Polysemy describes one word that means more than one thing depending on its context, where all of the meanings are related. Unlike humans, automated AI systems struggle with polysemes, partly because most structured knowledge cannot manage more than one meaning for each term. This can lead to AI making guesses and mistakes, or using important terminology inconsistently – especially when intelligent systems start to scale.
Put another way, most AI today cannot reliably manage multiple related meanings for the same word as well as it can handle synonyms, which are multiple words for the same meaning.
Why is polysemy a problem for AI?
Until natural language processing (NLP) came along, the polysemy problem was more of an annoyance than a serious challenge. If the finance department of an organisation had one definition of customer (someone who pays), and the sales department had another (someone who makes an enquiry), it could be managed by humans adjusting their siloed data manually when needed. It was ‘just semantics’.
But when an organisation plans to roll out AI, and interoperate with other domain intelligent systems by sharing data, polysemy presents a hard limit on what can be automated. Understanding a word’s meaning in different contexts (finance, engineering, sales) cannot be left to educated guesswork by a large language model (LLM), especially in critical domains like medicine.
Without the ability to manage meaning at scale, reliable and useful AI services risk semantic drift, where the meanings of words diverge from what they are supposed to mean. This presents a major roadblock to AI expansion across the global economy. It is hard to imagine a reliable agentic web, for example, unless polysemy can be effectively managed.
Why is polysemy harder than synonymy?
Many enterprise knowledge systems are implemented in ways that assume one meaning per term. This is because the common rules and constraints that enable structured data to be shared between systems, known as ontology standards, generally try to avoid the modelling and governance complexity that polysemy brings.
Instead of learning just one set of near-equivalent words that match a single meaning (for example, the synonyms ‘big’, ‘large’, or ‘enormous’) a polysemy-capable system must also learn separate clusters of synonyms for each difference in meaning.
So let’s say ‘customer’ has five different meanings corresponding to five different departments in an organisation. Each department might also have three synonyms for that particular use of the word (‘client’, ‘patient’, ‘user’), depending on its context. Legacy ontology software is unlikely to handle this level of complexity, especially at scale.
While humans have evolved to understand subtle differences in meaning based on how a word is used, AI needs to be exhaustively trained in the ways meaning is affected by context, and even then it can make mistakes. The verb ‘run’ has dozens of distinct meanings, for example.
How is polysemy managed currently by AI?
To some extent, specialised language models can differentiate meanings for a word when rich context is available, using statistical association. This approach tends to work well in addressing homonymy, where the same word – or the same pronunciation – has more than one unrelated meaning, such as a ‘bright’ student and a ‘bright’ light. However, when addressing multiple related meanings (polysemy), the margin for error is too high for enterprise use, especially in regulated or mission-critical domains.
Vertical AI agents that are focused on a single narrow domain, or application ontologies that are designed for a specific task, can help improve semantic accuracy and performance – but only by limiting exposure to polysemy, rather than mastering it.
A 2024 study published in Computational Linguistics addresses the complexity of polysemous sense extensions for both traditional computational approaches and modern neural models. It also surveys seminal theories on the mental processing of polysemes, and includes an estimate that around 40% of English language words are polysemous – a tough nut for AI to crack.