Lexical Tuning (LT) is closely related, but fundamentally different from, a group of related theories that are associated with phrases like ``lexical rules''; all of them seek to compress lexicons by means of generalizations, and we take that to include DATR [25], methods developed under AQUILEX [10], as well as Pustejovsky's Generative Lexicon [45] and Buitelaar's more recent research on under-specified lexicons [12]. All this work can be traced back to early work by Givon [29] on lexical regularities, done, interestingly to those who think corpus and MRD research began in the 1980s, in connection with the first computational work on Webster's Third Dictionary at SDC in Santa Monica under John Olney in 1966.
All this work can be brought under the heading ``data compressio'' whether or not that motive is made explicit. Givon became interested in what is now called ``systematic polysemy", and distinguished from homonymy (which is deemed unsystematic), with key examples like ``grain" which is normally given an PHYOBJ sense in a dictionary, dated earlier than a mass noun sense of ``grain in the mass", and this lexical extension can be found in many nouns, and indeed resurfaced in Briscoe and Copestake's famous ``grinding rule" [10] that added a mass substance sense for all animals, as in ``rabbit all over the road". The argument was that, if such extensions were systematic, they need not be stored individually but could be developed when needed unless explicitly overridden. The paradigm for this was the old AI paradigm of default reasoning: Clyde is an elephant and all elephants have four legs BUT Clyde has three legs. To many of us, it has been something of a mystery why this foundational cliche of AI has been greeted later within computational linguistics as remarkable and profound.
Gazdar's DATR is the most intellectually adventurous of these systems and the one that makes lexical compression the most explicit, drawing as it does on fundamental notions of science as a compression of the data of the world. The problem has been that language is one of the most recalcitrant aspects of the world and it has proved hard to find generalizations above the level of morphology--those to do with meaning have proved especially elusive. Most recently, there has been an attempt to generalise DATR to cross-language generalizations which has exacerbated the problem. One can see that, in English, Dutch and German, respectively, HOUSE, HUIS and HAUS are the ``same word"-a primitive concept DATR requires. But, whereas HOUSE has a regular plural, HAUS (HAUESER) does not, so even at this low level, significant generalizations are very hard to find.
Most crucially, there can be no appeals to meaning from the concept of ``same word": TOWN (Eng.) and TUIN (Dut.) are plainly the same word in some sense, at least etymologically and phonetically, and may well obey morphological generalizations although now, unlike the HOUSE cases above, they have no relation of meaning at all, as TUIN now means garden. Perhaps the greatest missed opportunity here has been any attempt to link DATR to established quantitative notions of data compression in linguistics, like Minimum Description Length which gives a precise measure of the compaction of a lexicon, even where significant generalizations may be hard to spot by eye or mind, in the time honoured manner.
The systems which seek lexical compression by means of rules, in one form or another, can be discussed by particular attention to Buitelaar, since Briscoe and Pustejovsky differ in matters of detail and rule format (in the case of Briscoe) but not in principle. Buitelaar continues Pustejovsky's campaign against unstructured list views of lexicons: viewing the senses of a word merely as a list as some dictionaries are said to do, in favour of a clustered approach, one which, in his terms, distinguishes ``systematic polysemy" [12] from mere homonymy (like the ever present senses of BANK). Systematic polysemy is a notion deriving directly from Givon's examples, though it is not clear whether it would cover cases like the different kinds of emitting and receiving banks covered in a modern dictionary (e.g. sperm bank,blood bank, bottle bank etc.)
Clustering a word's senses in an optimally revealing way is some- thing no one could possibly object to, and our disquiet at his starting point is that the examples he produces, and particular his related attack on word sense disambiguation programs (including the present author's) as assuming a list-view of sense, is misguided. Moreover, as Nirenburg and Raskin [47] have pointed out in relation to Pustejovksy, those who criticise list views of sense then normally go on in their papers to describe and work with the senses of a word as a list!
Buitelaar's opening argument against standard WSD activities could seem ill conceived: his counter-example is supposed to be one where two senses of BOOK must be kept in play and so WSD should not be done. The example is ``A long book heavily weighted with military technicalities, in this edition it is neither so long nor so technical as it was originally".
Leaving aside the question of whether or not this is a sentence, let us accept that Buitelaar's list (!) of possible senses (and glosses) of BOOK is a reasonable starting point (with our numbering added): (i) the information content of a book (military technicalities); (ii) its physical appearance (heavily weighted), (iii) and the events involved in its construction (long) (ibid. p. 25). The issue, he says, is to which sense of BOOK does the ``it" refer, and his conclusion is that it cannot be disambiguated between the three.
This seems to us quite wrong, as a matter of the exegesis of English. ``heavily weighted" is plainly metaphorical and refers to content (i) not the physical appearance (ii) of the book. We have no trouble taking LONG as referring to the content (i) since not all long books are physically large-it depends on the print etc. On our reading the ``it" is univocal between the senses of BOOK in this case. However, nothing depends on an example, well or ill-chosen and it may well be that there are indeed cases where more than one sense must remain in play in a word's deployment; poetry is often cited, but there may well be others, less peripheral to the real world of the Wall Street Journal.
The main point in any answer to Buitelaar must be that, whatever is the case about the above issue, WSD programs have no trouble capturing it: many programs, and certainly that of (Stevenson and Wilks, 1997) that he cites and its later developments, work by constraining senses and are perfectly able to report results with more than one sense still attaching to a word, just as some POS taggers result in more than one tag per word in the output. Close scholars of AI will also remember that Mellish [28], Hirst [33] and Small [52] all proposed methods by which polysemy might be computationally reduced by degree and not in an all or nothing manner. Or, as one might put it, under-specification, Buitelaar's key technical term, can seem no more than an implementation detail in any effective tagger!
Let us turn to the heart of Buitelaar's position: the issue of systematicity (one with which other closely related authors' claims about lexical rules can be taken together). If he wants, as he does, to cluster a word's senses if they are close semantically (and ignoring the fact that LDOCE's homonyms, say, in general do do that!) then what has that desire got to do with his talk about systematicness within classes of words, where we can all agree that systematicness is a virtue wherever one can obtain it??
Buitelaar lists clusters of nouns (e.g. blend, competition, flux, transformation) that share the same top semantic nodes in some structure like a modified WordNet: act/evt/rel in the case of the list just given(which can be read as action OR extent or relation). Such structures, he claims, are manifestations of systematic polysemy but what is one to take that to mean, say by contrast with Levin's [41] verb classes where, she claims, the members of the class share certain syntactic and semantic properties and, on that basis, one could in principle predict additional members. That is simply not the case here: one does not have to be a firm believer in natural kinds to see that the members of this class have nothing systematic in common, but are just arbitrarily linked by the same ``upper nodes". Some such classes are natural classes, as with the class he gives linked by being both animate and food, all of which, unsurprisingly, are animals and are edible, at least on some dietary principles, but there is no systemic relationship here of any kind. Or, to coin a phrase, one might say that the list above is just a list and nothing more!
In all this, we intend no criticism of his useful device, derived from Pustejovsky, for showing disjunctions and conjunctions of semantic types attached to lexical entries, as when one might mark something as act AND relation or an animal sense as animate OR food. This is close to older devices in artificial intelligence such as multiple perspectives on structures (in Bobrow and Winograd's KRL [6]), multiple formulas for related senses of a word in Wilks [55], and so on. Showing these situations as conjunctions and disjunctions of types may well be a superior notation, though it is quite proper to continue to point out that the members of conjuncts and disjuncts are, and remain, in lists!
Finally, Buitelaar's proposal to use these methods (via CoreLex) to acquire a lexicon from a corpus may also be an excellent approach. Our point here is that that method (capturing the content of e.g. adjective-noun instances in a corpus) has no particular relationship to the theoretical machinery described above, and is not different in kind from the standard NLP projects of the 70s like Autoslog [49] to take just one of many possible examples.