Machine translation methods

Machine translation methods can be divided into three approaches:  

Below are some notes from each of these methods.

The earliest machine translation systems were built with rule-based approaches. Rule-based machine translation engines require pre-coded linguistic rules.

https://machinetranslate.org/rule-based-machine-translation

Monolingual and bilingual dictionaries match input words to output words. In addition, rule-based machine translation engines require rules that present the structure of both the input language and the output language describing the grammatical structure of both languages.

Rule-based translating can be defined as an approach that includes a group of hardcoded linguistic rules that are used to analyse the grammatical input and create a representation into the output language structure. This approach requires knowledge of the source and target languages ​and the differences between them.

RBMT methods can be divided into three sub-logics: direct, transfer and interlingua approaches.

In the direct approach, the input text is translated word by word.

In the transfer approach, the input text is analysed sentence by sentence, after which the engine examines each word and sentence at a time and tries to understand its purpose. When the text input has been analysed and considered comprehensive, the engine translates the text using specific rules. Following the source sentence's structure, the system determines target words for each word and uses them to form target language sentences.

In the interlingua approach, the source text is transformed into semantic representation of the text, which will then form the basis for generating the target text.

Rule-based machine translation can be customised to a specific industry or topic. Challenges are, on the other hand, tightly connected with the nature of the method itself: all improvements in translation require either manually updating the built-in dictionaries or new hardcoded rules, which take a significant amount of time and human effort to create.

In the 2010s, the top method for machine translating was statistical machine translation.

https://machinetranslate.org/approaches

Statistical machine translator engines use substantial amounts of bilingual data to make as accurate translation as possible.

https://aws.amazon.com/what-is/machine-translation/

These engines require storing of text corpus (parallel corpora), which serve as the basis for the translation. Translation engine analyses the text looking for statistical relationships between the original texts and their existing human translations.

The SMT engines can be divided into four approaches:

All SMT engines follow the same logic. The input text is broken into sentences, after which the sentences are placed in their parallel counterparts, which are defined in the translation model. The language model then confirms that the translation is probable in the source language and translates the text.

SMT engines most commonly apply a phrase-based model, where words are translated in sentence-based sequences. The source text is segmented into sentences and compared to the targeted bilingual frame, after which statistical measurement is used to calculate the most likely target language segment based on the system's translation model and the collection of data of the target language. The translation model calculates the extent to which the source language word found in each sentence corresponds to target language words. It searches for this information from the text corpus input into the machine. The target language model calculates what is statistically the most probable translation.  

In the word-based model, the translation is created word-by-word. The syntax-based model translates syntactic units. The hierarchical sentence-based model combines a phrase-based model with syntax-based methods.

Challenges with statistical machine translation are related to creating the parallel data for each language pair – it is both costly and time-consuming as statistical machine translation requires massive parallel data. Specific errors in the translation are also difficult to fix. SMT is considered a more challenging method for language pairs with differences in word orders.

During the 2020s, the machine translation methodology has turned to neural machine translating.

https://machinetranslate.org/approaches

Neural machine translation (NMT) is the most used translation method in the commercial machine software industry, and it is widely seen as the most efficient and accurate method for machine translations.

https://aws.amazon.com/what-is/machine-translation/

Neural machine translation engines consider the entire sentences when creating the target sentence, where statistical translation engines calculate correspondence for individual words in sentences.

NMT engines use neural networks and teach themselves to recognise certain types of words and sentences. The process is time-consuming, but it is possible to train the translation process and thus make the processes more efficient. Neural machine translators are also improved as they are used – the more these machine translators are used, the better the translations become.