Question on Castorini UniCOiL model in HF (not sure where to ask.)

[PrithivirajDamodaran]

Hello Team,

I am looking at UniCOiL paper and the models in HuggingFace. I understand vanilla UniCOiL does just term weighting without any expansions and usually expansion models are added post-hoc. But there are castorini/unicoil-msmarco-passage and castorini/unicoil-noexp-msmarco-passage uploaded. Does that mean the former model has some form of expansion baked into it ? I could not figure out the difference between these two from seeing the config files. Please advice.

Also I didn't know what is the right place to ask this, hence here.

Thanks

[MXueguang]

Hi @PrithivirajDamodaran,
You are correct. castorini/unicoil-msmarco-passage takes text input together with document expansions (predicted queries).
castorini/unicoil-noexp-msmarco-passage is the vanilla uniCOIL without any expansion.

[lintool]

To be clear, with castorini/unicoil-msmarco-passage you have to expand the corpus first separately.

[PrithivirajDamodaran]

Thanks for the responses:

1.) So in the paper 2g denotes castorini/unicoil-noexp-msmarco-passage - am i right ? If so which entry denotes castorini/unicoil-msmarco-passage ?

To be clear, with castorini/unicoil-msmarco-passage you have to expand the corpus first separately.

2.) Do you mean you used castorini/unicoil-msmarco-passage for document / passage expansion for indexing and castorini/unicoil-noexp-msmarco-passage for query expansion at query time ? Sorry, I am not fully clear here.

[MXueguang]

1. Yes, 2g denotes castorini/unicoil-noexp-msmarco-passage and 2h denotes castorini/unicoil-msmarco-passage

2. Both castorini/unicoil-noexp-msmarco-passage and castorini/unicoil-msmarco-passage are encoding model, that assign token weights to each input tokens. They are not doing any expansion.
   To get predicted query for expanding document, need to use the query2doc-T5 model to generate. We have generated queries stored here https://huggingface.co/datasets/castorini/msmarco_v1_passage_doc2query-t5_expansions.

To be clarify, an end2end uniCOIL(with expansion) pipeline is:

1. Augment the msmarco corpus with predicted queries.

• The predicted queries are generated by query2doc-T5.

2. Concatenate the original passage and the tokens in predicted queries as augmented passage.
3. Encode augmented passage with castorini/unicoil-msmarco-passage to get sparse vector embedding
4. Index passage sparse vectors with pyserini Impact Indexer.
5. Encode queries with castorini/unicoil-msmarco-passage
6. Conduct search over the index.

On the other hand, an end2end uniCOIL(without expansion) pipeline is:

1. Encode msmarco passage with castorini/unicoil-noexp-msmarco-passage to get sparse vector embedding
2. Index passage sparse vectors with pyserini Impact Indexer.
3. Encode queries with castorini/unicoil-noexp-msmarco-passage
4. Conduct search over the index.

[PrithivirajDamodaran]

@MXueguang Thank you for the detailed response, I am independently reproducing "Improved Learned Sparse Retrieval with
Corpus-Specific Vocabularies" which is improving on uniCOIL. So I am trying to be informed before I start.