[论文解读] Introduction to Neural Network based Approaches for Question Answering over Knowledge Graphs

Nilesh Chakraborty, Denis Lukovnikov+4 more|arXiv (Cornell University)|Jul 22, 2019

Topic Modeling参考文献 97被引用 42

一句话总结

这篇文章综述基于神经网络的知识图谱问答，概述 KGQA 任务、数据集、范式（分类、排序、翻译）、训练设置以及未来方向。

ABSTRACT

Question answering has emerged as an intuitive way of querying structured data sources, and has attracted significant advancements over the years. In this article, we provide an overview over these recent advancements, focusing on neural network based question answering systems over knowledge graphs. We introduce readers to the challenges in the tasks, current paradigms of approaches, discuss notable advancements, and outline the emerging trends in the field. Through this article, we aim to provide newcomers to the field with a suitable entry point, and ease their process of making informed decisions while creating their own QA system.

研究动机与目标

• Introduce knowledge graph concepts and KGQA tasks and terminology.
• Review neural network based KGQA paradigms and their training/inference mechanisms.
• Summarize datasets used for KGQA and their characteristics.
• Discuss challenges such as spurious logical forms and search space management, and outline future directions.

提出的方法

• Classify KGQA neural approaches into three categories: classification, ranking, and translation.
• Explain training settings: fully supervised vs weakly supervised learning for semantic parsers.
• Describe key architectural components: encoders, span detectors, and scoring/ranking mechanisms.
• Discuss entity linking, relation identification, and operator/deduction steps necessary for forming logical forms.
• Highlight examples of systems and their training regimes to illustrate the approaches.

实验结果

研究问题

• RQ1 What are the main neural network paradigms used for KGQA (classification, ranking, translation) and how do they work?
• RQ2 How do KGQA systems handle training data availability (fully supervised vs weakly supervised) and the associated challenges?
• RQ3 What are the essential subtasks (entity linking, relation identification, operator detection) in KGQA, and how are they addressed in neural models?
• RQ4 What datasets are used for KGQA, and what characteristics (size, complexity, formal queries) differentiate them?
• RQ5 What are the current trends and future directions in neural network based KGQA research?

主要发现

• Neural KGQA approaches span simple classification to complex memory-augmented and translation-based models.
• KGQA training can be fully supervised or weakly supervised, each with distinct optimization and search challenges.
• Subtasks like entity linking and relation classification are critical components and are often handled with dedicated modules or end-to-end models.
• A large variety of KGQA datasets exist (e.g., SimpleQuestions, WebQuestions, LC-QuAD), differing in size and whether formal queries are provided.
• Spurious logical forms and exponentially growing search spaces are key challenges in semantic parsing for KGQA, motivating ranking and constrained search techniques.
• The article identifies emerging trends and potential future directions in KGQA research.