecorpus/orig/sme/wikipedia
1.8G        total

~1.6G training

suggestion:
* leave boundcorpus untouched
* use only freecorpus: 



Which kind of corpus can we use for evaluation of GramDivvun

test-howto.md

```bash
for i in naacl-2021-1 naacl-2021-2 naacl-2021-3 bisect1-993bbab 0695483-20210127 216d00d-20210127
do
  gramcheck_comparator.py se.$i.zcheck $GTFREE/nodalida2019/goldstandard/converted/sme $GTBOUND/nodalida2019/goldstandard/converted/sme 2> /dev/null
done
```

gramcheck_comparator.py  se.$i.zcheck $GTFREE/goldstandard/converted/sme  $GTBOUND/goldstandard/converted/sme 2> /dev/null

-> 
$GTFREE/goldstandard/converted/sme $GTBOUND/goldstandard/converted/sme 2> /dev/null
?

  505  cd $GTBOUND 
  506  convert2xml --goldstandard orig/sme 
  507  ccat -a -c goldstandard/converted/sme|sed 's/¶/\n/g'|wc -l

size of the error marked-up corpus:
❯ ccat -a -c goldstandard/converted/sme|sed 's/¶/\n/g'|wc 
  10754  105446  922199
  
how many syntactic errors do we got in the marked-up corpus:
 grep -r ¥ --include "*.correct.txt" orig/sme
 grep -r "<errorsyn" goldstandard/converted/sme
 
 grep -r "<errorsyn" ~/boundcorpus/goldstandard/converted/sme | fgrep cmp -c
93


TODO:
* script for changing the evaluation corpus into NN format - maybe it's already there
* we have to clear the error corpus of the other errors
* grammarchecker with only compound errors or evaluation only paying attention to syntax cmp - Børre?