KAMR: Grounding Generation via Knowledge-Aligned Multi-hop Retrieval

2026-07-29

AI Summary

This work addresses limitations in existing graph-based multi-hop retrieval methods, which independently rank triples without considering structural alignment between the query and the knowledge graph and suffer from coarse-grained supervision, often missing critical yet semantically weakly related facts. To overcome these issues, the authors propose the Knowledge-Aligned Multi-hop Retriever (KAMR), which explicitly distinguishes strongly constrained anchor triples from weakly aligned associative triples. KAMR constructs partially aligned training data without manual annotation by masking triple elements and leveraging large language models to generate queries. It further mitigates supervision sparsity through a dual-level contrastive learning framework operating at both element and pair levels. Evaluated across four multi-hop question answering benchmarks, three large language model backbones, and fourteen strong baselines, KAMR demonstrates substantial improvements in retrieval quality and downstream QA performance.

Abstract

Graph-based retrieval-augmented generation increasingly relies on multi-hop retrieval, where answering a query requires composing multiple connected knowledge-graph triplets. However, existing retrievers often rank triplets independently via global semantic matching. Moreover, many multi-hop benchmarks provide only final answers, which limits supervision for query--triplet alignment and causes structurally necessary but weakly aligned facts to be missed. To address these issues, we propose a knowledge-aligned multi-hop retriever, KAMR, which distinguishes anchor triplets that are strongly constrained by the query from connected triplets that are weakly aligned yet structurally linked to the anchors. To mitigate the lack of query-triplet alignment supervision, we build a partial alignment dataset by masking triplet elements and prompting an LLM to generate corresponding queries, and optimize two contrastive objectives for pair-level and element-level matching. At inference time, KAMR retrieves anchors globally and then expands locally to collect connected evidence. Across four benchmarks, three LLM backbones, and fourteen baselines, KAMR consistently improves multi-hop retrieval and downstream question answering performance.

Problem

• multi-hop retrieval
• retrieval-augmented generation
• knowledge graph
• query-triplet alignment
• weakly aligned facts

Innovation

• multi-hop retrieval
• knowledge graph
• retrieval-augmented generation
• contrastive learning
• query-triplet alignment