Autocorrect-Methodologies-Evaluation

In this repository I worked on a word correction evaluation problem, so I selected four auto-correction well-known libraries and evaluated them on a dataset, the dataset has been tackled to tweak the words to generate some incorrect words, and the data generated by swapping, removing, and adding letters.

Dataset

The error function:

def apply_error(word, remove=0.3, add=0.3, swap_letter=True):
  def swap(s, i, j):
    return ''.join((s[:i], s[j], s[i+1:j], s[i], s[j+1:]))

  if swap_letter:
    i = random.randint(0,len(word)-1)
    j = random.randint(0,len(word)-1)
    word = swap(word, i, j)

  letters_to_remove = int(remove*len(word))
  letters_to_add = int(add*len(word))

  for _ in range(letters_to_remove):
    i = random.randint(1,len(word))
    word = word[:i-1] + word[i:]

  lower_upper_alphabet = string.ascii_letters
  for _ in range(letters_to_add):
    random_letter = random.choice(lower_upper_alphabet)
    i = random.randint(1,len(word))
    word = word[:i] + random_letter.lower() + word[i:]

  return word

The deviated words:

	correct	wrong_1
0	perpetrated	perpetrated
1	hazard	hazaard
2	abandon	abandaondon
3	driven	drirviven
4	allow	allllow
...	...	...
195	shamed	smahed
196	singular	singulnagular
197	cabinet	cabinaebinet
198	righteous	rightheteous
199	requires	reruiqes

Models comparison

The used models:

• contextualSpellCheck
• textblob
• pyspellchecker
• autocorrect (Speller)

The model's accuracies:

The word correction of the best model

	correct	wrong_1	pyspellcheckerOutput
0	perpetrated	perpetrated	perpetrated
1	hazard	hazaard	hazard
2	abandon	abandaondon	abandaondon
3	driven	drirviven	drirviven
4	allow	allllow	allow
...	...	...	...
195	shamed	smahed	smashed
196	singular	singulnagular	singulnagular
197	cabinet	cabinaebinet	cabinaebinet
198	righteous	rightheteous	rightheteous
199	requires	reruiqes	requires