Lecture 13: Computational Morphology

Computational Morphology

Computational morphology studies how words are built from meaningful units and how natural language processing (NLP) systems can analyze or generate those forms.

Morphemes and Their Types

A morpheme is the smallest meaningful unit within a word. For example, dogs contains dog plus the plural marker -s. Unladylike can be analyzed as un-, lady, and -like.

• Free morphemes can stand alone as words, such as happy, house, and walk.
• Bound morphemes cannot stand alone and must attach to another form, such as un-, -s, -ly, and -ed.
• A stem is the central, meaning-bearing part of a word. Affixes attach to stems and may add grammatical information or help form new words. Stems are often free morphemes, while affixes are typically bound morphemes.
• Allomorphs are different forms that serve the same morphological function. Their choice depends on the word or its sounds. For example, English uses un- in unhappy, in- in incomprehensible, and im- in impossible. These forms generally cannot be substituted freely.

Morphemes can also be classified by function:

• Content morphemes contribute lexical or semantic meaning, as car does, or as -able does when it indicates ability.
• Functional morphemes primarily express grammatical information, such as the plural -s or the -s marking third-person singular verbs.

Affix Types

Affixes attach to stems in several ways:

• Prefix: placed before the stem, as in un-happy or pre-existing.
• Suffix: placed after the stem, as in walk-ing or quick-ly.
• Infix: inserted inside a stem. Infixes occur in some languages; an emphatic, nonstandard English example inserts bloody into absolutely.
• Circumfix: a combined affix with parts placed before and after the stem. One example comes from a Dutch plural form.

Prefixes and suffixes are common across languages. Infixes and circumfixes occur in particular languages and patterns.

Inflection and Derivation

Morphological changes can serve different purposes:

• Inflectional morphology adds grammatical information—such as tense, number, person, case, or gender—without changing a word’s part of speech. Examples include walk → walking and bring → brings/brought.
• Derivational morphology forms a new word and can change its part of speech. For example, drive → driver changes a verb into a noun. Logic can form logical, illogical, and logician.

Derivational patterns may appear regular but still have gaps. For instance, sincere → sincerity and curious → curiosity are possible, while the expected-looking form fiercity is not standard.

Other Word-Formation Processes

• Concatenation: joining morphemes, as in un + happy → unhappy or hope + less → hopeless.
• Sound and spelling changes at morpheme boundaries: combining forms may alter pronunciation or spelling. Examples include plural forms with different pronunciations and happy + -er → happier, where y changes to i.
• Reduplication: repeating all or part of a word to express a grammatical or lexical meaning. Examples from English and other languages include repeated forms used to convey meanings such as “examine carefully” or a future or perfect form.
• Suppletion: replacing a word with a form that has no obvious surface connection, as in go → went or good → better.
• Internal change: changing sounds within a word rather than adding an affix, as in sing → sang or man → men.
• Compounding: combining whole words, such as bittersweet, rainbow, and pickpocket. Compounds matter in machine translation because a word-for-word translation may not preserve the target language’s compound or phrase structure.
• Acronyms: forming a word from the initial letters of a phrase. The example intended in the lecture appears to be laser.
• Blending: combining parts of words, as in breakfast + lunch → brunch, smoke + fog → smog, and motor + hotel → motel.
• Clipping: shortening a longer word, as in laboratory → lab, advertisement → ad, examination → exam, bicycle → bike, and refrigerator → fridge.

Main NLP Tasks Involving Morphology

• Lemmatization: finding a word’s base or dictionary form. For example, saw could have the lemma see if it is a verb, or saw if it is a noun.
• Morphological analysis: identifying a word’s lemma together with its morphological features, such as part of speech and number or tense. A plural noun such as cats might be represented as the lemma cat with noun and plural features.
• Morphological tagging: choosing the correct analysis in context. In Peter saw her, context indicates that saw is a verb rather than a noun.
• Morphological segmentation: splitting a word into its component morphemes, as in an analysis of denationalization.
• Morphological generation: producing a word form from a lemma and requested grammatical features—for example, generating saw from see plus past-tense information. Generation is especially relevant to natural-language generation.

Why Morphology Matters in NLP

Morphological information is useful in several applications:

• Text-to-speech: spelling alone may not determine pronunciation. A system may need grammatical or contextual information to distinguish forms such as lead (the metal) and lead (the verb).
• Search and information retrieval: recognizing related forms can help handle word variation and organize searches.
• Machine translation: systems need to select appropriate target-language forms and account for compounds and grammatical differences between languages.
• Grammatical error correction: identifying a word’s grammatical features can help detect and correct incorrectly formed words.

Challenges in Automatic Morphological Analysis

Morphology is not always regular. Systems must cope with:

• Spelling alternations and irregular forms, rather than relying on one rule for every word.
• Ambiguity, as with saw, which can be a noun or a past-tense verb.
• Overgeneralization: a rule that works for one word may produce an invalid form when applied to another. A system must know, for example, which stems can take a particular suffix.
• Newly coined words: a system cannot rely only on forms already stored in a dictionary.

To analyze morphology, a system needs several kinds of knowledge:

1. A lexicon: which roots and words exist in the language.
2. Morphotactics: which morphemes can follow or combine with other morphemes, and in what order.
3. Constraints on affix use: which affixes are permitted with particular stems or word classes.
4. Spelling-change rules: how written forms change when morphemes combine.

A single, exhaustive dictionary of every possible form is not sufficient. English words have relatively few morphological variants on average, while languages such as Sanskrit can have many more. Languages also continually create new words, so their possible forms cannot all be listed in advance.

The lecture previews finite-state methods as a widely used approach to computational morphology, with more detail promised in the next lecture.

Speakers and Sources

• Speaker: One lecturer, whose name is not provided in the subtitles.
• Other source: Brief background music at the beginning.