Unit 7: Heads, Arguments, and Adjuncts

Words in a sentence are not related to each other in a flat, unstructured list. They form hierarchical groups — constituents — that function as units at intermediate levels between the word and the full sentence.

Evidence for constituent structure:

• Pronoun replacement: The large brown dog barked at the postman. → It barked at the postman. The whole sequence the large brown dog is replaced by it — evidence that it is a constituent.
• Movement: Constituents can move as units. At the postman, the dog barked. — the PP moves as a whole.
• Constituency questions: What did the dog bark at? — the postman. A constituent answers as a unit.
• Coordination: Only constituents of the same type can be coordinated: the large brown dog and the small grey cat — both NPs.

NLP relevance: Constituent identification (parsing) is a prerequisite for named entity recognition, coreference resolution, semantic role labelling, and question answering. A system that cannot identify constituent boundaries will make systematic errors in all of these downstream tasks.

The head is the most important word in a constituent. It determines both what the constituent contains (internal structure) and where it can appear (external distribution).

Tests for identifying the head:

1. Distribution test: The phrase can appear in the same positions as the head alone. A very good teacher arrived. → A teacher arrived. ✓ → A very good arrived. ✗ — teacher is the head.
2. Agreement/feature percolation: The morphological features of the head determine the features of the whole phrase. A big cat — cat is singular; the phrase takes singular agreement (a big cat is, not *a big cat are).
3. Obligatoriness: The head is the most obligatory element; a constituent cannot exist without its head.

Dependency grammar places the head-dependent relationship at the centre of all syntactic description. Every word (except the root) is a dependent of exactly one head; the entire sentence structure is a tree of head–dependent arcs. This formalism is the basis of most modern NLP parsers, including the Universal Dependencies annotation framework.

Every dependent of a head belongs to one of two fundamental types:

• Arguments — semantically required participants whose properties are specified in the head's lexical entry. They are part of the head's valency frame.
• Adjuncts — optional modifiers that are not listed in the head's lexical entry. They add information without being required.

Importantly, the same phrase can be an argument in one construction and an adjunct in another:

• She arrived in London. — in London is an adjunct: she arrived is a complete, grammatical sentence; the PP adds locative information optionally. Arrive is intransitive (valency 1); it does not select a location argument.
• She lives in London. — in London is an argument (or at least strongly expected): *She lives is awkward without a location because live in this sense strongly selects a locative complement. Whether this is a true argument or a very strong adjunct is debated — but it illustrates that the boundary is not always sharp.

NLP relevance: Distinguishing arguments from adjuncts determines how a semantic role labeller populates the argument slots of a predicate. Confusing an adjunct with an argument produces incorrect knowledge base entries.

Valency is the number of semantic arguments a head provides for — the number of participant slots it opens. It is stored lexically: knowing that a word is a verb tells you only that it selects a subject; the full valency frame specifies how many additional arguments are required and what semantic roles they bear.

English verb valency examples:

NLP systems performing semantic role labelling consult valency lexicons — VerbNet, FrameNet, PropBank — to know how many arguments each verb takes and what semantic roles (Agent, Patient, Theme, Goal, etc.) they bear. A system without valency information cannot reliably distinguish argument slots from adjunct positions.

Even obligatory arguments can be omitted under certain conditions — when they are recoverable from context, from discourse, or from pragmatic defaults.

English implicit arguments:

• She was eating. — the object (what she was eating) is implicit; eat is used intransitively. The participant slot is still there semantically; the argument is simply unexpressed.
• I've already read it. (diary entry) — subject omitted; acceptable in certain registers.

Pro-drop languages:

• Japanese — both subject and object arguments are freely dropped when contextually recoverable. Verb morphology does not mark agreement, but shared context makes the referents clear.
• Spanish — subject pronouns are routinely dropped (Habla español, "He/she speaks Spanish") because verb agreement suffixes identify person and number.
• Mandarin — arguments are frequently omitted in discourse when topics are established; no morphological agreement compensates.

NLP relevance: Unexpressed arguments must be recovered for semantic role labelling, coreference resolution, and entailment inference. In English, this is the task of implicit argument resolution or argument ellipsis resolution. In Japanese and Mandarin, zero-pronoun resolution is a major research challenge with significantly lower system performance than English coreference resolution.

Verbs are the prototypical argument-selecting heads, but they are not the only POS that selects arguments. Nouns, adjectives, and prepositions can also open argument slots.

Nouns selecting arguments:

• the arrival of the train — arrival is a deverbal noun derived from arrive; it inherits the verb's argument: the thing that arrives.
• a student of linguistics — student selects a PP complement specifying the field of study.
• the destruction of the city — the object argument of destroy is preserved in the nominalization.

Adjectives selecting arguments:

• proud of her — proud selects a PP argument specifying what the subject is proud of.
• similar to the original — similar selects a PP argument specifying what it is similar to.
• aware that prices had risen — aware selects a clausal argument.

Prepositions selecting arguments:

• on the table — on selects an NP argument.
• for the children — for selects an NP argument.
• despite the rain — despite selects an NP argument.

NLP relevance: Information extraction and semantic role labelling systems must handle argument structure across all POS heads, not just verbs. Nominalisation is particularly challenging: event nouns inherit verbal argument structure, and an NLP system that only looks at verb predicates will miss these argument relations.

Adjuncts have two diagnostic properties that distinguish them from arguments:

1. Not required: The head is grammatically and semantically complete without its adjuncts. Removing an adjunct from a grammatical sentence always yields a grammatical sentence (though a less informative one).
2. Iterable: Any number of adjuncts can be stacked onto the same head without creating ungrammaticality.

Demonstration:

She ran. ✓
She ran quickly. ✓
She ran quickly in the park. ✓
She ran quickly in the park on Tuesday. ✓
She ran quickly in the park on Tuesday despite the rain. ✓
She ran quickly in the park on Tuesday despite the rain in her new shoes. ✓

Contrast with arguments:

She put the book on the table. ✓
*She put on the table. ✗ (object argument missing)
*She put the book. ✗ (location argument of put missing in most dialects)

NLP relevance: The iterability diagnostic can be operationalised in information extraction pipelines: if removing an element still yields a grammatical predicate-argument structure, the element is likely an adjunct and should not be added to the argument slots in the extracted relation.

This is one of the most conceptually surprising insights about adjuncts. Syntactically, an adjunct is a dependent: it depends on its head. But semantically, the direction is reversed: the adjunct introduces a predicate, and the head (or its referent) is the argument of that predicate.

Example:

She ran quickly.

Syntactically: quickly modifies ran (dependent → head).
Semantically: quickly introduces the predicate QUICK(x); the running event is x. The meaning is: the running was quick.

Verbal adjuncts:

• She left before he arrived. — BEFORE(leaving-event, arriving-event).
• She ran because she was late. — CAUSE(being-late, running).

Nominal adjuncts:

• the dog in the garden. — IN(dog, garden): the PP introduces a predicate IN(x, garden) where x = the dog.
• the dog that bit the man. — the relative clause introduces BIT(dog, man) with the dog as the semantic argument of the embedded predicate.

NLP relevance: Accurate event extraction and knowledge graph construction require treating adjuncts as introducing additional propositions, not merely qualifying a single event representation. A system that represents she ran quickly in the park as a single event with properties attached will miss the fact that adjuncts contribute separate predicate-argument structures.

The obligatoriness test: remove the element in question from the sentence.

• If the result is ungrammatical, or if the meaning shifts from "less informative" to "expressing a different or incomplete proposition," the element is (likely) an argument.
• If the result is grammatical and still expresses a complete proposition (just a less informative one), the element is (likely) an adjunct.

Examples:

Caveats: The test is not always decisive. Some arguments can be left implicit (#55): She ate is grammatical without an object, even though eat can take one. And some very strong adjuncts (e.g. location with live) feel obligatory even though they are not strictly required by valency. The test is a diagnostic, not a proof.

The entailment test exploits the fact that adjuncts, as semantic predicates (#58), introduce entailments that are absent when the adjunct is removed.

Adjunct entailments:

• She ran in the park. entails She was in the park.
  She ran. does not entail She was in any specific place.
  → the entailment is introduced by the adjunct PP.
• She ran on Tuesday. entails The running occurred on Tuesday.
  She ran. does not entail this. → the temporal adjunct introduces the entailment.

Argument entailments:

• She gave him the book. entails He received the book. This entailment is part of the semantics of give — it is introduced by the verb, not an optional modifier.

The pattern: adjuncts introduce entailments that are not present in the adjunct-free sentence; arguments participate in entailments that are part of the verb's core semantics.

NLP relevance: Textual entailment systems — which must determine whether one sentence entails another — can use the argument/adjunct distinction to identify which propositions are contributed by optional modifiers versus core argument structure. This supports more accurate inference and question answering.

Adjuncts are not a single syntactic type — they can be realised at any level of structural complexity.

• Single word (adverb): She ran quickly.
• Phrase (PP): She ran in the park.
• Participial clause: She ran, breathing hard.
• Relative clause (nominal adjunct): the dog that bit the man
• Adverbial clause: She left before he arrived.
• Infinitival clause (purpose): She ran to catch the bus.

In all cases, the adjunct is structurally optional and semantically predicative (#58). The internal complexity of the adjunct does not change its status as an adjunct; a long relative clause is still an adjunct to its head noun.

NLP relevance: Parsers must identify adjuncts at all levels of complexity. Long clausal adjuncts are a major source of parsing errors because they introduce deeply embedded structures and multiple potential attachment points.

Nominal adjuncts modify noun phrases and contribute predications about the referent of the NP head.

• Adjective: the old dog — OLD(dog)
• Prepositional phrase: the dog in the garden — IN(dog, garden)
• Relative clause: the dog that bit the man — BIT(dog, man)
• Participial phrase: the dog running away — RUNNING-AWAY(dog)
• Genitive: the dog's owner — OWNS(owner, dog)
• Noun-noun compound modifier: the garden dog — located in the garden

NLP relevance: Nominal modification is central to information extraction. The attributes, properties, and relationships of named entities in text are typically expressed as nominal adjuncts. A system that fails to parse nominal adjuncts correctly will miss entity attributes and relationships. Relative clause attachment is a major source of parsing ambiguity (see Activity 4, Attachment Ambiguity tab).

Verbal adjuncts modify the verb phrase and contribute predications about the event or state described by the verb.

• Manner: She ran quickly. — QUICK(running)
• Frequency: She runs often. — FREQUENT(running)
• Temporal: She ran on Tuesday. — ON-TUESDAY(running)
• Locative: She ran in the park. — IN-PARK(running)
• Reason: She ran because she was late. — CAUSE(being-late, running)
• Purpose: She ran in order to catch the bus. — PURPOSE(running, catch-bus)
• Concession: She ran despite the rain. — DESPITE(running, rain)

NLP relevance: Verbal adjuncts carry the event properties that are essential for event extraction and temporal knowledge base construction. Timeline systems and question answering over events (when? where? why?) depend entirely on accurate verbal adjunct extraction and semantic interpretation.

Adjuncts are not limited to modifying NPs and VPs. They can also modify adjective phrases, adverb phrases, and even entire sentences.

• Adjective phrase (degree modifier): very fast — VERY modifies the AdjP fast.
• Adverb phrase (degree modifier): extremely quickly — EXTREMELY modifies the AdvP.
• Sentence adverb (speaker-oriented): Fortunately, she arrived on time. — the adverb scopes over the entire sentence; it expresses the speaker's evaluation of the proposition.
• Domain/frame adverb: In general, the system works. — scopes over the whole claim.
• Evidential adverb: Apparently, the meeting was cancelled. — indicates the speaker's information source.

Sentence-level adjuncts are semantically distinct from VP adjuncts: they evaluate the proposition expressed by the sentence, or characterise the speaker's epistemic state, rather than describing a property of the event.

NLP relevance: Sentence-level adjuncts are important for sentiment analysis (fortunately, unfortunately, disappointingly), hedging detection (apparently, seemingly, in general), and claim scope analysis. Systems that misanalyse these as VP adjuncts will misidentify their semantic contribution.

Adjuncts encode a rich variety of semantic relations between the adjunct predicate and its head argument. The principal semantic types are:

Each semantic type introduces different implications for event representation and knowledge base construction. Temporal and locative adjuncts ground events in space and time; causal adjuncts create causal links between propositions; concessive adjuncts signal that the expected causal or logical relationship is overridden.

Not every phrase can modify every other type. The syntactic category of a constituent determines where it can appear as a modifier:

• Adverbs (AdvP) modify verbs and adjectives, not nouns directly: She ran quickly ✓; *the quickly dog ✗
• Adjectives (AdjP) modify nouns directly, not verbs: the large dog ✓; *she ran large ✗ (as a manner modifier)
• Prepositional phrases (PP) can modify both NPs and VPs: the dog in the garden ✓ (NP); she ran in the park ✓ (VP)
• Relative clauses modify NPs, not VPs or AdjPs: the dog that barked ✓; *she ran that barked ✗

The modification potential is not a pragmatic convention but a structural property of the constituent's syntactic category. This is why the category (POS) assigned to a word or phrase directly constrains which attachment positions are available to it.

NLP relevance: POS tags constrain attachment possibilities in the parser. A POS tagger that misidentifies a category will produce downstream attachment errors. For example, if a word that is an adverb is mistagged as an adjective, the parser may attempt to attach it to a noun rather than a verb, producing an incorrect parse and downstream information extraction errors.

While many heads select NP arguments, some heads select clausal arguments — full embedded sentences (or sentence-like structures) as their complements.

Verbs with clausal arguments in English:

• Declarative complement (that-clause): She believed [that he was right]. — the entire bracketed clause is the object argument of believe.
• Interrogative complement (wh-clause): She knew [where he was].
• Polar interrogative complement: She asked [whether he would come].
• Infinitival complement: She wanted [him to leave].
• Counter-factual complement: She wished [that she had studied harder].

Cross-linguistically, clausal arguments are expressed differently: English uses complementisers (that, whether); some languages use subjunctive morphology on the embedded verb; others nominalise the embedded clause (the embedded proposition appears as a noun phrase). All of these are argument structures — the head selects a clausal complement — but the morphosyntactic realisation varies.

NLP relevance: Clausal argument extraction is substantially more complex than NP argument extraction and is a central challenge for relation extraction and knowledge base construction. Identifying the boundaries of a clausal complement, resolving the internal structure of the embedded clause, and determining the semantic relationship between the matrix verb and its clausal argument all require robust parsing and deep syntactic understanding. LLMs handle common clausal argument patterns reasonably well in English but struggle with deeply embedded clauses and with cross-lingual transfer.