ner-movie

Named Entity Recognition model on movie domain

Output

'2017년에 개봉한 미국 스릴러 영화 알려줘' =>

{'actor': [], 'country': ['미국'], 'director': [], 'genre': ['스릴러'], 
                      'month': [], 'movie': [], 'year': ['2017년']}

Motivation for making this model

C'est La Vie team of AI-LAB KOREA(한국인공지능연구소) worked on a chatbot project in 2018.
This model was made by me and it was used for the chatbot project.

How to train the ner-movie model

python main.py

Model Structure

- use five word embeddings : 
  (1) pretrained wiki-corpus word2vec embedding
  (2) pretrained movie domain word2vec embedding
  (3) POS(Part of Speech) embedding
  (4) syllable(char) embedding
  (5) entity dictionary embedding 
- bi-directional lstm 
- softmax 

Data on "data" folder

- intent_v3.txt: input data(anticipated user's questions about movies on a chatbot system)

Other Things to Know

- labels: in order to save time, label datas are not manually made. 
          I made codes to generate labels using entity dictionaries. 

Reference

- 송치윤 외, 「개선된 워드 임베딩 모델과 사전을 이용한 Bidrectional LSTM CRF 기반의 한국어 개체명 」, 한국소프트웨어종합학술대회 논문집, 2017
  => I obtained an idea of concatenating five different types of word embeddings