Dense Retrieval Method

A Dense Retrieval Method is an information retrieval method that uses dense vector representations to match queries with documents through semantic similarity computation.

• AKA: Dense Passage Retrieval Method, Neural Retrieval Method, Embedding-Based Retrieval Method.
• Context:
  • It can typically encode Text Sequences with neural encoder models.
  • It can typically compute Cosine Similarity between query vectors and document vectors.
  • It can typically perform Approximate Nearest Neighbor Search for efficient retrieval.
  • It can often utilize Bi-Encoder Architectures for independent encoding.
  • It can often employ Contrastive Training for representation learning.
  • It can often apply Hard Negative Mining for discrimination improvement.
  • It can often integrate Vector Databases for scalable indexing.
  • It can range from being a Single-Vector Dense Retrieval Method to being a Multi-Vector Dense Retrieval Method, depending on its representation granularity.
  • It can range from being a Symmetric Dense Retrieval Method to being a Asymmetric Dense Retrieval Method, depending on its encoder configuration.
  • It can range from being a In-Domain Dense Retrieval Method to being a Zero-Shot Dense Retrieval Method, depending on its training coverage.
  • It can range from being a Monolingual Dense Retrieval Method to being a Cross-Lingual Dense Retrieval Method, depending on its language capability.
  • ...
• Examples:
  • Foundation Dense Retrieval Methods, such as:
    • Dense Passage Retrieval (DPR), ANCE Method, ColBERT Method.
  • Domain-Specific Dense Retrieval Methods, such as:
    • Legal Dense Retrieval Method, Medical Dense Retrieval Method, Scientific Dense Retrieval Method.
  • Architecture-Based Dense Retrieval Methods, such as:
    • Bi-Encoder Model, Poly-Encoder Method, Late Interaction Method.
  • ...
• Counter-Examples:
  • Sparse Retrieval Method, which uses term-based rather than dense representations.
  • Lexical Retrieval Method, which matches keywords rather than semantics.
  • Boolean Retrieval Method, which uses exact matching rather than similarity.
• See: Bi-Encoder Model, Information Retrieval Method, Vector Database, Contrastive Learning Technique for Legal Text, Neural Information Retrieval, Semantic Search, Embedding Model.

---