Skip to content

Named Entity Recognition Model for Naver NLP Challenge 2018 : BiLSTM-CRF model based Korean named entity tagger

Notifications You must be signed in to change notification settings

toriving/naver-nlp-challenge-2018

Repository files navigation

Named Entity Recognition Model for Naver NLP Challenge 2018
Presentation file

NER model for Naver NLP Challenge 2018

Team : State_Of_The_Art (Dongju Park)
1st place on Naver NLP Challenge 2018 NER task

NER model architecture

model

  • The code is implemented based on baseline model
  • Bidirectional LSTM + CRF
  • Embedding layer consists of Word, character(LSTM), and named entity together
  • Shuffle training data on every epoch
  • Total data : 90000, training data : 80000, development data : 10000
  • Important to use RMSPropOptimizer as an optimizer and low value of learning rate
  • The model is an ensemble model that uses hard voting method for N different models

ensemble

Hyper-parameters

Hyper-parameter value
epoch 20
batch_size 128
learning_rate 0.001
keep_prob 0.65
word_embedding_size 128
char_embedding_size 128
tag_embedding_size 128
lstm_units 128
char_lstm_units 128
sentence_length 180
word_length 8
num_ensemble 3
  • Converge between 13 and 15 epochs
  • In this code, the default value of num_ensemble is set to 5, but I set it to 3 when submitting the model

Usage

$ python main.py
$ python main.py \
        --mode <Choice operation mode> \
        --dinput_dir <Input data directory> \
        --output_dir <Output data directory> \
        --necessary_file <necessary_file> \
        --epochs <num_epoch> \
        --batch_size <batch_size> \
        --learning_rate <learning_rate> \
        --keep_prob <dropout_rate> \
        --word_embedding_size <Word, WordPos Embedding Size> \
        --char_embedding_size <Char Embedding Size> \
        --tag_embedding_size <Tag Embedding Size> \
        --lstm_units <Hidden unit size> \
        --char_lstm_units <Hidden unit size for Char rnn> \
        --sentence_length <Maximum words in sentence> \
        --word_length <Maximum chars in word> \
        --num_ensemble <Number of submodels> 

Result

leaderboard

1st place on NER task

Link

Naver NLP Challenge
Changwon University Adaptive Intelligence Research Lab.
NER Leaderboard

About

Named Entity Recognition Model for Naver NLP Challenge 2018 : BiLSTM-CRF model based Korean named entity tagger

Topics

Resources

Stars

Watchers

Forks

Languages