# Fact-Tree Reasoning for N-ary Question Answering over Knowledge Graphs

**Authors:** Yao Zhang, Peiyao Li, Hongru Liang, Adam Jatowt, Zhenglu Yang  
**arXiv:** [2108.08297](https://arxiv.org/abs/2108.08297)  
**Date:** 2021-08-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2108.08297v2/

## Key points

- Significant focus on multi-hop reasoning frameworks in question answering to improve efficiency and interpretability on Knowledge Graphs
- Identified improvements: upgrading basic reasoning unit to fact and upgrading reasoning structure from chain to tree
- Fact-Tree reasoning framework proposed for complex n-ary fact questioning scenarios, using facts as basic reasoning unit and explicit tree structure for interpretability
- Integration of n-ary KG embedding model into Fact-Tree framework shows promising results in addressing complex reasoning scenarios caused by incomplete KGs
- Comprehensive evaluations demonstrate high accuracy in answer prediction and strong reasoning abilities compared to existing baselines

## Summary

, , , , 

In the field of question answering (QA), there has been a significant focus on multi-hop reasoning frameworks in recent years to improve the efficiency and interpretability of answer reasoning on Knowledge Graphs (KG). These frameworks aim to address limitations faced by traditional QA models, particularly when it comes to answering n-ary fact questions. However, current approaches still struggle with these types of questions due to their linear reasoning nature. To overcome this issue, two feasible improvements have been identified: upgrading the basic reasoning unit from entity or relation to fact, and upgrading the reasoning structure from chain to tree. Previous KGQA models have utilized structured query graphs synthesized from questions to match with KGs for answers, offering high interpretability but facing challenges due to incomplete KGs. On the other hand, some models compute semantic similarity between questions and candidate answers directly in latent space, overcoming incomplete KG limitations but lacking interpretability. The Fact-Tree reasoning framework proposed in this work uses facts as the basic reasoning unit and an explicit tree structure for highly interpretable reasoning. By transforming questions into fact trees and conducting iterative fact reasoning, this approach offers a novel solution for complex n-ary fact questioning scenarios. It integrates an n-ary KG embedding model into the Fact-Tree framework, showing promising results in addressing complex reasoning scenarios caused by incomplete KGs. Through comprehensive evaluations on various datasets, including n-ary fact KGQA datasets and binary datasets, this framework demonstrates high accuracy in answer prediction and strong reasoning abilities compared to existing baselines. This work provides a solid foundation for exploring complex reasoning scenarios in QA tasks and serves as a preliminary baseline approach for future research in this area.