bert as language model

？演示|案例-en | case-zh |

伯特作为语言模型

对于一句话，我们有

在传统语言模型中，例如rnn，

在双向语言模型中，它具有更大的上下文。

在此实现中，我们只采用以下近似值，

。

演示

在网络演示中尝试

测试案例

更多案例：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/

对于以下测试案例

$ 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.json

输出：

 # 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
  }
]