[Paper Review] A Survey on Complex Question Answering over Knowledge Base: Recent Advances and Challenges

Bin Fu, Yunqi Qiu+4 more|arXiv (Cornell University)|Jul 26, 2020

Topic Modeling5 references51 citations

TL;DR

This survey categorizes complex KBQA methods into Information Retrieval-based and Neural Semantic Parsing-based approaches, discusses datasets, and analyzes future challenges and frontier trends.

ABSTRACT

Question Answering (QA) over Knowledge Base (KB) aims to automatically answer natural language questions via well-structured relation information between entities stored in knowledge bases. In order to make KBQA more applicable in actual scenarios, researchers have shifted their attention from simple questions to complex questions, which require more KB triples and constraint inference. In this paper, we introduce the recent advances in complex QA. Besides traditional methods relying on templates and rules, the research is categorized into a taxonomy that contains two main branches, namely Information Retrieval-based and Neural Semantic Parsing-based. After describing the methods of these branches, we analyze directions for future research and introduce the models proposed by the Alime team.

Motivation & Objective

• Summarize datasets and their characteristics for complex KBQA.
• Categorize mainstream KBQA methods into IR-based and neural semantic parsing-based approaches.
• Analyze advantages, limitations, and applicability of traditional, IR, and NSP methods for complex questions.
• Identify frontier trends and practical challenges in real-world KBQA deployments.
• Highlight models and techniques used by the Alime team for industrial KBQA applications.

Proposed method

• Review and classify KBQA methods into two main branches: Information Retrieval-based (IR) and Neural Semantic Parsing-based (NSP).
• Describe traditional template/rule-based methods and their limitations.
• Detail IR-based methods with feature engineering and representation learning, including multi-hop and external knowledge integration.
• Describe NSP-based methods that generate executable logical forms via query graphs, STAGG/MultiCG extensions, and encoder-decoder approaches.
• Discuss encoder-decoder and neural-symbolic machines that address weak supervision and end-to-end training.
• Outline other emerging methods and improvements to KV-MemNN, multi-turn interaction, and robustness strategies.

Experimental results

Research questions

• RQ1 What are the dominant methodological branches for complex KBQA and their respective strengths and weaknesses?
• RQ2 What datasets and benchmarks drive complex KBQA research, and what characteristics do they exhibit?
• RQ3 What are the frontier challenges and future directions for robust, scalable, and industrial-grade KBQA systems?
• RQ4 How have industrial teams (e.g., Alime) approached complex KBQA in vertical domains?
• RQ5 What techniques improve entity linking, relation matching, and constraint handling in complex KBQA?

Key findings

• Two mainstream KBQA methods dominate: Information Retrieval-based approaches and Neural Semantic Parsing-based approaches.
• NSP methods typically achieve higher performance than IR methods on complex questions but require more training data or weak supervision.
• IR methods benefit from multi-hop reasoning and representation learning, and can be enhanced with external knowledge but often lack interpretability and robust constraint handling.
• Neural Semantic Parsing methods using query graphs and encoder-decoder architectures can better handle complex constraints and aggregations, but training challenges remain without gold logical forms.
• There is active integration of external knowledge sources and multi-hop reasoning to address KB incompleteness and complex question structures.
• Industrial applications, such as the Alime KBQA effort, demonstrate practical deployment, including multi-constraint questions and efficiency improvements.