bert as language model
1.0.0
? 데모 | 케이스-엔 | Cases-ZH |
문장을 위해, 우리는 가지고 있습니다
RNN과 같은 전통적인 언어 모델에서
양방향 언어 모델에서는 더 큰 맥락이 있습니다.
이 구현에서는 다음과 같은 근사치를 채택합니다.
.
웹 데모를 사용해보십시오
더 많은 사례 : 中文
export BERT_BASE_DIR=model/uncased_L-12_H-768_A-12
export INPUT_FILE=data/lm/test.en.tsv
python run_lm_predict.py
--input_file= $INPUT_FILE
--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=128
--output_dir=/tmp/lm_output/다음 테스트 사례
$ cat data/lm/test.en.tsv
there is a book on the desk
there is a plane on the desk
there is a book in the desk
$ cat /tmp/lm/output/test_result.json산출:
# prob: probability
# ppl: perplexity
[
{
" tokens " : [
{
" token " : " there " ,
" prob " : 0.9988962411880493
},
{
" token " : " is " ,
" prob " : 0.013578361831605434
},
{
" token " : " a " ,
" prob " : 0.9420605897903442
},
{
" token " : " book " ,
" prob " : 0.07452250272035599
},
{
" token " : " on " ,
" prob " : 0.9607976675033569
},
{
" token " : " the " ,
" prob " : 0.4983428418636322
},
{
" token " : " desk " ,
" prob " : 4.040586190967588e-06
}
],
" ppl " : 17.69329728285426
},
{
" tokens " : [
{
" token " : " there " ,
" prob " : 0.996775209903717
},
{
" token " : " is " ,
" prob " : 0.03194097802042961
},
{
" token " : " a " ,
" prob " : 0.8877727389335632
},
{
" token " : " plane " ,
" prob " : 3.4907534427475184e-05 # low probability
},
{
" token " : " on " ,
" prob " : 0.1902322769165039
},
{
" token " : " the " ,
" prob " : 0.5981084704399109
},
{
" token " : " desk " ,
" prob " : 3.3164762953674654e-06
}
],
" ppl " : 59.646456254851806
},
{
" tokens " : [
{
" token " : " there " ,
" prob " : 0.9969795942306519
},
{
" token " : " is " ,
" prob " : 0.03379646688699722
},
{
" token " : " a " ,
" prob " : 0.9095568060874939
},
{
" token " : " book " ,
" prob " : 0.013939591124653816
},
{
" token " : " in " ,
" prob " : 0.000823647016659379 # low probability
},
{
" token " : " the " ,
" prob " : 0.5844194293022156
},
{
" token " : " desk " ,
" prob " : 3.3361218356731115e-06
}
],
" ppl " : 54.65941516205144
}
]