Skip to content

Latest commit

 

History

History
73 lines (56 loc) · 2.66 KB

README.md

File metadata and controls

73 lines (56 loc) · 2.66 KB
Raw

Generative IR Survey Paper Structure

Introduction

  • Briefly introduce the history and trend in IR
  • Traditional DPR (Retrieve then Rank)
image
  • propose an alternative IR architecture called differentiable search index (DSI), direct seq2seq map query to document ID
image

image

Definition of GenIR (DSI)

image

Intro to DSI

Two Major Components

  • Indexing
  • Retrieval

Pros and Cons of DSI

  • Pros
    • End-to-end training
  • Cons
    • Not easy to scale DSI systems to handle large data volumes

Identifier Strategies in Generative IR

Document Identifiers

  • Transformer Memory as a Differentiable Search Index
  • DSI++: Updating Transformer Memory with New Documents
  • Learning to Tokenize for Generative Retrieval

String Identifiers

  • Autoregressive Search Engines: Generating Substrings as Document Identifiers
  • Semantic-Enhanced Differentiable Search Index Inspired by Learning Strategies
  • Learning to Rank in Generative Retrieval
  • TOME: A Two-stage Approach for Model-based Retrieval
  • Multiview Identifiers Enhanced Generative Retrieval

Enhancing Query Generation and Expansion

  • generating more training data
  • document representation as queries

Document Representation in DSI

  • DSI
    • Direct Indexing
    • Set Indexing
    • Inverted Index
  • Bridging the Gap
    • queries as representation
    • Bridging the Gap Between Indexing and Retrieval for Differentiable Search Index with Query Generation
    • Multiview Identifiers Enhanced Generative Retrieval

Add New Documents

  • DSI++: Updating Transformer Memory with New Documents
  • Continual Learning for Generative Retrieval over Dynamic Corpora

Evaluation of Generative IR

Compare the performance of traditional IR systems with DSI-based IR systems

  • Datasets
    • IR tasks
    • knowledge intensive tasks (eg. QA)
  • Metrics

Limitations and Future Directions

  • Challenges with DSI: Address potential issues
    • generating non-existent document IDs (FM Index)
    • scaling DSI systems to handle large data volumes.

This structure aims to provide a comprehensive overview of Generative IR and the pivotal role of Differentiable Search Indexes, making it accessible to newcomers while detailing the progress and challenges in the field.