Logos Model constitutes the technology that underlies both the commercial Logos System that is no longer available and the currently available, open-source variant known as OpenLogos. Further development of Logos Model linguistic knowledge base ended in 2000. The OpenLogos adaptation of the commercial version was accomplished by The German Research Center for Artificial Intelligence (DFKI) in 2005.34 This adaptation pertained exclusively to systems aspects. With one or two exceptions, all Logos Model translations in this book were run on OpenLogos. OpenLogos has reduced functions (e.g., no translation memory, no source format preservation in target output, no run-time Subject Matter Code options to govern lexical look-up at run time) and its translations, in rare instances, may not be identical with that of the commercial product (for unknown reasons). Logos Model translations in both versions have flaws as well as strengths, much as in any other MT system. We maintain, however, that translation flaws in Logos Model reflect where the system happened to be, development-wise, when work ceased in 2000, and do not imply inherent shortcomings in underlying methodology. Obviously, this is a crucial, distinguishing claim for the approach taken in Logos Model, one we try to make intelligible in the account given in this book.
With the end of American involvement in the Vietnam conflict in 1974, Logos Corporation turned to the development of new language combinations, some of which were then developed into products, others remaining simply as prototypes. For these efforts we received contractual assistance from a variety of sources: from the pre-revolutionary Iranian government for English-Farsi, from IBM and Xerox for English-French, from Siemens for German-English, from Walloon Provence of Belgium for German-French, from Swindell-Dressler for English-Russian, from the Italian government for English-Italian and German-Italian. In conjunction with these efforts, Logos Model was productized and ported from the IBM mainframe to other platforms, including Wang (VS), Sun (UNIX) and Microsoft’s Windows.
By the 1980s Logos Corporation had sales offices in Frankfurt, Germany, offering German source to English, French and Italian, and in the US (Boston) and Canada (Montreal), offering English to German, French, Spanish, Italian and Portuguese. In those early days, MT was still a more or less unproven novelty and translators often met it with skepticism. The experience we had with the Canadian government was not untypical. During a sales demonstration of the English-French system, one of the translators asked if he might enter a sentence. He sat down at the terminal and typed in:
It is quite unusual to find this type of machine in a non-creative environment that calls for less specific results and undoubtedly brings about a lot of confusion.
Needless to say, the Logos sales representative was relieved as the French translation began to appear on the screen:
Il est tout à fait inhabituel de trouver ce type de machine dans un environment non-créatif qui réclame des résultants moins spécifique et provoque indubitalement beaucoup de confusion.
Translations of this sort are standard fare nowadays, but 35 years ago, as the test sentence suggests, MT was still considered largely unproven. The Canadian government eventually purchased English-French systems for the Department of Foreign Affairs, Trade and Development, the Department of National Defense, and a number of other government entities. One private Canadian translation company used the Logos English-French system to translate tens of thousands of pages documenting all the various systems of a new Canadian frigate. “There is no doubt,” stated Manager Jean Gordon of the Centre d’Expertise at the Canadian Secretary of State, speaking of their experience with the Logos English-French system, “as long as the source text is well written, the raw translations will be quite good.”
In Germany, Siemens Nixdorf tested the viability of our system by having us translate one of its computer manuals from German into English. Two Logos staff members post-edited the machine’s output. Interestingly, one of the post-editors knew very little about computers, and the other had very limited knowledge of German. Despite this, the resultant man-machine combination won Logos a long-standing contract with Siemens Nixdorf. And we were told by Julian Cox, Nixdorf’s translation manager, that Logos Model MT motivated technical writers to compose their sentences more thoughtfully, resulting in machine output that required very little post-editing, with consequent savings in translation costs. A similar attention to source language style occurred at the German company Osram upon its acquisition of the American firm, Sylvania. Osram purchased the Logos system to facilitate online communications between the engineers on either side of the Atlantic. We were told engineers soon realized that attention to the quality of the source language paid immediate dividends in the translation quality of their messages.
SAP Germany, an early and long-standing customer, used Logos’ German system to translate documents into English and then, after editing the English output, entered it into our English system to produce documentation in French and other languages.
Undoubtedly, MT’s greatest attraction was the consistency in usage that MT by nature guarantees, a quality control function particularly needed in large translation projects. Sweden’s Ericsson was an example of this. This firm was so pleased with the effect of the Logos system on translation quality, it set up a translation bureau subsidiary to offer MT services to other companies.
Océ Technologies, the Netherland printer company, used the Logos system to translate its text, but Océ had outside translation bureaus do its post-editing. This arrangement guaranteed much sought-after consistency in usage while reducing translation costs.
In the United States, several translation agencies used the Logos System to take on huge translation jobs that other agencies could not handle. One, for General Motors, entailed over 10,000 pages. This is redolent of the very first translation contract that Logos received in its earliest days in 1970, when we translated 10,000 pages of helicopter documentation into Vietnamese, after several manual translation bureaus proved unable to address the government’s need. This was done on an emergency basis, even before the English-Vietnamese development effort was completed.
… having something to say must necessarily entail conceptual relationships, but does not yet involve lexicon and grammar, semantics and syntax, as linguists understand these matters.
Mentalese, while very different from external, spoken language, is nevertheless still language, “the language of thought” in Pinker’s terms, or I-language (for internal language) in Chomsky’s terms. According to them, mentalese is far simpler than what we usually understand by the term language. For example, mentalese is unconcerned with word order and even with words themselves. Yet mentalese is the basis for language. According to Pinker (1994, 73), when we learn a spoken language, what we are doing is translating mentalese into a string of ordered words in accordance with a particular sociolinguistic convention. The theories of the Russian psychologist Lev Vygotsky (1934/1986) on “inner speech” are relevant here. In his work Thought and Language he itemizes the following characteristics of inner language: (a) predominance of private meaning over public meanings; (b) abbreviation of syntax, e.g., where the subject or object is understood and never made explicit; (c) agglutination of multiple, complex concepts into a single abstract concept that has no conventional counterpart in public discourse. At the extreme, he envisions inner speech as “thinking in pure meanings,” i.e., conceptually, without the literalness, morphology and structure of so-called external language.
Brief overview of how Logos Model effects its translation of (13).
The literal German input string in (13) gets re-expressed as a SAL string during lexical lookup at the very start of source analysis.
As source analysis proceeds through its various stages (through pipeline modules or hidden layers, depending on the metaphor being used), this SAL input string gets progressively re-written (concatenated) to eventually look like (13)(i). To keep things simple, the SAL string in (13)(i) is shown in clear English, not in SAL symbols.
VT in (13)(i) stands for any transitive verb. The agentive, non-agentive designations for the two NP’s are derived from the SAL taxonomy for the head nouns in each case.
There are actually seven possible representation levels in a stored SAL pattern by which an input element may be represented and dealt with at any one time. These are: (i) set of POSs; (ii) POS; (iii) tagset of SAL elements within a POS; (iv) SAL superset; (v) SAL set (vi) SAL subset; (vii) literal string. For obvious reasons, the principle is always to deal with input at the abstraction level consistent with acceptable results, i.e., as near to literal as necessary and as abstract as otherwise possible.
SAL input string (13)(i) matches on a stored SAL pattern-rule having a commensurate SAL pattern, shown below in (13)(ii). Note that the agentive designation for NP2 in 13(ii), below, represents a tagset that comprises a list of agentive-like SAL supersets. This matching function is effected automatically; a process driven by the input pattern itself, not by some metarule or other form of intermediate, procedural logic. Moreover, the stored rule is never looked at if it is not potentially relevant, no more than are entries in a properly indexed lexicon looked at if not potentially relevant.
As a consequence of this match, the action portion of the pattern-rule links to a target action, and this target action, taking its clue from the semantics of the German source pattern, then transforms the German syntax into appropriate English target order preparatory to the final, literal translation.
In sum, the matchup of pattern-rule (13)(ii) with SAL input string (13)(i) occurs because both entail a transitive verb with the agent of the verb following rather than preceding the verb. That’s the anomaly (from the point of view of English) that this particular pattern-rule was designed to detect and deal with, doing so at the highest effective level of abstraction possible. In effect, the stored pattern-rule shown in (13)(ii) allows the literal German input string in (13)(i), and virtually any German sentence with reversed, subject-object word order, to be rendered in correct English order, as illustrated in (14) above, doing so purely on the basis of semantics.