Kor-Sentence-Similarity

Sentence/Text Similarity for Korean (In Simple Way)

Models

• Modification of Siamese recurrent architectures for learning sentence similarity
• Simple ver.
• Char-CNN & MLP for Siamese Networks

Details

• Data

  • In data, two questions are seperated by '\t'

• Preprocessing

  • Character Level (음소 or 음절)
  • Digits and Specials
  • For eumjeol(Syllable), use frequent 2350

• Configuration

  • main.py : main run file
  • --epochs : # of training epochs
  • --batch : Batch Size
  • --lr : Learning rate
  • --strmaxlen : Maximum Limit of String Length
  • --charsize : Vocab Size
  • filter_num : # of Filter of one CNN Filter
  • --emb : Embedding Dimension
  • --eumjeol : Use Eumjeol(Syllable-level) if specified
  • threshold : Threshold to determine Similar or not
  • --model : Model Selection (CNN, MLP)

To Run

• Set FC,layer and CNN layers in 'main.py'
• run 'main.py' with arguments as you wish