bridge entity

**Bridge Entity** is the **intermediate entity in multi-hop reasoning that connects the question's subject to the answer through an inferential chain — the implicit or explicit entity discovered during intermediate reasoning steps that bridges the gap between what is asked and what must be found** — the key concept in compositional question answering that determines whether a model can perform genuine multi-step reasoning or merely pattern-match to superficially similar single-hop questions. **What Is a Bridge Entity?** - **Definition**: In a multi-hop question requiring N reasoning steps, bridge entities are the intermediate entities discovered at each step that connect the starting entity in the question to the final answer entity — the "stepping stones" of the inference chain. - **Example**: "What language is spoken in the country where Einstein was born?" — Bridge entity: "Germany" (connects Einstein → country of birth → Germany → language → German). The question asks about a language; "Germany" is never mentioned but must be inferred. - **Implicit vs. Explicit**: Bridge entities may be explicitly mentioned in the question ("Einstein's birthplace") or entirely implicit (requiring world knowledge to identify the connecting entity). - **Chain Structure**: For N-hop questions, there are N−1 bridge entities forming a chain: Subject → Bridge₁ → Bridge₂ → ... → Answer. **Why Bridge Entities Matter** - **Multi-Hop Reasoning Validation**: If a model can identify the correct bridge entity, it demonstrates genuine multi-step reasoning rather than shortcut exploitation (e.g., guessing the answer from surface-level patterns). - **Interpretable Reasoning**: Explicit bridge entity identification creates an auditable reasoning chain — each step can be independently verified for correctness. - **Error Diagnosis**: When multi-hop QA fails, identifying which bridge entity was wrong pinpoints the exact reasoning step that broke — enabling targeted model improvement. - **Retrieval Guidance**: Knowing the bridge entity guides retrieval — the system can retrieve documents about "Germany" specifically rather than hoping a single retrieval captures the full reasoning chain. - **Question Decomposition**: Bridge entities correspond to the answer of sub-questions — "Where was Einstein born?" → "Germany" (bridge) → "What language is spoken in Germany?" → "German" (answer). **Bridge Entity in Multi-Hop QA** **HotpotQA Bridge Questions**: - Account for ~70% of multi-hop questions in HotpotQA. - Require identifying a bridge entity that connects two Wikipedia paragraphs. - Example: Para 1 about Person X → Bridge entity "City Y" → Para 2 about City Y → Answer. **2WikiMultiHopQA**: - Explicitly annotated bridge entities and comparison entities. - Enables evaluation of whether models find correct intermediate reasoning steps. - Question types: bridge, comparison, and inference — each requiring different intermediate entities. **Bridge Entity Detection Methods** **Entity Linking + Relation Extraction**: - Parse the question to identify all entities. - Use knowledge graphs to find entities that connect question entities to potential answers. - Select bridge entities based on relational path analysis. **Decomposition-Based**: - Decompose the multi-hop question into single-hop sub-questions. - Answer sub-questions sequentially — each intermediate answer is a bridge entity. - Tools: Least-to-Most prompting, DecompRC, question decomposition networks. **Retrieval-Guided**: - First retrieval step finds documents about the question's main entity. - Extract candidate bridge entities from retrieved documents. - Second retrieval step uses bridge entity to find documents containing the answer. **Bridge Entity Complexity** | Hop Count | Bridge Entities | Example | Difficulty | |-----------|----------------|---------|------------| | **2-hop** | 1 bridge | Person → Country → Language | Medium | | **3-hop** | 2 bridges | Ingredient → Dish → Country → Capital | Hard | | **4-hop** | 3 bridges | Author → Book → Film → Director → Birthplace | Very Hard | | **Comparison** | 0 bridges (parallel) | "Who is older, A or B?" | Different pattern | Bridge Entity is **the atomic unit of multi-hop reasoning** — the intermediate discovery that proves a model is genuinely chaining inferences rather than shortcutting to the answer, serving as both the mechanistic explanation of how multi-step reasoning works and the diagnostic tool for understanding when and why it fails.