Translation divergences, first systematically classified by Bonnie Dorr (1993), refer to the structural mismatches between languages that make literal translation inadequate. When translating between English and Spanish, for instance, a simple adjective-noun phrase may require reordering ("white house" becomes "casa blanca"), or an English verb may correspond to a verb-plus-preposition construction in the target language. These divergences are not merely surface phenomena but reflect deep typological differences in how languages encode meaning, and they pose fundamental challenges for all approaches to machine translation.
Dorr's Classification of Divergences
2. Promotional: modifier becomes head ("John swam across" → "John cruzó nadando")
3. Demotional: head becomes modifier (reverse of promotional)
4. Structural: differing phrase structure requirements
5. Categorial: part of speech changes ("be hungry" → "tener hambre")
6. Conflational: single word maps to multiple words
7. Lexical: different lexicalization patterns
Thematic divergences arise when the mapping between thematic roles (agent, patient, experiencer) and syntactic positions differs across languages. In English, "I like music" has the experiencer as subject, while in Spanish "me gusta la música" places the experiencer as an indirect object. Categorial divergences involve changes in part of speech: English expresses hunger with an adjective ("be hungry") while Spanish uses a noun construction ("tener hambre," literally "to have hunger"). These divergences are pervasive and interact in complex ways.
Impact on Machine Translation
Rule-based MT systems addressed divergences through explicit transfer rules that mapped between source and target representations at various levels of abstraction. Statistical MT systems handled many divergences implicitly through phrase tables that captured multi-word correspondences and reordering models that learned typical permutation patterns from parallel data. However, long-distance reordering — such as the verb-final order of German subordinate clauses — remained challenging for phrase-based systems.
Neural machine translation systems, particularly those based on the Transformer architecture, handle many translation divergences more gracefully than their statistical predecessors. The attention mechanism allows the model to learn arbitrary alignment patterns between source and target positions, naturally accommodating reordering divergences. However, even NMT systems can struggle with certain systematic divergences, particularly when training data is scarce or when divergences interact with rare lexical items.
Typological Considerations
The severity and nature of translation divergences depend on the typological distance between language pairs. Closely related languages (e.g., Spanish and Portuguese) exhibit relatively few divergences, while typologically distant pairs (e.g., English and Japanese) present pervasive mismatches in word order, morphological structure, and information packaging. The World Atlas of Language Structures (WALS) documents the cross-linguistic variation that underlies these divergences, and computational typology has emerged as a field that seeks to incorporate this knowledge into translation systems.
Understanding translation divergences remains important even in the neural era because they explain systematic error patterns, inform the design of evaluation metrics, and guide the development of linguistically informed architectures. Languages with rich morphology, free word order, or pro-drop properties continue to present challenges that purely data-driven approaches do not fully resolve without sufficient parallel data.