In machine translation, not all language pairs are created equal. While translating from a morphologically complex language like Finnish or Arabic into English is a well-trodden path, the reverse direction—translating from "morphologically poor" languages into "morphologically rich" ones—remains one of the field's most persistent hurdles.
This phenomenon is known as morphological asymmetry. As Philipp Koehn observed in his foundational work on the Europarl corpus, translation quality drops significantly when the target language has a complex system of inflections, cases, and genders. The difficulty isn't just about vocabulary; it is about the fundamental way information is packaged in different linguistic families.
The Information Deficit
The primary reason translating into a morphologically rich language (MRL) is harder than translating out of one is an information gap.
When translating from Czech to English, the source text is explicit. A Czech noun might carry a specific suffix indicating it is the instrument of an action (instrumental case). English doesn't have cases, so the model simply maps that suffix to the preposition "with." However, when Eleftherios Avramidis looked at the reverse (English to Czech), the problem became apparent: the English source "the cat" contains no explicit information about whether the cat is the subject, the object, or the instrument.
As noted in Improving Translation to Morphologically Rich Languages, the translation system must "invent" or infer grammatical features like gender, case, and number that are simply absent in the source text. This creates a one-to-many mapping problem, where a single English word might correspond to dozens of potential inflected forms in the target language.
Analysis is Easier than Synthesis
There is a mechanical difference between understanding a complex word and creating one. In agglutinative languages like Turkish or Uyghur, words are formed by "stacking" suffixes onto a root.
Jamshidbek Mirzakhalov and colleagues, in their Large-Scale Study of Machine Translation in Turkic Languages, found that while neural models can easily learn to "deconstruct" these suffix chains to understand a root, they struggle to "synthesize" them. Generating the precise sequence of suffixes required for grammatical correctness requires a level of linguistic accuracy that standard models often lack.
This is why Hassan Sajjad and Preslav Nakov found that unsupervised word segmentation works wonders for translating from Arabic dialects, but the reverse process of "desegmenting" or re-assembling those units into a fluid target word remains a major bottleneck.
The Lazy Decoder Problem
In modern Neural Machine Translation (NMT), there is a strange cognitive dissonance between the "brain" (the encoder) and the "mouth" (the decoder).
Research by Fahim Dalvi and Yonatan Belinkov in Understanding and Improving Morphological Learning in the Neural Machine Translation Decoder revealed that encoders are actually quite good at learning morphology. They internalize the rules of the language they are reading. However, decoders are "lazy." Because the decoder relies so heavily on the attention mechanism to look back at the source, it often fails to learn the internal grammatical consistency of the language it is supposed to be speaking.
This leads to "morphologically poor" output—sentences that might have the right words but the wrong endings, making the translation sound "broken" to a native speaker.
Strategies for Re-inflection
To solve this, researchers have moved beyond simple word-to-word mapping. Several successful strategies have emerged to help models handle target-side complexity:
- Source-Side Enrichment: Eleftherios Avramidis and Joachim Daiber have experimented with tagging the "poor" source language (English) with the grammatical markers it will need in the "rich" target language. By predicting the target's case or gender on the English syntax tree before translating, the model has a roadmap to follow.
- Two-Step Translation: Teams like LIMSI have used a "simplify then re-inflect" approach. They first translate English into a simplified, "lemmatized" version of a language like Czech, and then use a specialized second model to add the correct morphological endings (re-inflection).
- Asymmetrical Tokenization: While many systems use the same subword segmentation (like BPE) for both languages, Manish Shrivastava argues for Asymmetrical Byte Pair Encoding. By using different levels of granularity for the source and target, models can better capture the unique structural constraints of the MRL.
Ultimately, the goal is to move NMT from merely "mapping" words to truly "synthesizing" them, ensuring that the complexity of a language's grammar is treated as a feature to be mastered, rather than a noise to be smoothed over.