The general idea suggested by the discussion of the like-plaire example at the end of the previous section is that comparative grammar (hence transfer) becomes simpler as linguistic analysis goes deeper --- as the representations become more abstract. In fact, a major objective of MT research is to define a level of analysis which is so deep that the comparative grammar component disappears completely. Given such a level of representation, the output of analysis could be the direct input to the target synthesis component. Representations at such a level would have to capture whatever is common between sentences (and expressions of other categories) and their translations --- that is they would have to be representations of `meaning' (in some sense). Moreover, such a level of representation would have to be entirely language independent --- for example, if it preserved features of the source language, one would still require a transfer component of some kind to produce the corresponding features of the target language. For this reason, such a level of representation is normally called an Interlingua , and systems that use such a level are called Interlingual.
The relationship between transfer and interlingual systems can be pictured as in Figure . As one can see, the size of the contrastive grammar (hence the transfer component) between two languages decreases as the level of representation becomes more abstract. As this diagram perhaps suggests, the difference between transfer representations and interlinguas is a matter of degree rather than absolute distinction (for example, Chapter shows how one might combine an interlingual representation of tense and aspect with a transfer approach to other phenomena).
There are a number of clear attractions to an interlingual architecture. First, from a purely intellectual or scientific point of view, the idea of an interlingua is interesting, and exciting. Second, from a more practical point of view, an interlingual system promises to be much easier to extend by adding new language pairs, than a transfer system (or a transformer system) . This is because, providing the interlingua is properly designed, it should be possible to add a new language to a system simply by adding analysis and synthesis components for it. Compare this with a transfer system, where one needs not only analysis and synthesis , but also transfer components into all the other languages involved in the system. Since there is one transfer for each language pair, N languages require transfer components (one does not need a transfer component from a language into itself). For example, extending a system for 3 languages into one for 5 means writing 14 new transfer components (as one goes from 6 to 20 transfer components) , and going from a 5 language system to a 9 language system means going from 20 components to 72.
Figure: Transfer and Interlingua
Ideas about interlinguas are intimately tied up with ideas about the representation of meaning. We will look at this in more detail in Chapter . However, one can get a flavour of the problems that are involved in defining an interlingua by considering the following.
Producing an interlingual representation involves producing a representation that is entirely language independent (for the languages one wants to translate, at least). This involves producing a language independent representation of words, and the structures they appear in. Under the latter heading, one would have to make sure one could represent the difference in meaning between examples like those in ( ) --- assuming one does not want them all to translate alike, that is --- and find a way of representing the meaning that is expressed by various tenses, and by the distinction between definite, and indefinite NPs (e.g. a printer vs. the printer).
While this raises many unsolved linguistic problems, it is the language independent representation of word meaning that seems to pose the most difficult problems. The central problem is how to choose the vocabulary of the interlingua --- what are the primitive concepts of the meaning representation to be. Notice that this is not a question of what names we should give the concepts --- how we should write them down or represent them. Of course, we should make sure that we do not use one name for two concepts, which might be confusing, but beyond this, we can give them, for example, names from an existing language (e.g. English, or Esperanto ), or numbers, or codes in some invented language --- the only difference here will be how easy they are to write or remember. The problem is one of identity. For example, are we to include a concept that we might write as CORNER --- this being the interlingual representation of the English noun corner? This seems natural enough from the point of view of English, but from the point of view of, for example, Spanish it is not so natural, because in Spanish there are different words for inside corners ( rincón) and outside corners ( esquina). Is there any reason why we should not choose a more specific primitive word for our representation, for example, OUTSIDE-CORNER and INSIDE-CORNER. Similar problems will arise wherever one language has several words that correspond to one word in another. The point is that different languages `carve the world up' differently, so settling the choice of vocabulary for the interlingua will involve either (i) some apparently arbitrary decisions about which language's conceptualization to take as basic, or (ii) `multiplying out' all the distinctions found in any language. In the latter case one will have two interlingual items for English corner (because of Spanish ), two for English river (because of the distinction between rivière and fleuve in French ), and two for English eat, because of the distinction between essen (for humans) and fressen (for animals) in German . When one consider more distant languages like Japanese, even more distinctions will arise --- Japanese does not distinguish between wearing and putting on, as does English, but does make a distinction according to where the item is worn or put on (e.g. on the head vs on the hands). Of course, one solution to this multiplicity of concepts is to try to reduce the set of primitive concepts, defining complex concepts in terms of the primitive ones. For example, one might think that EAT is not a primitive, but that INGEST is, and that the interlingual representation of the meaning of eat should involve INGEST, and some other primitives. However, though this solves the problem of the number of concepts, it does not overcome the problem of arbitrariness, and it raises the problem of finding an adequate set of primitives to capture the relevant distinctions (the reader might, as an exercise, like to consider what a set of primitives would look like to distinguish a handful of verbs like eat, drink, gobble up, feed on, or find a set of primitives that will distinguish between different kinds of furniture (chairs, stools, tables, etc.)).
Figure: The Components of an Interlingual System
A further problem is that using an interlingua in MT can lead to extra, unnecessary work, in some cases. For example, suppose one has an interlingua intended for translation between English, French , and Japanese . Japanese distinguishes terms for older and younger brother and sister, and for various relatives depending on whether they belong to the speaker, or to the hearer (i.e. the term for my mother is different from that for your mother, or mothers in general). The problem is that this distinction has to be encoded in the interlingua , so one must decide if English brother is an older brother or a younger brother, even if one is not translating into Japanese. For example, translating Sam's brother has already left into French will involve dealing with an ambiguity, since there will be two interlingual representations differing as to whether the brother is older or younger than Sam. But of course, this is irrelevant for both English and French , and one can manage with a very simple transfer rule (along the lines of brother frère).
These are problems for general vocabulary. One should note, however, that these problems do not occur for all kinds of vocabulary. In particular, in domains where there is a codified system of terminology , the conceptual organization is generally relatively clear. In such cases, the set of concepts, and thus at least some of the vocabulary of the interlingua , is already settled. Interlinguas are rather metaphysical things. Implicitly or explicitly, they say what the universe is made of (events, processes, individuals, relations, etc.) and how it is put together. It is not at all surprising that many aspects of interlinguas are in dispute and are likely to remain so for some time to come. Given these difficulties, interlinguas in the sense described here are more popular as a basis for theoretical research in MT rather than for full-scale commercial development. For the next few years, most general purpose LK MT systems on the market are unlikely to analyse any deeper than to the level of semantic relations --- and even that will be considered impractically deep by many developers and vendors. Nonetheless, we can certainly expect a tendency towards increasingly deep analysis over the next decade or so.