PolyDPR

Introduction

• What is the name of the poly-DPR paper?

  Improving Biomedical Information Retrieval with Neural Retrievers

• What are the main contributions of the poly -DPR paper?

  • each context is represented by k vectors
    • still employ MIPS
  • TempQG: template based question generation method
    • can generate a large number of in domain questions
  • 2 pretraining tasks: ETM and RSM

Method

• Poly DPR uses colBERT style maxsim

• TempQG uses a seq2seq model with a passage and template as input

  

• TempQG templates are generated by masking low frequency tagged biological entities
  • also do DPR with text input to select

  

• PolyDPR pretraining task: Extended Title Mapping retreives an abstract based on title + concatenated top tfidf words

  

Results

• Poly-DPR results: best performance with most granular sub document representations
  • obvious tradeoff with time

  

Conclusions

• good experimental evidence about using subdocument representations vs encoding entire document
• Compare against T5query uses in GPL

Reference

@article{DBLP:journals/corr/abs-2201-07745,
  author    = {Man Luo and
               Arindam Mitra and
               Tejas Gokhale and
               Chitta Baral},
  title     = {Improving Biomedical Information Retrieval with Neural Retrievers},
  journal   = {CoRR},
  volume    = {abs/2201.07745},
  year      = {2022},
  url       = {https://arxiv.org/abs/2201.07745},
  eprinttype = {arXiv},
  eprint    = {2201.07745},
  timestamp = {Fri, 21 Jan 2022 13:57:15 +0100},
  biburl    = {https://dblp.org/rec/journals/corr/abs-2201-07745.bib},
  bibsource = {dblp computer science bibliography, https://dblp.org}
}