Example-based machine translation (EBMT) is a method of machine translation that relies on a bilingual corpus of parallel texts as its main knowledge base at run-time. It is essentially a translation by analogy, and can be viewed as an implementation of a case-based reasoning approach to Machine learning. Unlike rule-based systems that require deep linguistic analysis, EBMT translates by decomposing a source sentence into phrases, translating those phrases using examples from the corpus, and then composing the translated fragments into a target sentence.
The approach was first suggested by Makoto Nagao in 1984, who argued that it is especially suited for translation between two very different languages, such as English and Japanese. In such cases, a single sentence may map to several well-structured sentences in the target language, making deep linguistic analysis less useful. EBMT has since been explored as an alternative to statistical and rule-based methods, and it shares the use of bilingual corpora with statistical machine translation.
Translation by analogy
The core principle of EBMT is translation by analogy. This idea rejects the notion that human translators perform deep linguistic analysis of every sentence. Instead, it posits that translators decompose a sentence into phrases, translate each phrase by analogy to previously encountered translations, and then compose the translated phrases into a coherent whole. In EBMT, this principle is encoded through the example translations stored in the training corpus. The system retrieves analogous examples from the corpus to guide the translation of new input.
History and development
Makoto Nagao introduced EBMT in 1984, highlighting its potential for language pairs with very different structures, such as English and Japanese. His insight was that for such pairs, a sentence can often be rendered as several well-structured sentences in the target language, making deep linguistic analysis less effective than example-based methods. Over the following decades, EBMT was developed and refined, with contributions from researchers such as Michael Carl and Andy Way, who edited the 2003 volume Recent Advances in Example-Based Machine Translation. While EBMT has been largely overshadowed by Neural network-based approaches in recent years, its ideas about using parallel corpora and analogy continue to influence modern machine translation research.
How EBMT works
EBMT systems are trained on bilingual parallel corpora containing sentence pairs, such as:
- "How much is that red umbrella?" ↔ "Ano akai kasa wa ikura desu ka."
- "How much is that small camera?" ↔ "Ano chiisai kamera wa ikura desu ka."
From such minimal pairs, the system learns translation units: for instance, "red umbrella" corresponds to "akai kasa," and "small camera" corresponds to "chiisai kamera." These units can be composed to produce novel translations. For example, if the system has seen sentences like "President Kennedy was shot dead during the parade" and "The convict escaped on July 15th," it can translate "The convict was shot dead during the parade" by substituting the appropriate parts.
Phrasal verbs and sub-language phenomena
EBMT is particularly well-suited for sub-language phenomena like phrasal verbs, which have highly context-dependent meanings. Phrasal verbs in English consist of a verb followed by an adverb or preposition (the particle), and their meanings often cannot be derived from the individual words. For example, the phrasal verb "put on" can mean "switch on" (as in "Ram put on the lights") or "wear" (as in "Ram put on a cap"). In Hindustani, these meanings correspond to different verbs: "jalana" and "pahenna," respectively. EBMT can handle such ambiguities by retrieving examples that match the context, rather than relying on word-to-word translation.
Relationship to other approaches
EBMT is one of several data-driven approaches to machine translation. It differs from rule-based machine translation by avoiding deep linguistic analysis and instead using a corpus of examples. It also differs from statistical machine translation, which uses probabilistic models over large corpora, though both rely on bilingual data. In the 2010s, Deep learning and Transformer (architecture)-based models, such as those used in Large language models, largely superseded EBMT in practical applications. However, EBMT's emphasis on analogy and example retrieval remains relevant, and open-source platforms like Cunei have explored hybrid approaches that combine EBMT with statistical methods.