Derivational morphology encompasses the word-formation processes by which new lexemes are created from existing ones. Adding the suffix "-ness" to the adjective "kind" yields the noun "kindness"; adding the prefix "un-" to "happy" yields "unhappy"; adding "-ize" to "modern" yields the verb "modernize." Unlike inflection, which creates different grammatical forms of the same word, derivation produces new words that may differ in meaning, part of speech, and syntactic behavior. Computationally, derivation is more irregular and semantically varied than inflection, making it harder to model systematically.
Types of Derivational Processes
Suffixation: kind + -ness → kindness (Adj → Noun)
Category change: modern + -ize → modernize (Adj → Verb)
Argument structure: employ + -ee → employee (Agent → Patient)
Productivity: P(affix) = |types with affix| / |possible bases|
Some affixes are fully productive (-ness), others are not (*-th)
Derivational processes vary in productivity — the degree to which they can apply to new bases. English "-ness" is fully productive (any adjective can take it), while "-th" is unproductive (we have "warmth" and "growth" but cannot create new forms). Computational models must capture this gradient of productivity, which correlates with semantic transparency: more productive affixes tend to have more predictable meanings. Hay and Baayen (2005) proposed that relative frequency determines decomposability: a derived form is more likely to be decomposed if its frequency is lower than its base.
Computational Modeling of Derivation
Derivational morphology is harder to model computationally than inflection for several reasons. The meaning of derived words is not always predictable from their parts — "department" is not transparently derived from "depart." Derivation can apply recursively and in varying orders, creating words like "un+friend+li+ness" where the order of attachment matters for interpretation. Furthermore, derivational paradigms are incomplete: while "happiness" and "sadness" exist, "*angriness" is marginal despite "angry" being a common adjective.
Word embeddings provide a window into how derivation affects meaning. Lazaridou et al. (2013) showed that derivational affixes induce approximately linear transformations in embedding space: the vector offset from "quick" to "quickly" is similar to the offset from "slow" to "slowly." This regularity allows vector arithmetic to predict the embeddings of unseen derived forms. However, semantically opaque derivations (like "department" from "depart") do not follow these regular patterns, reflecting the dual-route nature of derivational morphology.
Derivation in NLP Applications
Derivational analysis benefits several NLP applications. In information retrieval, recognizing that "computerize," "computerization," and "computational" share the root "compute" improves recall. In text simplification, replacing derived forms with periphrastic alternatives ("nationalization" to "making something national") can improve readability. In vocabulary learning for second languages, understanding derivational families helps learners expand their vocabulary systematically.
Recent neural approaches to derivation leverage character-level models and pretrained embeddings. BERT and similar models implicitly capture derivational relationships through their subword tokenization and contextual representations. However, studies show that these models do not fully capture the semantic compositionality of derived forms, particularly for less frequent and semantically opaque derivations. Explicitly modeling derivational structure remains an open challenge in computational linguistics.