The Technology behind translate

The translation technology is based on the principles of slot grammar, a method of describing grammatical structures originally developed by IBM.  The basic idea underlying slot grammar is that every sentence and each of its constituents has a central element (its head) and modifiers. For each head it can be determined what kinds of slots there are availabe for modifiers (fillers).

The slots can be determined by the part of speech of by the word itself. Thus almost every noun can be modified by adjectives but only some nouns can be modified by phrases beginning with the preposition to, e. g.: message to the company. The slots which are dependent on single words must be defined in the dictionary, since otherwise source language sentences can neither be correctly analyzed nor correctly translated.

Slots

The concept of a slot is central to the slot-grammar system. Slot grammar is based on the premise that every word has its own specific slots. Every meaning of a word must be considered, since the word house, for example, has different slots as a verb than as a noun. The concept of a slot is related to existing concepts such as complementobject, and attribute.

The slots are filled by constituents known as fillers, which can be individual words or even complete sentences. The German verb schenken, for example, determines that the person giving the present appears as the subject of the sentence, the person receiving the present as the dative object, and the present itself as the accusative object:

Er schenkt dem Kind ein Auto.

Slots which belong to a specific word are included in the dictionary as part of the entry for that word. Because a word can have several slots, the word is described as having a slot frame. The slot frame of the verb schenken consists of subject, accusative object, dative object.

Slots can be optional, like the accusative and dative object in the case of schenken, or mandatory, like the accusative object of verursachen. Optional slots can be left empty, mandatory ones must be filled to create a complete sentence.

Fillers

Slots can be filled by different constituent parts of a sentence. These elements are known as fillers and must be specified for the individual slots associated with each word.

With a verb like schenken, the accusative and dative objects are noun groups (constituent parts of a sentence whose head is a noun). In contrast, the accusative-object slot of the verb forget can be filled by clauses starting with that and by infinitive clauses, in addition to noun groups:

Er hat das Prinzip vergessen.
Er hat vergessen, dass das Prinzip gilt.
Er hat vergessen, das Prinzip zu beachten.

The Way translate Operates

translate breaks a text down into individual sentences (or, in some cases, parts of sentences), and processes it sentence by sentence. Each sentence is first broken down into individual words, which are converted into the basic form and then looked up in the dictionary. The words are assigned their grammatical properties and possible translations. The sentence is then analyzed (or “parsed”), and split into its constituent parts. The translation itself takes place in two phases: the first phase involves lexical transfer, where each word is assigned the valid translation for the specific context, taking into account the options set by the user; the second phase is structural transfer, which ensures correct word order in the translation and carries out other necessary structural changes (see also Correlating source-language and target-language complements, Transformations). Finally, the correct word forms are generated and the final translation is produced.

Correlating Source Language and Target Language Complements

In order to translate a sentence it is necessary not only to know the target-language words corresponding to the source-language words, but also the respective slot frames. Default correlations have been defined for the German and English slots (complements). For example, a standard correlation has been defined between the German accusative object and the direct object in English. For example:

begleiten subject accusative object
accompany subject direct object

But note the following example:

bedürfen subject genitive object
require subject direct object

There is no default correlation defined for genitive objects.

Transformations

A German sentence and its English translation, or an English sentence and its German translation, often differ with regard to their syntactic structure. When translate translates, it uses transformations in such cases to make the required structural changes.

There are two kinds of transformations:

  1. Lexical transformations, which are defined for specific lexical items. The respective transformations are included in the translate dictionary together with the other information on the words or phrases concerned.
  2. Structural transformations, which describe general structural differences between English and German. They are part of translate‘s transfer component.

Structural Transformations

Position of the verb at the end of the sentence in German subordinate clauses compared with normal English word order:

It is good that he has come.

Es ist gut, dass er gekommen ist.

The do construction in English interrogative sentences:

Did you answer the letter?

Beantworteten Sie den Brief?

The do construction for negation:

I didn’t answer the letter.

Ich beantwortete den Brief nicht.

Preceding objects in German:

Diesen Brief beantwortete ich nicht.

I didn’t answer this letter.

Constructions involving the use of modal verbs:

Er hatte das Buch nicht lesen wollen.

He hadn’t wanted to read the book.

Lexical Transformations

Words or constructions cannot always be translated in such a way that the part of speech and slot frame correspond exactly. translate incorporates lexical transformations for handling a large number of structural differences between German and English constructions.

Some verbs do not require an object in the source language, but do so in the target language. Many examples can be found:

He golfed.

Er spielte Golf.

I bank at Barclay’s.

Ich habe ein Konto bei Barclay.

I inconvenienced him.

Ich bereitete ihm Umstände.

He hiccuped.

Er hatte den Schluckauf.

An English construction involving a predicate adjective can often be rendered in German using a verb:

He is aware of the new situation.

Er weiß von der neuen Situation.

An English subject can be replaced by a dative object in German, and the subject can become the accusative object. The word order changes for like, but not with lack:

I like it that the vase is red.

Es gefällt mir, dass die Vase rot ist.

I lack money.

Mir fehlt Geld.

Verbs which are causative in meaning when used transitively, such as drop, are replaced by lassen in German, for which there is no direct equivalent in the English sentence:

He dropped it.

Er ließ es fallen.

You should hear me out.

Sie sollten mich ausreden lassen.

Some English constructions with the infinitive are best rendered by an adverb in German:

I like to read books.

Ich lese gerne Bücher.

She happened to find the book he lost last week.

Sie fand zufällig das Buch, das er letzte Woche verlor.

Some verbs in English passive constructions are rendered in German by reflexive verbs in the active voice:

He said that he was injured.

Er sagte, dass er sich verletzte.

When defining words in translate, the most common lexical transformation can be specified, i.e. that have is translated as sein.

She has walked to the house.

Sie ist zum Haus gegangen.

References

Eberle, Kurt (2001)
Translation mismatches in lexically driven FUDR-based MT – Towards standardization of lexicon formalisms for MT
TALN, Tours, July 2001, pp. 267- 276.

Eberle, Kurt (2001)
FUDR-based MT, head switching and the lexicon
MT Summit VIII, Santiago de Compostela, 18-22 September 2001.

Lehmann, Hubert (2002)
Integration von automatischer Übersetzung und Translation-Memorys
Präsentation
tekom-Jahrestagung, November 2002