[Paper Review] A Question Answering System Using Graph-Pattern Association Rules (QAGPAR) On YAGO Knowledge Base

Wahyudi Wahyudi, Masayu Leylia Khodra+2 more|arXiv (Cornell University)|Feb 2, 2019

Semantic Web and Ontologies3 references4 citations

TL;DR

This paper proposes QAGPAR, a question answering system that leverages graph-pattern association rules on the YAGO knowledge base to infer answers when direct retrieval fails. By employing a waterfall-style architecture involving question classification, graph component generation, query generation, and processing, the system achieves accurate answers using high-confidence, logically sound association rules, demonstrating that rule quality directly determines answer correctness.

ABSTRACT

A question answering system (QA System) was developed that uses graph-pattern association rules on the YAGO knowledge base. The answer as output of the system is provided based on a user question as input. If the answer is missing or unavailable in the database, then graph-pattern association rules are used to get the answer. The architecture of this question answering system is as follows: question classification, graph component generation, query generation, and query processing. The question answering system uses association graph patterns in a waterfall model. In this paper, the architecture of the system is described, specifically discussing its reasoning and performance capabilities. The results of this research is that rules with high confidence and correct logic produce correct answers, and vice versa

Motivation & Objective

• To address the challenge of answering complex or incomplete queries in knowledge base question answering by extending beyond direct fact retrieval.
• To improve answer accuracy in cases where required facts are missing or unavailable in the YAGO knowledge base.
• To develop a reasoning mechanism using graph-pattern association rules that captures semantic relationships between entities and relations.
• To evaluate the effectiveness of association rules in enhancing QA system performance through confidence and logical consistency.
• To demonstrate that high-confidence, logically valid rules yield correct answers, while low-quality rules lead to errors.

Proposed method

• The system uses a waterfall architecture: question classification, graph component generation, query generation, and query processing.
• Graph-pattern association rules are mined from the YAGO knowledge base to represent common relational patterns among entities.
• Rules are evaluated based on confidence and support metrics to prioritize reliable patterns for inference.
• The system maps user questions to graph patterns by analyzing syntactic and semantic structures.
• Query generation involves translating the matched graph pattern into a SPARQL query for execution on the YAGO knowledge base.
• Answer retrieval combines direct lookup with rule-based inference when facts are missing, using high-confidence rules to derive plausible answers.

Experimental results

Research questions

• RQ1 Can graph-pattern association rules effectively infer answers when facts are missing in the YAGO knowledge base?
• RQ2 How does rule confidence and logical consistency affect the accuracy of inferred answers in a QA system?
• RQ3 To what extent does the proposed architecture improve answer retrieval beyond standard knowledge base queries?
• RQ4 What is the impact of rule quality on the reliability of the system’s output?
• RQ5 Can the system generalize across diverse question types using pattern-based reasoning?

Key findings

• Rules with high confidence and correct logical structure consistently produced accurate answers, confirming the reliability of rule-based inference.
• Incorrect or low-confidence rules led to erroneous answers, highlighting the importance of rule quality in the system’s performance.
• The system successfully retrieved answers for questions where direct knowledge base lookup failed, demonstrating the value of rule-based reasoning.
• The architecture effectively mapped natural language questions to semantic graph patterns, enabling structured query generation.
• The use of association rules significantly enhanced answer coverage and robustness in the absence of explicit facts.
• Performance was directly correlated with rule confidence, validating the use of confidence thresholds as a filtering mechanism.