Skip to content

Latest commit

 

History

History
26 lines (19 loc) · 902 Bytes

File metadata and controls

26 lines (19 loc) · 902 Bytes

IRIE_project_2018_relation_extraction

Relation Extraction with The Wall Street Journal

Prerequisite

  • Gensim
  • Keras
  • sklearn

workflow

  • Transfer documents into feature vectors by 3-billion words Google News pre-trained word2vec model
  • Train a Bidirectional LSTM following DNN model to predict Node Role/Label
  • Use the output of Bidirectional LSTM layer as input to train a DNN model for edge label prediction

model summary

node role prediction model

alt text

edge relation prediction model

alt text

result

alt text

source: https://arxiv.org/pdf/1809.02700.pdf