ANLP_Project

Dataset

• Full model dataset:

Link to Dataset

• Golden BERT dataset - Book8:
  • [Test](wget https://dissent.s3-us-west-2.amazonaws.com/data/discourse_EN_EIGHT_and_but_because_if_when_before_so_though_2017dec18_test.tsv)
  • [Train](wget https://dissent.s3-us-west-2.amazonaws.com/data/discourse_EN_EIGHT_and_but_because_if_when_before_so_though_2017dec18_train.tsv)
  • [Validate](wget https://dissent.s3-us-west-2.amazonaws.com/data/discourse_EN_EIGHT_and_but_because_if_when_before_so_though_2017dec18_valid.tsv)

Directory structure

Instructions for execution

To change:

• The stubbing (input into the 2nd transformer)
• Use the [WP] etc things in the first part of the source for each dataset

Doubts:

• 1st transformer - joining the prompt and outline (or smth like this)
• Loss with the exact connectives or probability distro?
• What do we do for 1 sentence (no punctuation)? Do we return 0 bert loss?
• threshold for labelling a label as unknown?