Prepare .m2 files

[osyvokon]

M2 specifics:

• Annotations are done on a sentence level.
• Texts are tokenized with Stanza.
• The error type annotations are copied from the corpus
• There's a special document heading sentence added to the beginning of each document.
  • It looks like this: # 0123, where 0123 is the document ID.
  • This adds opportunity to utilize document-level context.

[osyvokon]

Some baselines for reference:

gec-only

=========== Span-Based Correction =============================
                  TP      FP      FN      Prec    Rec     F0.5
LanguageTool      107     18166   2044    0.0059  0.0497  0.0071
Human 1    	  1420    1965    1195    0.4195  0.543   0.4395
Human 2           1420    1195    1965    0.543   0.4195  0.5128
------------ Span-Based Detection -----------------------------
                  TP      FP      FN      Prec    Rec     F0.5
Human 1           1675    1710    940     0.4948  0.6405  0.5184
LanguageTool      873     17393   1813    0.0478  0.325   0.0576
===============================================================

gec-fluency

=========== Span-Based Correction =============================
                  TP      FP      FN      Prec    Rec     F0.5
Human 1           1626    3227    1718    0.3351  0.4862  0.3573
LanguageTool	  121     18540   2735    0.0065  0.0424  0.0078

------------ Span-Based Detection -----------------------------
                  TP      FP      FN      Prec    Rec     F0.5
Human 1           2000    2853    1344    0.4121  0.5981  0.4394
================================================================

Human 1 -- annotator 2 is a hypothesis, annotator 1 is the reference
Human 2 -- annotator 1 is a hypothesis, annotator 2 is the reference
LanguageTool -- version 5.9

[osyvokon]

Merged in #26