1. I dont't have the PRE-TRAINED PARAMETER, so in paddle_train.py pretrained embedding is not loaded. Static parameter must be intialized by pre-training.

   • load_pretrained_parameters in this line should be implemented first, otherwise the embedding should be learned from data.
   • please check this: How to load pre-trained parameters that are trained by Paddle.

2. the lastes develop branch of Paddle is required (I fixed some bug.)

3. this fix is required to run paddle_infer.py: https://github.com/PaddlePaddle/Paddle/compare/develop...lcy-seso:fix_infer_dim （I will merge it soon after more test.）

---

• Below is example of inferring outputs:
  • one line is a word in document.
  • one line has four columns separated by TAB
    • 1. word
    • 2. sentence predictions: groudtruth [predicetd probability]
    • 3. start predictions: groudtruth [predicetd probability]
    • 4. end predictions: groudtruth [predicetd probability]

Super   0[0.9902 0.0098]        0[0.9879 0.0121]        0[0.9866 0.0134]
Bowl    1[0.9961 0.0039]        0[0.9954 0.0046]        0[0.9951 0.0049]
50      0[0.9968 0.0032]        0[0.9962 0.0038]        0[0.9960 0.0040]
was     0[0.9969 0.0031]        0[0.9963 0.0037]        0[0.9961 0.0039]
an      0[0.9969 0.0031]        0[0.9963 0.0037]        0[0.9961 0.0039]
...