Question-guided Knowledge Graph Re-scoring and Injection for Knowledge Graph Question Answering

TL;DR

A Question-guided Knowledge Graph Re-scoring method (Q-KGR) is proposed to eliminate noisy pathways for the input question, thereby focusing specifically on pertinent factual knowledge and introduces Knowformer, a parameter-efficient method for injecting the re-scored knowledge graph into large language models to enhance their ability to perform factual reasoning.

Abstract

Knowledge graph question answering (KGQA) involves answering natural language questions by leveraging structured information stored in a knowledge graph. Typically, KGQA initially retrieve a targeted subgraph from a large-scale knowledge graph, which serves as the basis for reasoning models to address queries. However, the retrieved subgraph inevitably brings distraction information for knowledge utilization, impeding the model's ability to perform accurate reasoning. To address this issue, we propose a Question-guided Knowledge Graph Re-scoring method (Q-KGR) to eliminate noisy pathways for the input question, thereby focusing specifically on pertinent factual knowledge. Moreover, we introduce Knowformer, a parameter-efficient method for injecting the re-scored knowledge graph into large language models to enhance their ability to perform factual reasoning. Extensive experiments on multiple KGQA benchmarks demonstrate the superiority of our method over existing systems.

Figures (8)

• Figure 1: Given the QA context (question and answers), we retrieve a subgraph including question entities (red), answer entities (blue) and other neighbor entities (grey). The correct answer is piggybank.
• Figure 2: The overall framework of our method. Given a retrieved subgraph from origin KG, we revise it through question-guided knowledge graph re-scoring (Q-KGR, I). Then we utilize question text and revised subgraph to conduct knowledge modeling, injection and reasoning (II). Knowformer (III) consists of self-attention layers and a customized feed-forward network (FFN) layer. In the FFN layer, $key$ and $value$ represent original weight matrices, $LoRA_k$ and $LoRA_v$ correspond to LoRA weight parameters, and $\phi_k$ and $\phi_v$ are the knowledge vectors mapped from graph latent space to the parameter space of FFN. Gray denotes frozen model parameters, while green indicates updated model parameters.
• Figure 3: Performance of different methods across divisions varying number of prepositional phrases.
• Figure 4: Qualitative analysis results.
• Figure 5: With the increase in training steps, it becomes apparent that the injected knowledge gradually integrates into the distribution of the FFN parameters.