Unit 4: Morphosyntax

Morphosyntax is the interface between morphology and syntax: the features encoded in word forms determine what those words can combine with. The word's internal structure (its morphemes) constrains its external syntactic behaviour (what it combines with).

A clear English example: the past tense morpheme -ed affects combinatoric potential in the auxiliary system. Had walked is grammatical; *is walked is not (unless passivised). The past tense form of the verb determines which auxiliaries can combine with it.

In case-marking languages, this is even more explicit. A German noun in the nominative case signals that it is the subject; a noun in the accusative signals the object. The case morpheme on the noun determines which syntactic position it fills — and which verbs or prepositions it can combine with.

NLP implication: a system that ignores morphosyntax cannot correctly determine what words can combine with. This affects every downstream task: parsing, semantic role labelling, machine translation, and grammatical error correction all depend on knowing the combinatoric potential of each word form.

TAM features encode temporal and modal information about the situation described. Languages differ substantially in how they encode these categories:

• English tense: walk / walked / will walk. Past and present are encoded morphologically; future is periphrastic (auxiliary will).
• Turkish TAM: Turkish encodes both tense and evidentiality as obligatory verb morphology. Gitti (he went — direct witness past) vs gitmiş (he went — hearsay/inference past). The morphology encodes whether the speaker witnessed the event or is reporting it secondhand.
• Japanese aspect: aspect particles mark the temporal structure of the event: tabete iru (is eating, ongoing) vs tabete ita (was eating, past ongoing).

NLP implication: MT between languages with different TAM systems is not a word-for-word problem. A system must transfer temporal and modal information across radically different grammatical frameworks — inferring what the source language leaves implicit when the target language requires it to be explicit.

Nouns and noun phrases carry features about the entities they refer to:

• Number: English cat / cats (singular/plural). Arabic has three-way number: singular, dual, and plural — kitāb (one book), kitābān (two books), kutub (books).
• Gender: French le chat (M, the cat) vs la chatte (F, the female cat). German has three genders: der (M), die (F), das (N). Crucially, grammatical gender is not biological sex — French la sentinelle (sentry) is grammatically feminine regardless of the biological sex of the person.
• Person: pronouns encode person (I / you / he / she / we / they). Some languages also encode person on nouns via possessive morphology.

NLP implication: grammatical gender is an arbitrary morphosyntactic feature, not a semantic one. NLP systems that conflate grammatical gender with biological sex produce systematic errors. MT systems must track noun gender to generate correct agreement on adjectives and verbs in the target language.

Case marking indicates the grammatical function of a noun phrase in the sentence — whether it is the subject (nominative), object (accusative), indirect object (dative), possessor (genitive), and so on. Case-marking languages include:

• German: four cases — Nominativ, Akkusativ, Dativ, Genitiv. The definite article changes form: der Hund (the dog, NOM) vs den Hund (ACC) vs dem Hund (DAT) vs des Hundes (GEN).
• Japanese: case is marked by postpositional particles: -ga (nominative subject), -wo (accusative object), -ni (dative/locative), -no (genitive). A parser reading Japanese reads the case particle — not the word order — to determine grammatical function.
• Turkish: six cases: nominative (-Ø), accusative (-i), dative (-e), locative (-de), ablative (-den), genitive (-in). Case suffixes undergo vowel harmony with the root.

NLP implication: in case-marking languages, word order is freer than in English because grammatical function is marked morphologically rather than positionally. Parsers for these languages must read from case markers, not from position; parsers trained on English-like word order constraints will fail.

Many languages encode negation as a morphological feature on the verb or as a derivational prefix on adjectives:

• English derivational negation: un- (unhappy, unkind), -less (harmless, careless), in-/im-/ir- (inefficient, impossible, irregular). These are attached to adjectives and some nouns.
• Turkish verbal negation: the negative morpheme -me/-ma is inserted into the verb complex between the root and TAM suffixes: git-me-di (did not go). Negation is morphologically inside the verb, not a separate syntactic word.
• Japanese verbal negation: -nai is a negative suffix: ika-nai (does not go), tabe-nai (does not eat).
• Swahili: the negative prefix ha- attaches to the verb: anakwenda (he/she is going) vs hakwendi (he/she is not going).

NLP implication: negation scope is one of the hardest problems in NLP sentiment analysis and natural language inference. A system that processes words without tracking morphological negation markers will systematically misjudge polarity. In morphologically complex languages, negation is fused with the verb — the system must parse the morphological structure to detect it.

Evidentiality is the grammatically obligatory marking of the speaker's source of information — whether they witnessed the event directly, inferred it, or heard it from another person. In languages with grammatical evidentiality, the verb morphology encodes this distinction on every utterance.

• Turkish: -di marks direct witness past; -miş marks hearsay or inference past. Ahmet geldi (Ahmet came — I saw it) vs Ahmet gelmiş (Ahmet came — I was told / I infer from evidence).
• Bulgarian: has a grammatical reportative form used for events the speaker did not witness.
• English: has no grammatical evidentiality. Source of information is expressed lexically if at all: apparently, reportedly, I heard that. A speaker can assert any proposition as fact, regardless of evidence.

NLP implication: LLMs generate statements with equal linguistic confidence whether the proposition is well-established, inferred, or fabricated. This is the grammatical encoding of the problem of hallucination: in a language with evidentiality, every assertion would be obligatorily marked for its epistemic status. LLMs trained on English have no such grammatical obligation — and this structural fact is part of what enables confident-sounding confabulation.

Definiteness encodes whether the referent of a noun phrase is assumed to be identifiable by the hearer (definite) or not (indefinite). Languages encode it in different ways:

• English: a free article system — a/an (indefinite) vs the (definite). Articles are separate words.
• Romanian: definiteness is a suffix: om (a man) → omul (the man); casă (a house) → casa (the house). The definite article is morphologically integrated into the noun.
• Arabic: definiteness is a prefix: al-bayt (the house) vs bayt (a house). The prefix triggers assimilation with certain initial consonants.
• Languages without articles: Japanese, Mandarin, and Russian have no grammatical articles; definiteness is inferred from context and word order.

NLP implication: article generation is a known difficulty for NLP systems. Speakers of article-free languages learning English frequently omit or misplace articles; NLP grammatical error correction must model definiteness to correct these errors. MT from article-free languages into English must generate articles from contextual inference.

Honorifics encode the social relationship between the speaker, the hearer, and/or the referent — respect, familiarity, formality, status. In some languages this is obligatory morphological marking on the verb:

• Japanese keigo (polite/formal speech): iku (go, plain form) vs irassharu (go, honorific form used when referring to the actions of a person the speaker defers to). Multiple levels of keigo encode distinct social relationships. Using the wrong level is a serious social error.
• Korean: verb endings encode the speaker-hearer relationship across six speech levels. The choice of ending signals the relative status of interlocutors.
• Thai: politeness particles (khráp for male speakers, khâ for female speakers) are appended to utterances as obligatory politeness markers.

NLP implication: NLP systems that generate text in Japanese or Korean must track the social register of the interaction and select the appropriate honorific forms throughout the response. Mistaken honorific use is socially significant and perceived as insulting or inappropriate — not merely grammatically wrong. Most NLP systems default to a single register, missing the full communicative function of the language.

Possessive relationships — ownership, belonging, association — can be encoded morphologically on the noun that is possessed:

• Turkish possessive suffixes: ev (house) → evim (my house) / evin (your house, sg.) / evi (his/her house) / evimiz (our house) / eviniz (your house, pl.) / evleri (their house). The possessor's person and number are encoded as a suffix on the possessed noun — no separate possessive pronoun is needed.
• Hungarian: a similar system of possessive suffixes on the possessed noun.
• English: the possessive clitic -'s attaches to the possessor noun phrase: the dog's bone, the teacher's pen.

NLP implication: information extraction systems looking for ownership and relational information must recognise morphological possessive marking. In Turkish, the possessive suffix is fused with the noun — a morphological analyser is required to segment and interpret the possessive relationship encoded in the surface form.

The categories in concepts #29–#36 are not an exhaustive list. Languages encode a wide range of additional grammatical notions morphologically, many of which have no equivalent category in English:

• Directional morphemes: some Indigenous American languages mark whether motion is toward or away from the speaker as obligatory verb morphology. English expresses this lexically (come vs go), but the grammatical obligation is absent.
• Switch reference: found in several Australian Aboriginal and Indigenous American languages, this morphological system marks whether the subject of the next clause is the same as or different from the subject of the current clause. It is an elaborate morphological anaphora-tracking system.
• Mirativity: some languages mark grammatically whether the speaker considers the information new or surprising.
• Telicity: the distinction between events with a natural endpoint (telic: build a house) and events without one (atelic: run) can be morphologically marked in some languages.

NLP implication: English morphosyntax is not a universal template. NLP systems built around English grammatical categories will lack the representations needed for these language-specific features. This is a fundamental obstacle to genuinely multilingual NLP — the categories themselves differ, not just the forms.

Agreement is not the duplication of information — it is the marking of a feature on a controller (the element that semantically "owns" the feature) and the copying of that feature value onto one or more targets. The feature belongs to the controller; the targets merely carry agreement morphology.

A classic example from Latin: bonus vir (a good man, nominative masculine singular). The noun vir (man) is the controller; it is nominative, masculine, and singular because of its grammatical role and inherent gender. The adjective bonus agrees with it — it carries nominative masculine singular endings not because it is nominative, masculine, and singular in its own right, but because the noun it modifies is. The accusative form of the same phrase is bonum virum: the case marker appears on both noun and adjective, but it belongs to the noun.

NLP implication: a parser must identify agreement relationships to correctly assign grammatical structure. Tracking which element is controller and which is target requires knowledge of morphosyntactic feature hierarchies — which elements own features, and which merely copy them.

Verb–argument agreement is found in the vast majority of the world's languages. The typical pattern is verb–subject agreement, but many languages extend this:

• Spanish verb–subject agreement: como (I eat) / comes (you eat, sg.) / come (he/she eats) / comemos (we eat) / coméis (you eat, pl.) / comen (they eat). The verb encodes person and number; the subject pronoun is often omitted because the verb form already identifies the subject.
• Basque: verbs agree with both the subject AND the object — the verb form encodes the person and number of all core arguments simultaneously. This is cross-linguistic evidence that agreement is not restricted to the subject.
• Swahili: both subject and object agreement are marked as prefixes on the verb, from a system of noun classes (not just person/number): ni-na-ki-soma (I am reading it — subject prefix ni-, tense na-, object prefix ki- for a class-7 noun).

NLP implication: verb–argument agreement must be tracked across the full span of the sentence, not just at the immediately adjacent subject. In pro-drop languages (like Spanish) where the subject pronoun is frequently omitted, the parser must recover the subject from the verb's agreement morphology — not from an overt NP.

Within the noun phrase, agreement cascades from the noun (as controller) to all dependents — determiners, adjectives, numerals, and demonstratives:

• French noun-adjective agreement: le grand chat (M SG, the big cat) vs la grande chatte (F SG, the big female cat) vs les grands chats (M PL). The adjective grand takes four forms depending on gender and number: grand / grande / grands / grandes.
• Russian: adjectives agree with nouns in number, gender, and case, yielding a paradigm of approximately 24 distinct agreement forms (6 cases × 2 numbers × 2 paradigm types, with gender distinctions in the singular). Russian NLP generation must select the correct form from this paradigm for every adjective in the text.

NLP implication: AI-generated French text frequently contains agreement errors — an adjective in the wrong gender or number form. This is a documented failure mode of large language models. Because LLMs learn statistics over surface forms rather than tracking agreement features structurally, they misapply agreement in less common contexts (unusual noun gender, long distance between noun and adjective, complex NP structure).

Agreement tracks grammatical and conceptual features that may not be visible on the controller's surface form. This is called agreement with a covert feature or notional agreement:

• English collective nouns: The committee has decided (singular — treating the group as a unit) vs The committee have decided (British English — treating the group as a collection of individuals). The noun committee is morphologically singular, but the verb can agree with a conceptual plural interpretation. The agreement target (the verb) is sensitive to an abstract feature, not the overt form of the controller.
• Gender agreement with epicene nouns: in Spanish, el/la estudiante (the student — common gender noun). The gender of the referring pronoun and of agreement on adjectives must track the actual individual's social gender, which is not overtly marked on the noun.

NLP implication: agreement cannot be resolved by surface pattern-matching alone. Tracking covert features requires richer linguistic representations — including conceptual and social information that is not recoverable from the text surface. This is a deep challenge for NLP systems that rely only on observable token sequences.

Every language marks some grammatical information morphologically and expresses other information lexically, syntactically, or not at all. The selection differs dramatically across languages:

• Turkish marks evidentiality morphologically; English does not. A Turkish speaker must grammatically encode the source of their information on every past tense verb; an English speaker need not.
• Japanese marks politeness register morphologically; English does not.
• English marks tense morphologically; Mandarin Chinese does not. In Mandarin, temporal information is conveyed through time adverbials and aspect particles, not tense morphemes.
• German and Japanese mark case morphologically; English does so only for pronouns.

NLP implication: when translating from a language that morphologically marks feature X into one that does not (or vice versa), the translation system must either infer and generate information that was absent in the source, or accept that information present in the source must be lost or demoted. Cross-linguistic NLP cannot assume that the source and target languages share a set of morphological categories.

Even when two languages both mark the same category, they may draw different numbers of distinctions within it:

• Number: English marks singular vs plural. Arabic marks singular / dual / plural — requiring a separate form for exactly two of something. Slovenian also has a dual number. A MT system from Slovenian to English must collapse the dual into the plural and lose the two-ness information.
• Gender: English has no grammatical gender (only pronoun gender). French has two genders (masculine/feminine). German has three (masculine/feminine/neuter). Some languages have ten or more noun classes (sometimes called genders), each requiring its own agreement paradigm.
• Person: English distinguishes first/second/third person. Some languages also distinguish inclusive vs exclusive first person plural — we including the hearer vs we excluding the hearer.

NLP implication: the finer the distinctions within a morphological category, the larger the inflectional paradigm and the more data an NLP system needs to learn each form reliably. Languages with 10+ noun classes, 6+ cases, and complex TAM systems — such as Finnish or Georgian — require substantially more annotated training data than English for the same task performance. This contributes to the performance gap between high-resource and low-resource languages in NLP.