mindojune / kor-depparser

Transition-based Universal Dependency Parser for Korean: in progress

Pipeline

1. preprocessing stage
   1. morphological analysis, lemmatize
   2. tokenizer: which tokenizer?
2. word embedding vector: which embedding?
3. transition classifier - a seq2seq model or a transformer

Oracle for generating training examples

Evaluation Metrics

• UAS (Unlabeled Attachment Score): measure if the tree has a correct head/arc structure, percentage of words that have been assigned the correct head,arc
• LAS (Labeled): also measures the accuracy of arc labels