Regarding the spans in the contrastive loss calculation

[hyleemindslab]

Hello,

In the paper it is stated that

... given a random list of n documents [d1, d2, ..., dn], we extract randomly from each a pair of spans, [s11, s12, ..., sn1, sn2].

I was wondering how the spans were extracted from a document. Are they sentences, each of which is split by nltk.sentence_tokenizer? Or, are they equally sized chunks extracted using a sliding window? Maybe they are the same as the Condenser pretraining blocks but annotated with a document id to which they belong?

Thank you.

[luyug]

I used random non overlapping sequences. Technically what is desired here is a good passage tokenizer; you may get better performance if you can do a better job separating out the passages.

[eugene-yang]

@luyug I'm wondering how long are these spans? From what I understand, you were using $MAX_LENGTH in the scripts for setting the length. Can you share the values you used when training the models?

[luyug]

The length should be selected to align roughly with text lengths in your actual search task's (rounded according to your accelerator's requirement). For passage retrieval, we used 128.