Morphology and Word Formation › Morphological Typology: Why Language Choice Changes Your Pipeline

Languages differ systematically in how much work their morphology does, and this typology predicts, surprisingly well, which NLP techniques will struggle. The categories below form a spectrum rather than a set of boxes, and most languages sit somewhere in the middle, but the extremes clarify the trade-offs sharply.

Analytic (isolating) languages, like Mandarin Chinese and Vietnamese, pack little into each word. Grammatical information lives in word order and separate particles, so a large fraction of word forms are monomorphemic. Vocabulary is manageable, but syntax carries the load. Text-processing effort shifts accordingly: segmentation of continuous script becomes the hard problem, while inflection handling nearly vanishes.

Agglutinative languages, like Turkish, Finnish, Hungarian, and Swahili, string together long sequences of clearly delimited affixes, each contributing one meaning. Turkish evlerimizden decomposes cleanly as ev-ler-imiz-den: house, plural, our, from ("from our houses"). The combinatorics explode: Turkish verb paradigms contain millions of theoretically possible forms, so enumerating them in a vocabulary is hopeless, but rule-based decomposition works well because the affix boundaries are regular. Agglutinative languages are thus simultaneously the worst case for vocabularies and the best case for explicit morphological modeling.

Fusional languages, like Russian and Latin, compress multiple grammatical meanings into a single unsegmentable ending (Arabic layers this kind of fusion on top of the templatic root-and-pattern system described above). Russian книга, книги, книге, книгу encode case and number jointly in endings you cannot cleanly split into one-morpheme-one-meaning pieces. Paradigm tables, not string rules, traditionally describe these systems. Fusion defeats the natural decomposition strategy, so systems for fusional languages tend to lean on lookup tables and learned classifiers over whole endings.

Polysynthetic languages, like Inuktitut and Mohawk, push fusion to extremes, packing what English expresses with a full clause into a single word. At this end of the spectrum the very notion of a "word" as a processing unit comes into question, since one word may warrant an entire parse tree of its own.

The engineering lesson is direct. Recall the vocabulary problem and its subword solutions from Part V. For English, a 30,000-item vocabulary covers most running text comfortably. Apply the same budget to Turkish or Finnish and a large share of tokens fall back to tiny fragments, inflating sequence lengths and diluting each fragment's representation. Longer sequences mean more positions for attention to cover, more memory per document, and weaker signal per token. This is one reason multilingual models allocate vocabulary unequally across languages, and why morphologically aware tokenization remains an active research area.