Unit 2: Morphology

A morpheme is a form–meaning pairing. The word cats contains two morphemes: cat (the animal) and -s (plurality). Neither can be divided further without destroying the form–meaning relationship.

Note the qualification usually: the definition of morpheme accommodates edge cases (non-contiguous roots, null forms, tonal contrasts) that the simple phone-sequence prototype does not cover — these are addressed in #8–#10.

NLP implication: NLP systems that treat cats and cat as completely unrelated tokens miss the systematic relationship encoded by morphology. Recognising morpheme boundaries is the first step toward lemmatisation, morphological tagging, and linguistically informed feature extraction.

In some languages, a single morpheme is realised as a discontinuous sequence of sounds distributed across a word. The classic example is the Arabic triliteral root system.

The root k-t-b carries the abstract meaning "write/writing". Vowels (and sometimes consonants) are interleaved to create distinct word forms:

NLP implication: morphological analysers designed for concatenative languages (where morphemes are simple linear sequences) cannot handle non-contiguous roots. Segmenting Arabic requires a fundamentally different architecture that recognises the root–pattern template system.

In some languages, morphological distinctions are encoded not in a sequence of phones but in suprasegmental features such as tone. Mandarin Chinese provides the canonical example: the same segmental sequence ma has four distinct lexical meanings depending on which of the four tones is used:

Here the tonal contour is itself a morpheme (or rather, a component of lexical form) — and it is not a sequence of phones.

NLP implication: text-based NLP for Mandarin typically uses characters, and tonal diacritics (pīnyīn romanisation) are rarely present in natural running text. The tonal distinction is therefore usually invisible to standard text-processing pipelines. Speech processing requires explicit tonal modelling.

A zero morpheme (or null morpheme) is a morpheme with no phonological form — it encodes grammatical information, but nothing is pronounced or written in its place.

English examples:

• sheep (singular) / sheep (plural) — the plural is marked by a zero morpheme, not by the addition of -s. The plural meaning is present; the phonological form is absent.
• I run — English first-person singular present tense has no agreement suffix (contrast third-person she runs). The absence of a suffix is the zero morpheme.

NLP implication: a system that learns morphological patterns from observed forms will not learn zero morphemes — there is nothing to observe. Yet the grammatical information they encode must still be recovered. Rule-based morphological analysers handle this explicitly; neural systems must learn it implicitly from contextual patterns.

The root is the morpheme that carries the central, irreducible lexical content of a word. All other morphemes attach to it (or are interleaved with it, in non-concatenative systems).

Examples: cat in cat / cats / catfish / catlike; Turkish git- (go) in gidiyorum / gitmedim / gitmek.

NLP implication: identifying the root of a word is the goal of stemming — reducing all surface forms of a word to a common representation. Stemming is used in information retrieval to improve recall: a search for run should also retrieve documents containing runs and running.

Derivational morphology creates new lexical items from existing ones. It frequently changes the grammatical category of the base, as well as its meaning:

NLP implication: part-of-speech tagging and lemmatisation must model derivational morphology. A word-form lookup approach will fail for novel derivationally complex forms, which can in principle be created productively (a speaker can coin unhappifiable and be understood). Morphological analysers must generalise across productive derivational patterns.

Even when both the root and the derivational affix are known, the meaning of the derived word may not be compositionally predictable:

• blackbird ≠ a bird that is black (it is a specific species, Turdus merula)
• hotdog ≠ a dog that is hot (it is a type of food)
• understanding ≠ the act of standing under something

These forms are lexicalised: their meaning has diverged from what strict compositional analysis would predict.

NLP implication: a purely compositional morphological analyser will generate wrong glosses for lexicalised compounds and derivatives. NLP systems need access to lexical entries for such forms, rather than relying on rule-based decomposition alone.

Inflectional morphology adds features such as tense, aspect, number, gender, case, person, or mood to an existing word, without creating a new lexical item. The resulting forms all share the same lemma and basic meaning.

English examples: walk / walks / walked / walking — all forms of the verb walk. Spanish: hablo / hablas / habla / hablamos / habláis / hablan — all forms of hablar (to speak), inflected for person and number.

Inflectional features are syntactically required by the grammar: a finite English verb must be inflected for tense; a Spanish verb must agree with its subject in person and number.

NLP implication: morphological tagging assigns inflectional feature values to word tokens. This is a prerequisite for syntactic parsing in morphologically rich languages, where inflectional information determines grammatical role. Lemmatisation must strip inflectional affixes correctly to group related forms under a single representation.

A single morpheme form can encode multiple, distinct grammatical meanings — it is ambiguous when the context is not sufficient to distinguish between them, and underspecified when it genuinely does not differentiate between possible values.

The English suffix -s is famously ambiguous across three distinct morphemes:

• cats — plural noun marker
• she walks — third person singular present tense
• John's cat — genitive/possessive marker

Mandarin Chinese provides a different type of underspecification: there is no tense morphology, so verbs do not indicate whether an event is past, present, or future. The temporal interpretation must be inferred from context.

NLP implication: morphological ambiguity must be resolved before many downstream tasks can proceed correctly. Part-of-speech disambiguation and morphological tagging perform this resolution using contextual information. Underspecification in the source language can cause problems for machine translation: translating from Mandarin to English requires choosing a tense for every verb — information that is simply absent from the source.

In English, the word seems like a natural unit: it is a string of characters between spaces in the written language. But this orthographic convention does not reflect a universal linguistic reality.

• Turkish gidiyorum ("I am going") encodes subject, tense, aspect, and the verb in a single orthographic word — it could reasonably be analysed as a phrase.
• French je t'aime ("I love you") involves a clitic t' that some analyses treat as a separate word, others as a prefix on aime.
• Mandarin Chinese is written with no whitespace between characters; "word" boundaries must be inferred.

NLP implication: word tokenisation — the first step in virtually every NLP pipeline — is a non-trivial task in many languages. The unit of analysis (character, morpheme, word, phrase) must be chosen deliberately for each language and task.

Within a word, morpheme order is typically fixed by the grammar: you cannot permute the affixes on a Turkish verb without producing an ungrammatical or meaningless form. Between words in a sentence, ordering constraints are often more flexible (particularly in case-marking languages where role is indicated by the morpheme, not the position).

English example: adjective order within a noun phrase is relatively fixed (big old red Italian sports car — a specific ordering must be respected), yet adverbials can move more freely within a sentence. At the morpheme level, the English verbal suffix sequence is entirely fixed: - VOICE - TENSE - NUMBER is not permutable.

NLP implication: a language model that learns word-level co-occurrence patterns does not automatically learn morpheme-level ordering constraints within words. Sub-word models (BPE, WordPiece) approximate morpheme boundaries but do not enforce morphotactic constraints.

Over time, free words can become bound morphemes — a process called grammaticalisation. What starts as an independent word becomes phonologically reduced and grammatically dependent.

Examples:

• English -ly (adverb suffix in quickly, happily) derives from Old English -līce, itself from lic (meaning "body, form"). A free noun has become a bound morpheme over several centuries.
• English going to → gonna: a phrase is in the process of grammaticalising into a tense/aspect marker. In spoken English, this is already functionally a bound morpheme in informal registers.

NLP implication: NLP tools trained on contemporary data may behave differently on historical texts, where a form that is now a bound morpheme was still a free word. Historical NLP and digital humanities applications must account for diachronic morphological change.

A clitic occupies an intermediate position between a free morpheme (an independent word) and a bound morpheme (an affix). It has its own syntactic function and meaning, but it cannot bear stress and must attach phonologically to an adjacent word — its host.

English examples:

• 've in I've — a clitic form of have
• 's in John's — a clitic marking genitive (but also a form of is/has)
• n't in don't, can't — a clitic form of not

Spanish object clitics are another example: lo veo ("I see him/it") and véalo ("see him/it!") — the clitic lo attaches in different positions in declarative vs imperative contexts.

NLP implication: clitics challenge tokenisation. The form don't must be split into do + n't for correct morphological and syntactic analysis, but a simple whitespace tokeniser will treat it as a single token. Clitic handling is a standard preprocessing step in many NLP pipelines.

The morpheme-per-word ratio is a key typological parameter that directly determines the scale of vocabulary in a language's text:

Turkish evlerinizden = ev-ler-iniz-den: house-PL-2PL.POSS-ABL "from your houses"

NLP implication: a fixed-vocabulary model trained on Turkish text will encounter vast numbers of unseen word forms — because any verb or noun root can in principle combine with dozens of suffixes, producing thousands of distinct forms. Subword models (BPE, WordPiece) mitigate this by splitting unknown words into known subword units, but they approximate — not replicate — morphological structure.

Cross-linguistically, suffixing is far more common than prefixing as the dominant mode of affixation — this is one of the most robust typological universals. Most languages that use affixes at all attach them after the root.

Examples:

• Suffixing majority: Turkish, Finnish, Japanese, Swahili (for many categories)
• Prefixing tendency: many Bantu languages (e.g. Swahili verb prefixes for subject agreement: a-na-soma = 3SG-PRES-read, "he/she reads")
• Mixed: English uses both (prefix un-, suffix -ed); Arabic uses both prefixes and suffixes alongside its non-concatenative root system

NLP implication: morphological analysers must be configured for the directionality of the target language. A suffix-stripping stemmer designed for English will not work for a Bantu language where grammatical information is primarily prefixal.

The clarity of morpheme boundaries varies substantially across morphological types:

• Agglutinative (clear boundaries): Turkish ev-ler-im-den (house-PL-1SG.POSS-ABL) — each suffix encodes one feature; boundaries are unambiguous.
• Fusional (fused forms): Latin -at in amat ("he/she loves") — this single suffix encodes third person, singular, present tense, active voice, and indicative mood simultaneously. There is no way to separate these features into distinct segments.
• Non-concatenative (no linear boundaries): Arabic roots interleaved with vowel patterns — the "boundary" between root and pattern does not exist in linear sequence.

NLP implication: morpheme segmentation algorithms perform well on agglutinative languages but poorly on fusional and non-concatenative ones. BPE and other subword algorithms learn from data and therefore approximate the boundaries that are actually present — they work best when those boundaries are consistent and learnable.