SA-RGLP: Syntax-Aware Residual Graph Label Propagation with LLM Soft Priors for Metaphor Detection

Language is a living thing that often speaks in riddles. When a writer says a student "grasped" a difficult idea, they are not describing a physical hand closing around a solid object. They are using a metaphor, a cognitive shortcut where we understand one thing by borrowing the language of another. For computers, which process language as strings of data, these riddles are notoriously difficult to solve. A machine can easily recognize that "grasped" usually means holding a tool, but it struggles to know when that same word has shifted meaning to describe an abstract understanding. This difficulty matters because metaphors are not just poetic flourishes; they are fundamental to how humans think, feel, and communicate. If a computer cannot tell the difference between a literal grip and a mental one, it cannot truly understand a story, a news report, or a conversation.

For years, researchers have tried to teach computers to spot these hidden meanings. Most methods have treated each sentence as an isolated island, analyzing the words around a target word to decide if it is being used metaphorically. While this works well enough for simple cases, it misses the bigger picture. Metaphors are not random accidents; they follow patterns. The way we use the word "grasp" to mean "understand" is a recurring theme across thousands of different sentences. By looking at only one sentence at a time, computers miss the connections between them. Furthermore, while large language models—powerful AI systems trained on vast amounts of text—have a deep knowledge of how words are used, they often struggle to apply that knowledge consistently to specific, tricky cases without human guidance.

A team of researchers at Xinjiang University has developed a new approach to bridge this gap, calling it SA-RGLP. Instead of treating every sentence as a separate puzzle, their system looks at the entire collection of sentences as a single, connected map. Imagine a vast library where every book is a sentence. Traditional methods read one book, close it, and move to the next. This new system, however, walks through the aisles, noticing that the book about "grasping a concept" sits on the same shelf as the book about "grasping a tool," and that the book about "grasping a rope" is nearby too. By seeing how these different usages relate to one another, the system can make a much smarter guess about what a word means in a specific context.

The process begins with a careful, step-by-step analysis of the sentence itself. The system first looks at the target word and its immediate surroundings, paying close attention to the grammar and the specific roles the words play. It asks: does the physical action described by the word clash with the abstract idea it is paired with? If a student "grasps" a concept, the system notes the tension between the physical act of holding and the abstract nature of an idea. This initial check produces a first guess, but the researchers knew this local view was often not enough to be certain.

To improve this guess, the system brings in a second layer of intelligence: a large language model acting as a knowledgeable consultant. This consultant does not just give a simple "yes" or "no" answer. Instead, it weighs the evidence and offers a nuanced probability, saying, for instance, that there is a seventy-five percent chance the word is being used metaphorically and a twenty-five percent chance it is literal. This "soft" advice preserves the uncertainty that is so common in human language, rather than forcing a rigid decision too early.

The most significant innovation, however, happens in the third stage. The system takes all the sentences it has analyzed and builds a global network, connecting them based on how similar their meanings are. If one sentence uses "grasp" to mean "understand" and another uses it to mean "hold," the system recognizes the link. It then uses a process called residual graph label propagation to share information across this network. Think of it as a group of people trying to solve a mystery; if one person is unsure, they look to their neighbors for clues. If the neighbors have strong evidence, that evidence helps clarify the doubt. In this digital network, the system allows the confident predictions of clear examples to gently nudge the uncertain predictions of ambiguous ones. It does not overwrite the initial guess but refines it, using the collective wisdom of the entire dataset to correct errors.

The researchers tested this method on three different sets of data, ranging from small collections of a few hundred sentences to massive datasets with thousands of examples. The results were striking. On the smaller datasets, the system performed well, but on the larger, more complex ones, it significantly outperformed previous methods. It achieved the highest accuracy scores ever recorded for two of the major benchmarks used in this field. The study showed that by combining a deep understanding of local grammar, the broad knowledge of a large language model, and the power of cross-sentence connections, the system could detect metaphors with a level of precision that isolated methods could not match.

Crucially, the researchers also tested what happened when they removed these different parts of the system. When they took away the ability to look at connections between sentences, the system's performance dropped sharply, proving that the global map was essential. When they removed the nuanced advice from the large language model, the system became less accurate, showing that the external knowledge was a vital supplement. The study concluded that the best way to understand a metaphor is not to look at a single word in a vacuum, but to see it as part of a vast, interconnected web of human expression. This approach moves the field of natural language processing closer to a true understanding of how humans use language to build meaning.