Controlled Languages: The DLT Intermediate Language (DLTIL) and Its Ambiguity Tests
Take one Esperanto sentence about “specific actions” that must be taken to raise saturation and advance adoption “in all humanitarian-type contexts.” Where does the phrase “in all of them” attach? It can modify the whole condition, the first goal, the second goal, or the verb. A test set for the DLT intermediate language, often shortened to DLTIL, can be built from variants like these. The words stay the same, but the position of one phrase moves. The sentence in each variant means something slightly different.
That is the kind of material the evaluation “Controlled Languages: DLTIL” worked with. Its page consisted of lettered variants of one long conditional sentence. In some, an adverbial phrase sits at the start of the clause. In others it sits before the first goal, before the second goal, or at the very end. A few variants also add a marker that separates the two coordinated goals. The question behind the set is whether a language designed as a pivot for translation can make such differences visible and fixed.
The idea: an unambiguous pivot
An intermediate language, or interlingua, is a middle step in translation. Text goes from the source language into the pivot, and from the pivot into each target language. Distributed Language Translation (DLT) was a project to develop an interlingual machine translation system for twelve European languages, which ran between 1985 and 1990. Its distinctive feature was the use of a version of Esperanto as the intermediate language, with translation divided into two stages: from L1 into IL and then from IL into L2.
The reasoning was that the pivot must be more reliable than any ordinary human language. Klaus Schubert’s description of the design argues that since normal human languages do not allow this, a special language that machines can translate onward especially reliably must be chosen. He also notes that the lightly modified Esperanto of DLT is a syntactically unambiguous language. This is why the language works as a controlled language. Its grammar is regular enough to be restricted further, so that each sentence has one reading.
Why attachment is the hard case
Word-level ambiguity is the familiar problem, such as one word with two senses. Attachment ambiguity is harder to see. The words are all clear, but the grammar allows a phrase to hang from more than one place. Esperanto’s regular endings help with word class, since every noun ends in -o and every adjective in -a. They do not by themselves say which head a free-floating phrase belongs to. A pivot that claims to be syntactically unambiguous has to settle this through word order, punctuation or explicit markers. Variants like those above test whether it does.
DLT placed much of the disambiguation work before the pivot stage. Witkam’s project overview describes a main semantic analysis, carried out entirely in the intermediate language, that searches a knowledge base of some 75,000 semantically related Esperanto word pairs. That overview also admits that the pivot did not remove every problem. On the French-generation side, parallel translations are less numerous because of the syntactic unambiguity of Esperanto and its lack of homonyms, but they are not absent.
What remains relevant
The pivot idea has survived in other forms. Neural systems have shown that a model can translate between a language pair it was never trained on directly, as in Google’s 2016 “zero-shot translation,” for example Ukrainian to English with Russian as an intermediate. A recent ACL Anthology paper describes Esperanto as relatively underexplored in the context of modern machine translation approaches and evaluates open-source systems for it. The older design question remains: how much ambiguity can be removed before text reaches the translator, and how much must the system resolve itself?
FAQ
Why was Esperanto chosen for DLT?
According to Schubert, it has a very clear morphological and syntactic structure, which makes it easy to parse. It has also been used for over a hundred years in an international community of second-language speakers.
Did DLT translate in one step?
No. Translation ran in two stages through the intermediate language. The intermediate translation could be transmitted over a network to any number of workstations, which handled the translation into the desired language.
