Types of machine translation

2.1. Direct machine translation systems

Boitet et al. say machine translation types were categorized differently in different time periods of research. At first, in the 1970s, machine translation methodologies were categorised according to the technology used in the three translation steps listed above (Boitet et al. 2009). In the period after the Second World War, researchers, especially computer scientists, naively delved right into the world of machine translation out of enthusiasm about artificial intelligence.

Using cryptography, they created the first direct machine translation systems that simply substituted words with their dictionary-based equivalents instead of completing the phases of decoding and encoding. The system completely skipped the transcoding phase. This methodology was obviously unsuccessful but research and development at the time considered the role of linguistics to be superfluous and therefore did not even resort to applying linguistics as a science. Theories on formal language had yet to be written and scientists did not even consider the possibility of having strict and formal rules pertaining to language that could help computers create acceptable translations, while computer scientists had no adequate knowledge about linguistics (Somers 2000).

This is hardly surprising given that the ALPAC report issued in 1966 was very sceptical of research done in machine translation up to that point, and although it emphasised the need for basic research in computational linguistics, it did not express great expectations about the usefulness of machine translation in the near future. The prediction made in the ALPAC report has since been contradicted as, for example, the translation system METEO was used to translate weather forecasts as early as 1977 in Canada (Thouin 1981). In the cases of certain language pairs, machine-translated texts were also used in practice, albeit with post-editing.

2.2. Indirect machine translation systems

Although research on machine translation at this point came to a halt in the United States and the UK, mostly due to a lack of funding, research continued throughout the rest of Europe. It was not until the 1960s that V. Yngve (Boitet 1998) suggested that machine translation should be broken up into three stages instead of using direct methods. Yngve said the process should be broken up into source language analysis, bilingual transfer and target language synthesis, which corresponds to the three-step approach to the translation process. The representation resulting from analysis had to be an intermediary representation not relying on the target language. This method is known as the transfer approach.

In 1951, Bar-Hillel suggested the use of an interlingua for intermediary representation, for example Esperanto or an abstract intermediary language so that synthesis could then be carried out into any other language. Schubert also recommends the use of Esperanto as an interlingua (Schubert 1992: 79). This is known as the interlingua approach, and although its popularity peaked in the 80s and the 90s, it is still used today in experiments (e.g. Dave et al. 2001).

The transfer and interlingua approaches are the second generation of machine translation and are also known as indirect machine translation systems. These approaches are illustrated below in the Vauquois machine translation triangle (Vauquois 1976: 131):

Figure 1

The first and second generation of machine translation

2.3. Knowledge-based machine translation systems

The third generation of machine translation methods are the ones that apply knowledge about the world. These are known as the knowledge-based machine translation systems, but they did not gain ground at all.

After 2000, emphasis gradually shifted to Rule-Based Machine Translation (RBMT) and Statistics-Based Machine Translation (SMT). In the 1980s, machine translation research started using corpus-based methods, which apply statistics-based and example-statistics-based machine translation (EBMT) approaches. Both SMT and EBMT use bilingual parallel corpora in research.

Statistics-based methods are mainly based on the following model (Hutchins 2005b): words and sentences are aligned to create a bilingual parallel corpora to create a translation model (based on frequency of co-occurrence) and a linguistic model (based on probability  for certain word sequences). This is followed by selecting the most probable target language equivalent for each source language word and the most probable sequence of these words in a sentence. The basic units of translation are words. The translation model determines the probability of a target language word being the equivalent of a given source language word, while the linguistic model determines the most probable acceptable sequence of a given set of target language words in the target language.

Example-based systems are similar (Hutchins 2005b), but in this case the units of translation are phrases. In the analysis phase, the system divides the input sentence into segments which are then aligned with the correct source language segments in the database. These segments are patterns that may contain variables. In the pre-synthesis stage, the source language segments are aligned with the target language segments in the database and templates are derived. In the synthesis stage the derived target language segments are transformed, combined and the output sentences are formed. Example-based machine translation systems integrate several different methodologies and techniques from other types of systems (RBMT, SMT, translation memories). Any system can combine the various methodologies.

Rule-based systems are based on linguistic information about source and target languages retrieved from rules and grammars. The rules are used to analyse the source language text and then to generate the target language text based on that analysis. RBMT can also make use of dictionaries. The systems use the rules to generate an intermediary representation and then a target language text.

2.4. The MetaMorpho translation system

The highest quality translation system in terms of English-Hungarian translation is the pattern-based MetaMorpho (www.webforditas.hu) (Prószéky & Tihanyi 2002). The system aims to find the optimal solution between example-based and rule-based systems. The system’s ‘knowledge’ is comprised of patterns that possess certain attributes. If these attributes have specific values, these patterns are called examples but if the values of the attributes are not ‘filled out’ the patterns are considered rules. The generalized rules are rules with certain defined attributes.

The examples are generated from dictionaries, corpora or collocation databases and the rules are created manually. Every source language pattern has a corresponding target language pattern, which are then paired up in the translation process. There is a chance that a given target language unit corresponds to more than one target language unit. In this case more specific patterns override general ones, in other words the system will choose the source language pattern whose attribute has a more defined value.

In the translation process (Figure 2) the analysis phase is followed directly by the target language synthesis phase, thus generating the target language tree. The transfer phase is skipped in this process. The target language tree is immediately used to generate the target language sentence. The three phases of analysis are the following:

1. Sentences are tokenized into words; morphological analysis.

2. Syntactic analysis of the input sentence with a bottom-up analyser which generates the target language tree with terminal and non-terminal symbols. If the sentence is correct, the analyser will create one or more root symbols.

3. The target language tree is formed, the morphological generator then creates the output sentence from the terminal symbols on the leaves of the tree. The set of terminal symbols is a finite one containing the elements that make up the sentences of the language (Prószéky & Kis 1999: 115).

This process does not require a transfer phase to transform the source language representation into a  target language representation, nor does it require an intermediary language.

Figure 2

The MetaMorpho translation process

The system uses unification grammar and the morphological analysis that follows the tokenization is done by the HUMOR morphological analyser (Prószéky 1994).

MetaMorpho can be used to translate smaller texts and entire webpages (http://

www.webforditas.hu).