When building systems to understand human language, researchers have long faced a fundamental challenge: how to represent the messy, interconnected web of real-world knowledge. One approach relies on hand-coding formal logic; another relies on training massive neural networks on ungrounded text corpora. However, a compelling alternative lies in a resource humans have spent centuries perfecting: the dictionary.
Dictionary definitions (or glosses) are not just arbitrary strings of text. They are highly dense, semi-structured explanations designed to map complex concepts to simpler ones. By treating a dictionary as a self-contained, recursive semantic system, we can extract structured knowledge bases, discover semantic primitives, and learn how phrases compose into meanings.
The Dictionary as a Structured Graph
At its core, a dictionary definition follows a predictable linguistic formula. To explain a word, a lexicographer typically states its general category (the genus) and its distinguishing characteristics (the differentia). For example, defining a sedan as "a car with four doors" immediately establishes a hierarchical link (sedan $\rightarrow$ car) and a semantic relationship (has-part $\rightarrow$ four doors).
This inherent structure means that dictionaries can be decoded as a metalanguage to reveal underlying linguistic frameworks. Early in the history of computational linguistics, Giacomo Ferrari proposed the "dictionary-machine" hypothesis in Dictionnaire Automatique Et Dictionnaire-Machine: Une Hypothese, arguing that definitions encode structured syntactic and semantic data rather than purely unstructured explanations. Rather than manually building semantic lexicons from scratch, systems could programmatically parse human-readable definitions to resolve linguistic ambiguities.
This idea was put into practice by Jean-Louis Binot et al. in A Semantic Expert Using an Online Standard Dictionary. The authors demonstrated that a system could automatically resolve complex syntactic ambiguities—such as prepositional phrase attachments—by parsing definitions in an online standard dictionary.
[Sedan] ──(definition)──> "A car with four doors"
│ │
▼ ▼
(Hypernym) (Semantic Relation)
[Car] [with four doors]
To transform these definitions into machine-readable knowledge bases, algorithms must systematically untangle specific phrasal patterns. For instance, verb definitions frequently use the structure "to VERB with NP." In Disamibiguating and Interpreting Verb Definitions, Yael Ravin showed that an automated system could recursively look up the semantic categories of constituent words to classify the role of the preposition "with" (distinguishing between an instrument, like "cutting with a knife," or a manner, like "acting with caution"). By identifying these key linguistic anchors, as also explored by Ildikó Pilán et al., unstructured lexicographical entries can be converted into highly accurate, categorized datasets.
Decoding the Metalanguage of Primitives
If you follow the chain of definitions in a dictionary—using words to define words—you eventually run into a problem of circularity. To avoid infinite regress, dictionaries naturally rely on a restricted, highly optimized set of core words to define everything else. This is known as a controlled defining vocabulary.
These restricted vocabularies are incredibly valuable for computer science because they function as natural semantic primitives, or "sememes"—the smallest indivisible units of meaning. In Automatic Construction of Sememe Knowledge Bases via Dictionaries, Fanchao Qi et al. leveraged this insight to automatically construct Sememe Knowledge Bases for English and French. By treating the controlled defining vocabulary of a standard dictionary as a set of foundational semantic building blocks, their system bypassed the need for decades of manual expert annotation, achieving higher consistency and stronger performance on downstream language tasks than human-curated baselines.
This evolutionary process of refining dictionaries into structured primitives allows for the creation of Machine-Tractable Dictionaries (MTDs). As detailed by Cheng-ming Guo et al. in The Evolution of Machine-Tractable Dictionaries, systematically processing existing machine-readable dictionaries allows systems to extract these definition primitives automatically, shifting the paradigm away from manual resource creation toward automated extraction.
Natural Supervision for Phrasal Composition
In modern machine learning, dictionary definitions offer a unique, naturally occurring supervision signal for a difficult problem: phrasal composition. How do we teach a model that the combination of the words "a", "large", "grey", "mammal", "with", "a", "trunk" equals the single concept of an "elephant"?
Dictionaries provide this exact mapping for free. Every entry is a paired equation:
$$\text{Multi-Word Definition} = \text{Single Lexical Concept}$$
In Learning to Understand Phrases by Embedding the Dictionary, Felix Hill et al. utilized this bridge by training neural networks to map multi-word dictionary definitions directly to the pre-trained vector representation of the single word being defined. Because the model is forced to compress a phrase into a specific semantic point, it learns generalizable rules of phrasal composition without requiring any manually labeled sentence-similarity datasets.
"A large grey mammal with a trunk" ──(Neural Network)──> [ Vector Space ] <── [ Elephant ]
This mapping also works in reverse to improve Word Sense Disambiguation (WSD). Words are often polysemous (having multiple meanings), but each meaning has a distinct dictionary gloss. Jie Wang et al. demonstrated in Learning Sense Representation from Word Representation for Unsupervised Word Sense Disambiguation (Student Abstract) that by training models to project dictionary glosses into a shared embedding space, they could derive high-quality representations for specific word senses. This knowledge-based approach allows models to accurately identify which sense of a word is being used in a sentence, purely by leveraging existing lexical resources.
Scaling Beyond Traditional Dictionaries
The core intuition of dictionary-based extraction—that explanatory text can be mined as structured data—extends naturally to modern, collaborative online resources like Wikipedia and Wiktionary. While traditional dictionaries are authoritative but limited in scope, collaborative web resources provide massive, multilingual sense inventories that scale to specialized domains.
For example, Mei Yang et al. showed that Wikipedia and Wiktionary entries can be treated as "pseudo-glosses" to generate semantic similarity signals for unsupervised word sense disambiguation in machine translation. Similarly, Sara Tonelli et al. demonstrated in Wikipedia as Frame Information Repository that Wikipedia can be used as a massive sense inventory to automatically populate expert semantic networks like FrameNet with new examples and lexical units across multiple languages.
Whether dealing with a centuries-old print lexicon or a modern collaborative wiki, the underlying principle remains identical: human explanations are inherently structured. By treating definitions as data, natural language processing systems can unlock a vast, self-contained network of common sense and semantic relationships that humans have already explicitly mapped out.
Key Foundational Work
Dictionnaire Automatique Et Dictionnaire-Machine: Une Hypothese (1973) — contributed 2 concept elements to this direction.