[TOC]

BB-KBQA

File change

file	note
ner/	add BERT-CRF model
run_classifier.py	add NlpccpmProcessor, NlpccpmwmProcessor and NlpccxqProcessor
chinese_L-12_H-768_A-12/vocab.txt	[unused1] 替换为entity
data_utils.py	数据预处理
evaluate.py	结果处理

TODO: (huangziqi) 加入计算模型结果的代码等其他部分代码

Model

BERT

BERT-Base, Chinese（chinese_L-12_H-768_A-12）:Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters

BERT-Softmax

BERT-CRF

Dataset

data	train	dev	test	infer
mention detection	13609	1000	9870	9870
entity disambiguation	67246	5033	36219	34764
predicate mapping	175121	10601	121960	342208
predicate mapping(with mention)	175121	10601	121960	342208

dataset	train	dev	test	infer
md	MD/train.txt	MT/dev.txt	MT/test9870.txt
ed(raw)	XQ/candidate_entity_train.txt	XQ/candidate_entity_dev.txt	XQ/candidate_entity_test_bert_9870_mc.txt	XQ/candidate_entity_test_bert_9870.txt
ed	XQ/train.csv	XQ/dev.csv	XQ/test9870.csv	XQ/infer9870.csv
pm(raw)	PM/train.txt	PM/dev.txt	PM/candidate_predicate_bert_9870_ec.txt	PM/candidate_predicate_bert_9870_all.txt
pm	PM/train.csv	PM/dev.csv	PM/test9870.csv	PM/infer9870_all.csv
pm(with mention)	PM/train_with_mention.csv	PM/dev_with_mention.csv	PM/test9870_with_mention.csv	PM/infer9870_all_with_mention.csv

Modules

mention detection

ner/ : BERT-CRF model

运行下面的命令进行训练或者测试：run_md.sh

#!/usr/bin/env bash
export BERT_BASE_DIR=./chinese_L-12_H-768_A-12
export MY_DATASET=./data/MD
python run_md.py \
  -do_train \
  -do_eval \
  -do_predict \
  -data_dir $MY_DATASET \
  -vocab_file $BERT_BASE_DIR/vocab.txt \
  -bert_config_file $BERT_BASE_DIR/bert_config.json \
  -init_checkpoint $BERT_BASE_DIR/bert_model.ckpt \
  -max_seq_length 60 \
  -batch_size 32 \
  -learning_rate 1e-5 \
  -num_train_epochs 30 \
  -output_dir ./md_output/ \
  -device_map 1

• train. ./run_md.sh train
• predict. ./run_md.sh predict
• evaluate. ./run_md.sh eval

entity disambiguation

run_classifier.py: 添加NlpccxqProcessor

运行下面的命令进行训练或者测试：run_xq.sh

#!/usr/bin/env bash
export BERT_BASE_DIR=./chinese_L-12_H-768_A-12
export MY_DATASET=./data/XQ
python run_classifier.py \
  --task_name=nlpccxq \
  --do_train=false \
  --do_eval=false \
  --do_predict=true \
  --data_dir=$MY_DATASET \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --max_seq_length=60 \
  --train_batch_size=32 \
  --learning_rate=5e-5 \
  --num_train_epochs=3.0 \
  --output_dir=./xq_output/

predicate mapping

run_classifier.py: 添加NlpccpmProcessor, NlpccpmwmProcessor

运行下面的命令进行训练或者测试：run_pm.sh, run_pm_with_mention.sh

#!/usr/bin/env bash
export BERT_BASE_DIR=./chinese_L-12_H-768_A-12
export MY_DATASET=./data/PM
python run_classifier.py \
  --task_name=nlpccpm \
  --do_train=false \
  --do_eval=false \
  --do_predict=true \
  --data_dir=$MY_DATASET \
  --vocab_file=$BERT_BASE_DIR/vocab.txt \
  --bert_config_file=$BERT_BASE_DIR/bert_config.json \
  --init_checkpoint=$BERT_BASE_DIR/bert_model.ckpt \
  --max_seq_length=40 \
  --train_batch_size=32 \
  --learning_rate=5e-5 \
  --num_train_epochs=3.0 \
  --output_dir=./pm_output/

Results

NLPCC-ICCPOL 2016 KBQA results (%)

Models	Averaged F1
Baseline System (C-DSSM)	52.47
Wang et al. [15]	79.14
Xie et al. [16]	79.57
K et al. [17]	80.97
Zhou et al. [14]	81.06
Yang et al. [13]	81.59
Xie et al. [11]	82.43
Lai et al. [12]	82.47
BB-KBQA	84.12

Mention detection results (%)

Models	F1
BiLSTM-CRF	90.28
BERT-Softmax	96.61
BERT-BiLSTM-CRF	96.90
BERT-CRF	97.34

Entity disambiguation results (%)

Models	Accuracy@1	Accuracy@2	Accuracy@3
BiLSTM-DSSM [19]	85.89	88.50	90.81
Siamese BiLSTM [29]	87.85	92.58	94.59
Siamese CNN [30,31]	88.04	92.68	94.88
BERT-Softmax	89.14	93.19	95.05

Predicate mapping results (%)

Models	Accuracy@1	Accuracy@2	Accuracy@3
Siamese BiLSTM	92.54	96.74	98.12
Siamese BiLSTM(2)	93.74	97.46	98.38
Siamese CNN	86.47	93.80	96.16
Siamese CNN(2)	90.61	95.57	97.01
BERT-Softmax	94.81	97.68	98.60
BERT-Softmax(2)	94.66	97.63	98.41

Setup

1. cd source root.
2. Install python library. pip3 install -r requirements.md
3. Download bert model from BERT-Base, Chinese（chinese_L-12_H-768_A-12）:Chinese Simplified and Traditional, 12-layer, 768-hidden, 12-heads, 110M parameters. Place it to chinese_L-12_H-768_A-12/

Knowledge dataset

You can download it from kaggle, but I have not know how to use it.

Reference

@article{devlin2018bert,
  title={BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding},
  author={Devlin, Jacob and Chang, Ming-Wei and Lee, Kenton and Toutanova, Kristina},
  journal={arXiv preprint arXiv:1810.04805},
  year={2018}
}

• https://github.com/google-research/bert
• https://github.com/macanv/BERT-BiLSTM-CRF-NER
• https://zhuanlan.zhihu.com/p/53796189
• https://www.kaggle.com/kerneler/starter-nlpcc-kbqa-7dd5b660-f?select=nlpcc2018.kbqu.train

Contact information

For help or issues using BB-KBQA, please submit a GitHub issue or email (164504252@qq.com) or weixin (yyaadet2002).