Skip to content

Latest commit

 

History

History
127 lines (109 loc) · 4.51 KB

README.md

File metadata and controls

127 lines (109 loc) · 4.51 KB

SKE_BERT_RE

BERT based solution for IE problem in 2019 Language and Intelligence Challenge.

Requirements

  • Python 3.7
  • pytorch 1.10
  • OpenKE

Data

Download from 2019 Language and Intelligence Challenge Put data resources into raw_data/chinese

Solution

two steps solution

Firstly do the multi-label relation extractionpso_1, secondly do the entity extractionpso_2 with Bert structure.

pso_1 pso_2
image image

multi head selection solution

Replace the lstm structure with Bert for the paper Joint entity recognition and relation extraction as a multi-head selection problem.

Run

preprocess

python main.py --mode preprocessing --exp_name chinese_bert_re
python main.py --mode preprocessing --exp_name chinese_bert_pso --type pso_1

train the model

python main.py --mode train --exp_name chinese_bert_re
python main.py --mode train --exp_name chinese_bert_pso --type pso_1
python main.py --mode train --exp_name chinese_bert_pso --type pso_2

reload the model from some epoch

python main.py --mode reload --exp_name chinese_bert_re --epoch $epoch
python main.py --mode reload --exp_name chinese_bert_pso --type pso_1 --epoch $epoch
python main.py --mode reload --exp_name chinese_bert_pso --type pso_2 --epoch $epoch

evaluate the model - NER and Triplets F1 score

python main.py --mode evaluation --exp_name chinese_bert_re
python main.py --mode evaluation --exp_name chinese_bert_pso --type pso_1
python main.py --mode evaluation --exp_name chinese_bert_pso --type pso_2

postcheck

  • build triple knowledge graph
  • distant supervision to enrich triples
  • triple classification with XGBOOST
Features Descriptions Resources
score the triple confidence score for two steps / multi head selections
rank the relative rank for triple confidence score in one sample candidate triples
transe the transe score for one triple (s,p,o) Triple Trustworthiness Measurement for Knowledge Graph
sdvalidate the sdvalidate value for one triple (s,p,o) Improving the Quality of Linked Data Using Statistical Distributions
one hot label predicates for one sample candidate triples
seg whether the subject/object boundaries consistent with word segment

distant supervision

cancel comment on self.spo_search(text,result)

triple classification

  • prepare positive negative samples
  • prepare TransE score

prepare positive negative samples for Training data

cancel comment on tester.run() , get pos_neg.json output

python lib/kg/negative_sample.py

Modify tag xgb_train_root in experiments to according run .json, like pos_neg_score.json, Then rerun the model to get the model confidence score for pos/neg data and get the ouput pos_neg_score.json.

# multi-head selection
python main.py --mode postcheck --exp_name chinese_bert_re
# pso two steps model
python main.py --mode postcheck --exp_name chinese_bert_pso --type pso_2

prepare for dev data

For dev data, modify tag xgb_train_root to dev.json.
If use multi-head model, just run

python main.py --mode postcheck --exp_name chinese_bert_re

If pso two steps model

python main.py --mode postcheck --exp_name chinese_bert_re

then comment on tester.spo_search_res(), scp error/chinese_bert_pso/dev.json data/chinese_bert/pso/dev.json,then rerun

python main.py --mode postcheck --exp_name chinese_bert_re

prepare TransE score

prepare train/dev data proper format for OpenKE, get the data in data/transe fold

python prepare_transe.py --mode train/dev

place the entity2id.txt,relation2id.txt,train2id.txt to base_path in OpenKE packages and run

python lib/transe/run_transe.py --mode train/dev

get the embedding of (s,p,o) triple, scp the output(*.pickle) to the path lib/transe

python lib/transe/get_embedding.py

XGBOOST Model

Firstly, for the training data, then for the dev data.

python kg/post_check.py

Others

Model Ensemble

Bert attention visualization

see in lib/metrics/attn_vis.py