Throughout most of this book, we have assumed a model of the translation machine which involves explicit mapping rules of various sorts. In the `translation by analogy', or `example-based' approach, such mapping rules are dispensed with in favour of a procedure which involves matching against stored example translations. The basic idea is to collect a bilingual corpus of translation pairs and then use a best match algorithm to find the closest example to the source phrase in question. This gives a translation template, which can then be filled in by word-for-word translation.
This idea is sometimes thought to be reminiscent of how human
translators proceed when using a bilingual dictionary:
looking at the examples given to find the source language example that
best approximates what they are trying to translate, and constructing
a translation on the basis of the target language example that is
given. For example, the bilingual dictionary entry for printer
which we discussed in Chapter 
gave the following
as examples.
's error faute f d'impression, coquille
f;
's reader correcteur m, -trice f
(d'épreuves).
Given a sentence like (
) to translate, a human translator would
certainly choose faute d'impression or coquille as the
translation, on the basis that a mistake is much more like an
error than it is like a reader.
.
The distance calculation, to find the best match for the source
phrase, can involve calculating the closeness of items in a hierarchy
of terms and concepts provided by a thesaurus . To give a flavour of
the idea, and the sort of problem it addresses, consider the problem
of translating Japanese phrases of the form A no B ( no is
a particle indicating the relation between A and B) into English.
Among the forms to choose from are AB, A's B, B of
A, B on A, B in A, and B for A, cf
Table which gives English paraphrases of examples
involving no, together with the correct translations for these
different patterns. The problem is certainly not an esoteric one,
since the expression is claimed to occur in around 50% of Japanese
sentences.
Table: Alternative Translations for the Particle no
For a given input, the system will then calculate how close it is to various stored example translations based on the distance of the input from the example in terms of the thesaurus hierarchy (this involves finding the `Most Specific Common Abstraction' for the input and the alternative translations --- i.e. `closest' concept in the thesaurus hierarchy) and how `likely' the various translations are on the basis of frequency ratings for elements in the database of examples. (Notice this means we assume that the database of examples is representative of the texts we intend to translate.)
The following is an extension to this basic idea: pairs of equivalent source and target language expression are given, along with example translations, written in parentheses, and interpreted as stating `conditions' under which the given equivalence holds. For example, the rule for the Japanese word sochira (`this', or `this person' --- i.e. the addressee, you), given below, indicates that sochira translates as this when the example involves desu, (translating as be), and as you, when the input involves something like okuru (translating as send). In translating an input like sochira ni tsutaeru, the English pronoun you would be selected as the translation of sochira, because tsutaeru (convey) is closest to okuru (send) in the thesaurus.
sochira
this (( desu {be}),...)
you (( okuru {send}),...)
this (( miru {see}),...)
This rule uses only information about the surrounding string, but one
could imagine other sorts of example, where information is given in
terms of patterns of strings, or of grammatical information. An
example involving string patterns is
given below, which would be
involved in translating examples involving the expression
o-negaishimasu along the lines of (
) ( o-negaishimasu
(`please') is a general expression indicating that a request is being
made, or a favour requested, o indicates that the preceding
noun phrase is an OBJECT).
To deal with this, rules like the following use information about surrounding string patterns:
X o o-negaishimasu
May I speak to X' ((jimukyoku {office}),...)
Please give me X' ((bangou {number}),...)
It should be evident that the feasibility of the approach depends crucially on the collection of good data. However, one of the advantages of the approach is that the quality of translation will improve incrementally as the example set becomes more complete, without the need to update and improve detailed grammatical and lexical descriptions . Moreover, the approach can be (in principle) very efficient, since in the best case there is no complex rule application to perform --- all one has to do is find the appropriate example and (sometimes) calculate distances. However, there are some complications. For example, one problem arises when one has a number of different examples each of which matches part of the string, but where the parts they match overlap, and/or do not cover the whole string. In such cases, calculating the best match can involve considering a large number of possibilities.
A pure example-based approach would use no grammar rules at all, only example phrases. However, one could also imagine a role for some normal linguistic analysis, producing a standard linguistic representation. If, instead of being given in simple `string' form, examples were stated in terms of such representations (i.e. given as fragments of linguistic representations), one would expect to be able to deal with many more variations in sentence pattern, and allow for a certain amount of restructuring in generation. In this way, one would have something that looked more like a standard LK architecture. The chief difference would be in the level of specificity of the rules. In particular, where in a traditional transfer system the rules are stated in as general a form as possible, to cover entire classes of case, what one would have here is a system where the rules are stated in highly particular forms (each one for essentially one case), but there is a general procedure for estimating, for each case, which rule is most appropriate (i.e. by estimating which example is closest). Of course, what this suggests is that there is no radical incompatibility between example-based, and rule-based approaches, so that the real challenge lies in finding the best combination of techniques from each. Here one obvious possibility is to use traditional rule-based transfer as a fall back, to be used only if there is no complete example-based translation.
