/NER_Knowledge

Primary LanguagePython

NER_Knowledge

Environment

python:3.5 tensorflow: 1.11

MRC for NER

use machine reading comprehension (MRC) model to solve NER task. each data is a tuple (question,passage,start_pisition,end_position) In NER, question is the lable definition for each entity type, passage is each sentence, start_position is the start position of each entity and end_position is the end position of each entity.

Single One-pass Model (SOne) for NER

use single one-pass model to solve NER task. Each data ia a tuple (passage, start_position1, end_position1, start_position2, end_position2, ...)

Notice

Because we use the last checkpoint of BERT to predict, so the development set is just to verify the performance of model.

data_mrc

we just set an example for mrc data

data_sone

we just set an example for SOne data for SOne model, the type information is defined in advance. For example, normalize_bert.npy is bert representation of guideline information.

RUN

  1. run MRC-based model
  • bash run_mrc.sh
  1. run SOne-based model
  • bash run_sone.sh

decode

get the answer of submit file format python trans2answer.py