기존 nbt 에서 임베딩을 BERT 추가하여 실험함
An implementation of the Fully Data-Driven version of the Neural Belief Tracking (NBT) model (ACL 2018, Fully Statistical Neural Belief Tracking).
This version of the model uses a learned belief state update in place of the rule-based mechanism used in the original paper. Requests are not a focus of this paper and should be ignored in the output.
The config file in the config directory specifies the model hyperparameters, training details, dataset, ontologies, etc.
train.sh and test.sh can be used to train and test the model (using the default config file). track.sh uses the trained models to 'simulate' a conversation where the developer can enter sequential user turns and observe the change in belief state.
BERT MODEL: BERT-Base, Uncased
Bidirectional Encoder Representations from Transformers
단어/문장 두 레벨로 나누어 테스트
문장 단위 임베딩은 bert-as-service 를 사용하여 진행함
코드에 768 붙어있는 것들이 bert version 코드
현재는 문장단위로 실험했던 코드들이 남아있음
jupyter notebook 으로 embedding visualization 확인 가능