Factor Process

I. Introduction

We exam each error types (training set) to:

• Find the cause of error types
• Fix the error types
  • Enhance the data or algorithm if it is a generic pattern
  • Correct the gold standard (if that is the cause)
• Rerun the program

II. Detail Process

• Tokens are not in Brat annotation data (correct spelling)
  • A2.2.1. Not in checkDic, corrected wrong, by dictionary
    => Need to add these words to checkDic
    => Combo-6 (best performance) is chosen as the base model for further enhancement

    Model & Enhancement	TP|Ret|Rel	P|R|F
    Combo-6	543|737|814	0.7368|0.6671|0.7002
    Add Shorthand

  • A2.2.2. Not in checkDic, corrected wrong, by preCorrection
    => TBD (this analysis focus on dictionary-based correction first)

• Tokens are in Brat annotation data (spelling error)
  • B1.1. PreCorr (T)
  • B1.2. PreCorr (F)
  • B2.1. DicCorr (T)
  • B2.2. DicCorr (F)
    • B2.2.1. Not detect, real-word (error tag)
    • B2.2.2. Not detect, spelling error (non-word)
    • B2.2.3. Detect, not candidates by edit-distance
    • B2.2.4. Detect, not candidates by suggestion Dic
    • B2.2.5. Detect, not candidates by multi-corrections
    • B2.2.6. Detect, candidates, wrong (not top) rank
    • B2.2.7. Detect, candidates, wrong top rank