Subcategorization is a fundamental concept in generative linguistics, introduced by Noam Chomsky, referring to the syntactic restrictions that lexical items—particularly verbs—impose on the types and structures of their complements or arguments within a sentence.[1] These restrictions, formalized as subcategorization frames, specify the permissible syntactic contexts for a word's insertion into a phrase structure, ensuring grammatical well-formedness by linking lexical properties to broader syntactic rules.[1] For instance, the verb put requires both a noun phrase (NP) and a prepositional phrase (PP) as complements, as in "She put the book on the table," whereas an intransitive verb like sleep takes no such complements.[2]In Chomsky's framework from Aspects of the Theory of Syntax (1965), subcategorization forms part of the base component of generative grammar, where context-sensitive rules categorize lexical items into subclasses based on their valency—the number and category of arguments they select.[1] This mechanism distinguishes between broad categories like transitive (requiring one NP object) and ditransitive (requiring two NPs) verbs, but extends to more nuanced frames involving prepositional phrases, clauses, or adjectives.[2] Subcategorization frames are often represented formally, such as hit: [__ (NP)] to indicate an optional NP complement, highlighting how verbs "subcategorize" their syntactic environments rather than the reverse.[2]Beyond theoretical syntax, subcategorization plays a crucial role in computational linguistics and language acquisition studies, where building comprehensive verb subcategorization lexicons aids in natural language processing tasks like parsing and machine translation.[3] Research shows that subcategorization preferences vary across corpora and domains, influencing verb usage frequencies—for example, business English may favor certain frames over literary texts—underscoring its empirical basis in corpus data.[4] Automatic acquisition of these frames from large-scale corpora has become a key area, addressing challenges in manual lexicon development due to evolving language use across sublanguages and time periods.[5]
Core Concepts
Definition and Overview
Subcategorization refers to the specification in a lexical entry of the syntactic categories and types of complements that a word—typically a verb—requires or allows. This property, often termed c-selection, governs the insertion of lexical items into syntactic structures by dictating compatible phrasal contexts, such as the need for a direct object or prepositional phrase.[6] It is distinct from s-selection, which imposes semantic restrictions on the thematic compatibility of complements rather than their syntactic categories.[7]The concept was coined by Noam Chomsky in the 1960s, specifically in his seminal 1965 book Aspects of the Theory of Syntax, where "strict subcategorization rules" were introduced as context-sensitive mechanisms within the base component of generative grammar to regulate lexical insertion and ensure structural coherence.[8] This innovation built on earlier ideas from structural linguistics, particularly the notion of verb valency developed by Lucien Tesnière in Éléments de syntaxe structurale (1959), which described the obligatory arguments (actants) that predicates demand to form complete syntactic units.A classic example illustrates subcategorization: the verb put requires both a noun phrase (NP) and a prepositional phrase (PP) as complements, yielding well-formed sentences like "put the bookon the table," while omitting the PP results in ungrammaticality, as in "put the book."[9] In contrast, an intransitive verb like sleep subcategorizes for no complements, making "sleep the book" ill-formed due to the absence of any required arguments.[9]Subcategorization is essential for determining syntactic well-formedness, as it enforces lexical constraints that prevent invalid combinations and supports efficient parsing by resolving structural ambiguities in sentence processing. It forms the syntactic foundation within the broader argument structure framework, outlining the obligatory slots for predicate participants.[10]
Relation to Argument Structure
Argument structure refers to the lexical representation of a predicate's arguments, specifying their number, syntactic expression, and hierarchical relations, such as external arguments realized as subjects and internal arguments as objects. In frameworks like Lexical Functional Grammar (LFG), argument structure is encoded at an a-structure level that maps to functional structures, ensuring the syntactic realization of arguments without transformational derivations.[11] Similarly, in Head-driven Phrase Structure Grammar (HPSG), argument structure is captured through valence lists in lexical entries, which detail the syntactic features and order of arguments a head subcategorizes for.[12]Subcategorization integrates directly with argument structure by providing the syntactic specifications for how these arguments are realized in phrase structure. For instance, subcategorization frames dictate the categories and positions of complements, as in the verb give, which requires an NP for the external argument (subject), an NP for the first internal argument (theme), and a PP headed by to for the second internal argument (recipient), yielding constructions like "She gave the book to him."[1] This lexical specification enforces the predicate's argument positions, linking the abstract argument structure to concrete syntactic configurations.[13]Examples illustrate how subcategorization distinguishes verb types based on argument licensing. Ditransitive verbs like give or send subcategorize for two internal arguments, permitting structures with both a direct object and an oblique complement, as in "He sent the letter to her." In contrast, monotransitive verbs like hit or break subcategorize solely for one internal NP argument, restricting them to patterns like "She hit the ball," without allowance for an additional complement.[13] These differences highlight subcategorization's role in defining the syntactic skeleton of argument structure across verb classes.[14]Theoretically, subcategorization accounts for argument alternations by permitting multiple frames for the same predicate, as seen in the dative shift alternation. For verbs like give, one frame realizes the recipient as a PP (She gave the book to him), while an alternative frame promotes it to a direct object, demoting the theme to a second object (She gave him the book). This lexical flexibility in subcategorization frames allows the same argument structure to project varying syntactic realizations, resolving apparent variations in argument positions without altering the underlying predicate-argument relations.[13]
Selection Mechanisms
Thematic Roles
Thematic roles, also known as theta-roles, represent the semantic relations that arguments bear to a predicate, providing essential interpretive content in subcategorization by specifying how participants contribute to the event described by the verb.[15] In linguistic theory, these roles help determine the meaning of a sentence by linking the verb's lexical properties to its arguments, ensuring that the syntactic structure aligns with semantic requirements.[16]Core thematic roles include the agent, which identifies the initiator or instigator of an action, typically a volitional entity exerting control; the patient, denoting the entity affected or undergoing change as a result of the action; the theme, referring to an entity that undergoes motion or relocation without inherent change; the goal, marking the endpoint or recipient toward which motion or transfer is directed; and the experiencer, capturing a perceiver or sentient being involved in states of consciousness or emotion.[15] For instance, in the sentence "John broke the window," John serves as the agent, initiating the action, while the window functions as the patient, being directly affected and altered by it.[17] These roles are prototypical rather than strictly discrete, allowing for gradations based on contextual properties like volitionality or affectedness.[18]A fundamental constraint on thematic role assignment is the theta-criterion, proposed by Chomsky, which mandates a one-to-one correspondence: each argument must receive exactly one theta-role, and each theta-role specified by the predicate must be assigned to precisely one argument.[18] This principle, introduced in Chomsky's 1981 framework, ensures that predicates saturate their semantic requirements without redundancy or omission, thereby regulating the well-formedness of argument structures.In relation to subcategorization, lexical entries for predicates explicitly associate syntactic argument positions with specific theta-roles, guiding how verbs select and interpret their complements.[19] For example, transitive verbs like "break" subcategorize for a subject-agent and object-patient pair, embedding these theta-role assignments within the verb's lexical specification to enforce semantic coherence in syntax.[16] This linkage, as explored in early generative semantics, allows subcategorization frames to incorporate theta-grid information, where the verb's argument-taking properties are tied to role distribution.[20]Thematic role assignment exhibits variations across languages, with greater fluidity in nominative-accusative systems like English, where agents typically align as subjects regardless of transitivity, permitting flexible mappings through passivization or adverbial modifications.[17] In contrast, ergative languages enforce more rigid alignments, often marking agents of transitive verbs distinctly from patients or themes (which pattern together as absolutive), leading to less neutralization between roles and stricter subcategorization constraints on case assignment.[16] This difference highlights how theta-roles interact with morphological case to influence predicate interpretation in diverse grammatical systems.
Selectional Restrictions
Selectional restrictions, often termed s-selection, represent the semantic dimension of subcategorization, imposing constraints on the complements of a predicate based on their semantic properties rather than solely syntactic categories. These restrictions mandate that arguments fulfill predicate-specific semantic features, such as animacy, concreteness, or edibility, to avoid type clashes that render sentences semantically anomalous or ungrammatical. For instance, the verb "kick" selects for a concrete, physical object, making "She kicked the ball" acceptable but "She kicked freedom" infelicitous due to the abstract nature of the complement.This framework originates in Noam Chomsky's work, where he distinguished semantic selectional rules from syntactic subcategorization rules in his 1965 analysis, later elaborated in the context of lexical properties in "Remarks on Nominalization" (1970), separating what would later be formalized as s-selection from c-selection. In Chomsky's 1965 account, selectional restrictions arise from semantic markers associated with lexical items, supplementing subcategorization frames to account for semantic well-formedness.[21][22]Classic examples illustrate these constraints: the verb "eat" requires an edible object, permitting "eat an apple" but not "eat a rock," as the latter violates the [+edible] feature, leading to a perceived ungrammaticality. Similarly, "regret" selects for a [+human] subject, explaining the oddity of "*The rock regrets falling" compared to "John regrets falling." Such type mismatches highlight how s-selection enforces semantic coherence, often resulting in judgments of deviance even when syntactic structures are intact.Theoretical debates center on whether s-selection is strictly lexical, encoded as fixed features in the lexicon as per Chomsky's generative approach, or derived from broader ontological categories reflecting world knowledge. Critics from cognitive linguistics, such as George Lakoff, argue that selectional restrictions are not rigid lexical rules but emerge from prototype-based, encyclopedic semantics, where compatibility is graded and context-dependent rather than binary.[23]
Syntactic Principles
Projection Principle
The Projection Principle, introduced by Noam Chomsky as part of Government and Binding theory, posits that subcategorization frames— the lexical specifications dictating the complements required by a head—must be satisfied at every syntactic level of representation, including D-structure, S-structure, and Logical Form (LF). This principle ensures that the argument structure projected from the lexicon remains invariant throughout the syntactic derivation, preventing any transformations from altering or ignoring these requirements.[24]Formally, the Projection Principle is stated as follows: Representations at each syntactic level (i.e., LF, S-structure, and D-structure) are projected from the lexicon, in that they observe the subcategorization properties of lexical items. This formulation implies that lexical properties, such as selectional requirements for arguments or sisters, must project fully into phrase structure representations, with no addition or deletion of elements permitted during the derivation process. In essence, if a lexical item like a verb subcategorizes for a particular complement at one level, that complement must appear as an immediate constituent or satisfy the selectional relation at all levels.[24]The principle is exemplified in passive constructions, where subcategorization is preserved despite NP-movement. Consider the verb put, which subcategorizes for both an NP object and a PP locative complement. The active sentence John put the book on the table satisfies this at D-structure, and the corresponding passive The book was put on the table retains the PP at S-structure; omitting it, as in the ungrammatical The book was put, violates the Projection Principle by failing to project the full lexical frame. Similarly, for ditransitive verbs like give, which require both a direct object (DO) and an indirect object (IO), the passive She was given a gift promotes the IO to subject position while preserving the DO, ensuring semantic and syntactic satisfaction of the subcategorization frame—unlike ill-formed variants that drop required arguments. These cases demonstrate how the principle constrains derivations to maintain lexical integrity.[25][26]The Projection Principle thus plays a crucial role in explaining the uniformity of subcategorization across syntactic transformations, interacting with related constraints like locality of selection to limit how projections occur within immediate structural domains.[27]
Locality of Selection
In generative syntax, the locality principle of subcategorization requires that a lexical head selects only its immediate complements, which function as sister nodes to the head in the phrase structure tree under X-bar theory, excluding selection across specifiers or adjuncts.[28] This constraint ensures that syntactic dependencies remain structurally proximate, maintaining the hierarchical organization of phrases without permitting non-local projections that could violate phrase structure integrity.[28]The theoretical foundation for this locality derives from Government and Binding theory, where subcategorization features are part of a head's lexical properties that must be satisfied within its local domain, and from the Minimalist Program, which reinforces it through mechanisms like Merge and phases to avoid long-distance selection violations, such as a verb subcategorizing a prepositional phrase embedded inside a subjectnoun phrase.[29] For example, the verb "put" subcategorizes for both a direct object noun phrase and a locative prepositional phrase as immediate sisters, rendering "*Sue put the book" ungrammatical due to the missing complement.[28] In contrast, control structures like "Susan hopes to sleep" allow local selection within the embedded clause via a silent PROsubject coindexed with the matrix subject, preserving locality without direct long-distance dependency.[28]Extensions of this principle appear in analyses of incorporation languages, where the Head Movement Constraint limits head-to-head adjunction to adjacent positions that govern the trace without intervening barriers, enabling incorporated nouns to satisfy subcategorization locally through syntactic incorporation into the verb.[30] In such cases, noun incorporation, as in polysynthetic languages, adheres to locality by treating the process as successive-cyclic head movement, ensuring the incorporated element remains within the selecting head's minimal domain.[31] The Projection Principle briefly ensures that this locality is preserved derivationally across levels of representation.[29]
Formal Representations
Subcategorization Frames
Subcategorization frames provide a formal mechanism in lexical entries to specify the syntactic categories and positions of complements that a word—most commonly a verb—requires or permits within a phrase structure. These frames are typically represented as sets of bracketed options indicating the verb's position and the categories of its arguments, such as [__ NP] for a transitive verb requiring a noun phrase object or [__ PP] for one selecting a prepositional phrase. This notation originates from generative grammar, where verbs are subcategorized based on the selectional properties of their immediate syntactic contexts, ensuring compatibility with the verb phrase (VP) structure. For instance, the verb "hit" has the frame V, [__ (NP)], where the parentheses denote an optional noun phrase complement, as in "She hit" or "She hit the ball."[2]Frames exhibit variability in obligatoriness and multiplicity, reflecting both the syntactic demands of individual verbs and their polysemous nature. Obligatory frames mandate specific complements for grammaticality, such as the transitive frame [+ __ NP] for "like," which requires an object as in "She likes apples," whereas optional frames allow but do not always require additional elements. Polysemous verbs often possess multiple frames corresponding to distinct senses; for example, "run" can appear in an intransitive frame [+ __] for the motion sense ("Birds run") or a transitive frame [+ __ NP] for the management sense ("She runs a company.")[2]Detailed examples illustrate how frames capture complex complementation patterns. The verb "persuade" typically selects the frame V, [__ NP to VP], specifying an object noun phrase and an infinitive clause complement, as in "She persuaded him to leave." Frames also accommodate syntactic alternations, such as particle shift, where a particle like "up" can precede or follow the object depending on the verb's properties; for instance, "pick up the book" alternates with "pick the book up," a pattern shared among certain transitive verbs grouped by their alternation behaviors. These representations draw from classifications that group verbs by shared frame options, enabling predictions about acceptable constructions. For "put," the frame is V, [__ NP PP[+loc]], requiring both an NP and a locative PP, as in "She put the book on the table."[2]In computational linguistics, subcategorization frames are integral to natural language processing (NLP) systems, particularly parsers, where they predict and validate complement structures during sentence analysis to improve accuracy in tasks like dependency parsing and semantic role labeling. Automated acquisition of these frames from corpora has enabled the construction of large-scale dictionaries, supporting applications in machine translation and text generation by modeling verb argument preferences empirically.
Valency
Valency refers to the degree or arity of a predicate, defined as the number of syntactic arguments it requires to form a complete clause. This concept, central to dependency grammar, was formalized by Lucien Tesnière in 1959, who analogized verbs to atoms whose "bonds" determine the complements needed for syntactic saturation. Predicates are classified by valency levels: avalent (zero arguments, as in impersonal verbs like "rain" with a dummy subject "it"), monovalent (one argument, typical of intransitive verbs), divalent (two arguments, as in transitive verbs), and higher valencies for predicates like ditransitives that demand additional complements.[32]Cross-linguistically, valency operations modulate this arity, with reduction mechanisms such as passivization (demoting the agent to an oblique) and anticausativization (eliminating the external causer) decreasing the number of core arguments, often to monovalency. In contrast, valency-increasing processes like applicativization add a beneficiary or other participant as a core argument, raising divalent predicates to trivalency or beyond. These patterns vary typologically; for instance, in ergative languages like Basque, valency interacts with case alignment, where monovalent predicates mark their single argument in the absolutive case, while divalent ones distinguish the transitive subject (ergative) from the object (absolutive), shaping auxiliary selection and argument realization.[33]In relation to subcategorization, valency offers a broader, numerical assessment of predicate complementation, whereas subcategorization details the specific syntactic categories involved; subcategorization frames thus implement valency at a finer-grained level. Theoretical debates highlight non-equivalence between the approaches: generative grammar emphasizes subcategorization for hierarchical phrase structure and selection, while dependency theories prioritize valency for linear head-dependent relations, viewing them as complementary rather than isomorphic levels of syntactic organization.Contemporary applications in construction grammar treat valency patterns as entrenched argument structure constructions that license participant roles and syntactic forms, often overriding lexical verb specifications to ensure clause-level coherence. This perspective integrates valency into usage-based models, where patterns emerge from conventionalized form-meaning pairings across languages.[34]