Neural-Symbolic Reasoning over Knowledge Graphs: A Survey from a Query Perspective
This survey provides a comprehensive review of neural-symbolic reasoning over knowledge graphs, focusing on query types and classification while exploring the innovative integration of large language models to bridge the gap between deep learning and symbolic reasoning.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine a vast, digital library where every fact about the world is stored not in books, but as a massive web of connections. In this library, a person is linked to their job, a city is linked to its country, and a movie is linked to its director. This structure, known as a knowledge graph, acts as a map of human understanding, allowing computers to navigate from one piece of information to another. For decades, scientists have tried to teach machines to reason through this map, to ask a question and find a new truth hidden within the connections. However, the real world is messy. Facts are often missing, and data can be noisy or contradictory. Traditional methods, which rely on strict, rigid rules like a legal code, often break down when faced with this imperfection. They cannot handle the gray areas of reality.
To solve this, researchers have been developing a new kind of intelligence that blends two different ways of thinking. One way is like a computer program following a strict recipe, precise but brittle. The other is like a human brain learning from patterns, flexible but sometimes vague. By merging these approaches, scientists are creating systems that can reason through the messy web of knowledge with both accuracy and adaptability. A recent survey by researchers at Wayne State University and the University of Illinois at Urbana-Champaign takes a deep look at how this is being done. They examine how these systems answer questions, from simple one-step inquiries to complex, multi-layered puzzles, and how they are now being joined by a new, powerful tool: large language models, the same technology that powers modern chatbots.
The researchers organized their review by looking at the types of questions these systems are asked to answer. The simplest questions are single-hop queries, where the answer is just one step away. If you ask, "Who is the CEO of Company X?", the system looks for a direct link. For these tasks, older methods used strict rules written by experts, but these often failed when the data was incomplete. Newer methods use neural networks, which are computer models that learn from examples, to guess the missing links. The most promising approach combines both: it uses the learning power of neural networks to handle uncertainty while keeping the logical structure of rules to ensure the answers make sense. This hybrid method allows the system to infer that if a person works for a company in a specific city, they likely live in that city, even if that specific fact was never explicitly written down.
When the questions become more complex, involving multiple steps or logical conditions like "and," "or," and "not," the challenge grows significantly. These are not just about finding a single link but about navigating a path through the web. The survey explains that researchers have developed ways to treat these questions as mathematical problems that can be solved by moving through a geometric space. Instead of just checking if a fact is true or false, the system assigns a degree of truth to different possibilities. Some methods visualize these possibilities as shapes, like boxes or spheres, in a multi-dimensional space. By seeing how these shapes overlap or fit together, the computer can determine the answer to a complicated question, such as "Find all actors who have worked with a director from France but have not won an Oscar." The researchers found that while these methods are powerful, they still struggle with the most intricate logical structures, suggesting that the field is still maturing.
The survey also explores how these systems handle questions asked in everyday language, rather than in a formal code. When a human asks, "Who played the villain in the movie that won the most awards last year?", the system must first understand the sentence, break it down, and then search the knowledge graph. Early attempts tried to translate the sentence into a rigid command, but this often failed if the phrasing was unusual. Newer approaches use reinforcement learning, a technique where the system learns by trial and error, to navigate the graph and find the answer. In conversations where a user asks a follow-up question, the system must remember the context of the previous turn. The researchers noted that while progress has been made, these systems still sometimes get lost in long chains of reasoning or fail to distinguish between similar paths in the data.
Perhaps the most exciting development discussed in the survey is the integration of large language models with knowledge graphs. These models are incredibly good at understanding human language but often make things up, a problem known as hallucination, because they lack access to a verified database of facts. Conversely, knowledge graphs are full of facts but are terrible at understanding natural language. The researchers found that combining them creates a powerful partnership. In one approach, the language model acts as a guide, helping to translate a human question into a search query for the knowledge graph. In another, the knowledge graph acts as a fact-checker, providing the language model with verified information to prevent it from making things up. The most advanced methods allow the two to work together in a loop, where the language model helps the graph find relevant information, and the graph helps the language model refine its answer. This mutual assistance creates a system that is both fluent in conversation and grounded in reality.
Looking ahead, the researchers suggest that the next frontier lies in expanding these systems beyond text. They envision knowledge graphs that include images, videos, and audio, allowing the computer to reason about the world in a way that is closer to human perception. They also see a future where these systems can bridge different languages, finding connections between facts expressed in English, Chinese, or Spanish without needing a human translator. While the current technology has made significant strides in making machines smarter and more reliable, the survey concludes that there is still much work to be done. The goal is to build systems that can not only retrieve information but truly understand the complex, interconnected nature of human knowledge, handling the messiness of the real world with the same grace as a human mind.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.