H+ Embedding: Harmonizing Global and Token-Level Retrieval with Context-Dependent Phrases
H+ Embedding introduces a unified multi-granularity retriever that bridges the gap between single-vector global retrieval and token-level late interaction by using context-dependent phrase partitions as intermediate retrieval units The model predicts variable-length phrase partitions, preserves uncovered tokens as singletons, and applies importance-guided unit selection with weighted MaxSim interaction Across 16 scientific, medical, and bilingual tasks, the phrase retrieval branch outperforms th
Analysis
TL;DR
- H+ Embedding introduces a unified multi-granularity retriever that bridges the gap between single-vector global retrieval and token-level late interaction by using context-dependent phrase partitions as intermediate retrieval units
- The model predicts variable-length phrase partitions, preserves uncovered tokens as singletons, and applies importance-guided unit selection with weighted MaxSim interaction
- Across 16 scientific, medical, and bilingual tasks, the phrase retrieval branch outperforms the global retrieval branch by 6.91 macro nDCG@10
- H+ Embedding nearly matches token-level performance while using 13.7% fewer document vectors, offering a superior quality-cost tradeoff
- Context-dependent phrase interaction provides a practical intermediate point between global compression and token-level interaction for real-world retrieval systems
Why It Matters
This work directly addresses a critical bottleneck in terminology-intensive retrieval—particularly in medical and scientific domains—where preserving multi-word entities, abbreviations, and compositional concepts is essential. By offering a middle ground between over-compressed global vectors and computationally expensive token-level interaction, H+ Embedding gives practitioners a scalable alternative that maintains high retrieval quality without the storage and scoring overhead of full token-level approaches.
Technical Details
- Multi-granularity architecture: H+ Embedding operates as a unified retriever with both a global vector branch and a phrase-based branch, enabling flexible deployment depending on resource constraints
- Context-dependent phrase partitioning: The model predicts variable-length phrase boundaries from the input context rather than relying on fixed or content-independent grouping rules, allowing it to capture domain-specific multi-word entities like medical terminology and abbreviations
- Importance-guided unit selection with weighted MaxSim: After phrase partitioning, uncovered tokens are preserved as singletons, and the system applies importance-weighted MaxSim interaction to score relevance between query and document units
- Benchmark evaluation: Tested across 16 scientific, medical, and bilingual retrieval tasks, demonstrating a 6.91 macro nDCG@10 improvement over global retrieval and near-parity with token-level retrieval at 13.7% fewer document vectors
Industry Insight
- Retrieval systems in regulated or terminology-heavy domains (healthcare, legal, scientific literature) should evaluate multi-granularity approaches like H+ Embedding to balance accuracy and infrastructure costs, especially where token-level late interaction is prohibitively expensive
- The context-dependent phrase partitioning strategy could be adapted beyond retrieval to other NLP tasks requiring preservation of compositional entities, such as question answering over specialized corpora
- As vector budgets remain a practical constraint for large-scale deployment, the 13.7% vector reduction while maintaining near-token-level performance makes this approach attractive for production systems seeking to optimize the quality-cost frontier
Disclaimer: The above content is generated by AI and is for reference only.