Machine translation and challenges

Machine translation (MT) is the use of computer software to translate text or speech automatically from one natural language (source) to another (target), for example English to Nepali, while preserving the meaning.

Approaches: rule-based (direct, transfer, interlingua), statistical, and neural MT.

 Source text -> Analysis -> Transfer/Interlingua -> Generation -> Target text

Challenges of machine translation

1. Lexical ambiguity: one word, many meanings. "Bank" (river or money), "bat" (animal or cricket bat). The right sense must be chosen.
2. Structural (syntactic) ambiguity: "I saw the man with a telescope" has two readings. Languages also differ in word order: English is SVO (Ram eats rice), Nepali is SOV (Ram bhat khanchha).
3. Idioms and metaphors: "kick the bucket" cannot be translated word by word.
4. Pronoun and reference resolution: "The trophy does not fit in the suitcase because it is too big": what is "it"?
5. Morphological richness and gender/number/honorifics: Nepali has several honorific levels (ta, timi, tapai) with no English equivalent.
6. Missing words (lexical gaps): a word in one language may have no single equivalent in the other (cultural terms such as "tika", "dashain").
7. Context and pragmatics: the same sentence can mean different things in different situations; sarcasm, tone.
8. Low-resource languages: little parallel text for languages like Nepali limits statistical and neural approaches.
9. Named entities and transliteration: names must be recognised and transliterated, not translated.
10. Fluency and quality evaluation: the output must be both accurate and natural; automatic evaluation (BLEU) is imperfect.
11. Speech issues: accent, noise and disfluencies for spoken translation.