Assignment 4

Competition Deadline (CodaLab): 12/15/2016 09:00
Code Deadline (GitHub): 12/15/2016 21:00:00

CSIE 5431 - Applied Deep Learning, instructed by Yun-Nung (Vivian) Chen

Assignment Release

12/01/2016

  • implement seq2seq model
  • Acheiving baseline will get 6% and 4 % each (Translation & Natural Language Generation)
  • 3% will be graded via report
  • Achieving strong baseline will get 1% each

Deadline

12/15/2016 9:00am

  • Competition end
  • Start to count free day quota

Assignment 4 (15%)

Free HTML5 Template by FREEHTML5.co

Assignment 4 has two main part: Translation & Natural Language Generation. Details are shown below.

Assignment 4-1 Translation (7%)

Assignment 4-1 is to implement machine translation using seq2seq model.
Machine Translation can be achieved using seq2seq model, which contains two RNNs, encoder RNN and decoder RNN. It can learn to form a fixed sized vector that contains the information of the source English sentence, and then decode the vector into another language (Spanish).

Github :

Write a bash script named 'run_translation.sh'. TA will run your code by using command :
' bash run_translation.sh [input_file] [answer_file] '
Make sure your code can be executed by this bash script with correct arguments setting.

Dataset :

English to Spanish dataset, in here.(containing testing data)

Evaluation :

TA will evaluate by BLEU score.
The submission must follow the pattern below: Free HTML5 Template by FREEHTML5.co

Assignment 4-1 Grading Policy :

- achieve weak baseline 0.22 get 6%
- achieve strong baseline 0.27 get 1% more

Assignment 4-2 Natural Language Generation (5%)

Assignment 4-2 is to implement Natural Language Generation.
Natural Language Generation needs first to convert the action-slot vector into a fixed-sized vector and then, like machine translation, using decoder RNN to generate a target sentence.

Github :

Write a bash script named 'run_generation.sh'. TA will run your code by using command :
' bash run_generation.sh [input_file] [answer_file] '
Make sure your code can be executed by this bash script with correct arguments setting.

Dataset :

Parsed DSTC2 dataset, in here.(containing testing data)

Evaluation :

TA will evaluate by BLEU score.
The submission must follow the pattern below: Free HTML5 Template by FREEHTML5.co

Assignment 4-2 Grading Policy :

- achieve weak baseline 0.32 get 4%
- achieve strong baseline 0.60 get 1% more

Report (3%)

- Describe your model in detail.
- Describe what you learned and how you improve the performance.
- Explain the main function of your code.

Submission

The result files must be 'answer.txt'.
Package in a zip file and upload.


FAQ

Any frequent question will be shown here. Post your question on FB Group or contact with TA via e-mail: adl2016ta@gmail.com

Assignment Submission

GitHub

Open-source your code.

CodaLab

Language Generation

A: Deadline 設定為12/15上課前,在此之前Assignment皆能夠按照規則繳交。

Teaching Assistants

Free HTML5 Template by FREEHTML5.co

莊舜博

General questions; GitHub submission.

Free HTML5 Template by FREEHTML5.co

李致緯

General questions; CodaLab submission.