Naver NLP Challenge 2018

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

• 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

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

1st place on NER task

• Named Entity Recognition(NER) Leaderboard

Link

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