Adaptive ToR: Complexity-Aware Tree-Based Retrieval for Multi-Intent NLU
A recent study published on arXiv introduces the Adaptive Tree-of-Retrieval (Adaptive ToR), an innovative retrieval framework tailored for multi-intent natural language comprehension. This architecture adjusts its retrieval structure based on the specifics of the query, utilizing a Query Tree Classifier that generates a Query Complexity Index from weighted linguistic cues. This enables the system to direct queries through either a fast single-step route or a flexible-depth hierarchical route. Additionally, it features a Tree-Based Retrieval component that systematically breaks down intricate queries into targeted sub-queries, along with an Adaptive Pruning Module that applies a two-stage filtering process. The goal is to establish Pareto-optimal balances between accuracy and computational efficiency, overcoming the drawbacks of uniform single-step retrieval and rigid hierarchical decomposition.
Key facts
- Adaptive ToR is a complexity-aware retrieval architecture for multi-intent NLU.
- It dynamically configures retrieval topology based on query characteristics.
- Includes a Query Tree Classifier computing a Query Complexity Index.
- Routes queries to either a rapid single-step path or an adaptive-depth hierarchical path.
- Tree-Based Retrieval module recursively decomposes complex queries.
- Adaptive Pruning Module employs two-stage filtering.
- Aims for Pareto-optimal trade-offs between accuracy and computational efficiency.
- Addresses limitations of uniform single-step retrieval and fixed-depth hierarchical decomposition.
Entities
Institutions
- arXiv