This is an attempt to reproduce the Don't Parse, Generate! A Sequence to Sequence Architecture for Task-Oriented Semantic Parsing

You need to download data from http://fb.me/semanticparsingdialog and unzip the data, your file structure should like the following

- top-dataset-semantic-parsing
    - test.tsv
    - train.tsv
    - eval.tsv
    ...
- TOP
    train.py
    - data
        ...
    ...

To run the code, you need to install the lastest fastNLP by the following command

pip install git+https://github.com/fastnlp/fastNLP

Then you can run the code as

python train.py

The implementation is different from the paper in two folds:
(1) We don't generate(copy) the token from the source sentence, since in this task, the copy occurs sequentially. Therefore, in our code, during generation, the model only needs to judge whether to copy at this position. An index indicator is kept to record which position in the source sentence has been copied (CopyTransformerState.copy_src_len is for this purpose).
(2) Instead of keeping an embedding for the word copied from the source sentence, we use the embedding from the source sentence to represent the copied tokens in the decoder sequence (the CopyTransformerSeq2SeqDecoder.get_real_embed() is used to assign corresponding embedding for different tokens.).

The performance in the TOP datasets are