[Paper Review] UHop: An Unrestricted-Hop Relation Extraction Framework for Knowledge-Based Question Answering

Ziyuan Chen, Chih-Hung Chang+3 more|arXiv (Cornell University)|Apr 2, 2019

Topic ModelingComputer Science35 references48 citations

TL;DR

UHop introduces an unrestricted-hop framework for KBQA that splits relation extraction into single-hop prediction and a termination decision, enabling hop-free maximums and reducing search space while staying competitive with state-of-the-art models.

ABSTRACT

In relation extraction for knowledge-based question answering, searching from one entity to another entity via a single relation is called "one hop". In related work, an exhaustive search from all one-hop relations, two-hop relations, and so on to the max-hop relations in the knowledge graph is necessary but expensive. Therefore, the number of hops is generally restricted to two or three. In this paper, we propose UHop, an unrestricted-hop framework which relaxes this restriction by use of a transition-based search framework to replace the relation-chain-based search one. We conduct experiments on conventional 1- and 2-hop questions as well as lengthy questions, including datasets such as WebQSP, PathQuestion, and Grid World. Results show that the proposed framework enables the ability to halt, works well with state-of-the-art models, achieves competitive performance without exhaustive searches, and opens the performance gap for long relation paths.

Motivation & Objective

• Relax the fixed hop limit in knowledge-based relation extraction to handle long and unknown-length relation paths.
• Decompose relation extraction into single-hop relation prediction and a termination (halt) decision.
• Enable compatibility with existing state-of-the-art models by integrating into a transition-based search framework.
• Reduce the search space from exponential to polynomial without sacrificing performance.

Proposed method

• Two-subtask framework: (i) single-hop relation extraction via pairwise classification of candidate outbound relations; (ii) comparative termination decision to decide when to halt based on scores.
• Greedy transition-based search that traverses from the topic entity along chosen relations, updating the current entity and candidate set at each hop.
• Use of a termination decision that compares the current hop score to outbound options to decide whether to continue or stop.
• Dynamic question representation that updates with each chosen relation to focus on relevant information for the next hop.
• Joint end-to-end training that optimizes both relation extraction and termination decision through a combined loss.

Experimental results

Research questions

• RQ1 Can an unrestricted-hop framework achieve competitive KBQA performance without predefining a maximum hop count?
• RQ2 Does a transition-based search with a termination decision effectively reduce search space while maintaining accuracy on real and long-path questions?
• RQ3 How does dynamic question representation affect relation extraction and termination decisions across multi-hop questions?

Key findings

• UHop yields competitive results with state-of-the-art models when used within the framework, while removing the need for a fixed hop limit.
• The search space with UHop is reduced from roughly n(n-1)^(L-1) to about n(L+1), and experiments show a 30% reduction in candidate relations on WebQSP.
• On WebQSP (1–2 hops) and PathQuestion/PQL (2–3 hops) datasets, HR-BiLSTM and ABWIM within UHop perform comparably to their standalone versions and outperform several baselines.
• Grid World experiments demonstrate UHop enables long-path reasoning (up to 10 hops) where baseline models struggle without the unrestricted-hop framework.
• Dynamic question representations (+DQ) generally improve relation extraction on PQ/PQL datasets, particularly at the first hop, and aid termination decisions when trained jointly within UHop.