mootezsaad / bugbert Goto Github PK

Go through the DeCLUTR repository and adapt their replication package to our needs.
To reiterate, during our last meeting we agreed not to follow DeCLUTR's input representation that is based on anchor and positive spans (within the same document) and negative spans (within the other documents). We agreed that we will use the vanilla BERT input structure: [CLS] BR_DESC [SEP] DUP_BR_DESC [SEP]. The inclusion of the title would depend on the results of the EDA.
In addition, we agreed that we will pretrain the encoder in a supervised manner since we have the labels. As of writing this issue, we may want to use InfoNCE as a potential loss function.

What fields to return from the index?

What fields will be stored in the index?

Transform MongoDB documents into the JSON Structure, which was discussed yesterday.