Christiane Fellbaum, ed., WordNet: An Electronic Lexical Database.  MIT Press, 1998, xxii + 423 pp.  ISBN 0-262-06197-X.

 

Reviewed by:

Geoffrey Sampson, School of Cognitive and Computing Sciences, University of Sussex.

 

 

 

For automatic natural-language processing researchers who need a reasonably comprehensive machine-usable database of the meanings of English words, at present WordNet is the only available possibility. 1

 

There are electronic dictionaries derived from printed English dictionaries, including machine-usable versions of various editions of the Oxford Advanced Learner’s Dictionary (OALD) and of the Longman Dictionary of Contemporary English (LDOCE), but these are useful mainly for grammatical and pronunciation information; the definitions are either left as unformalized passages of ordinary English, as in the printed originals, or even omitted altogether (in the case of Roger Mitton’s computer-usable version of the third edition of OALD, Mitton 1986).  The electronic LDOCE does includes some limited semantic classifications, independent of the contents of the printed version; but this resource has the drawback, for many researchers, of being a commercial product to which access is tightly controlled by the proprietors and, I understand, very expensive.  (Chapter 16 of the volume under review suggests that LDOCE may be the most widely-used machine-readable dictionary, nevertheless.)

 

By contrast, WordNet, initiated by George Miller of Princeton University in 1985 and developed under a series of publicly-funded research grants, is freely available to all.  A flyer from MIT Press slipped into my review copy of this volume offers a CD-ROM containing PC, Mac, and Unix versions of WordNet Release 1.6, with associated software, for US $25, with no mention of restrictions forbidding commercial exploitation or the like; alternatively, Internet addresses are given from which the packages can be downloaded by ftp.  (Release 1.6 is one version on from the release described in the volume under review.) One cannot really call WordNet a “dictionary” in the usual present-day sense, because (apart from a very simple part-of-speech classification) it contains only semantic information; but, within its purview, it is rather comprehensive.  According to this book, the then-current WordNet release contained about 168,000 entries, on the order of double the corresponding figures for electronic versions of OALD or LDOCE.  For research that needs grammatical or phonological information in addition to meanings, perhaps it might be possible to link the semantic data in WordNet with the grammatical and pronunciation data in one of the traditional electronic dictionaries.

 

The volume under review contains an introduction and sixteen chapters by a total of 24 authors (the overwhelming majority of them based in North America) who have been involved either in different aspects of the development of WordNet, or in attempts to exploit it for various NLP applications.  It thus offers an excellent opportunity to take stock of the resource and assess its value.  At present, as already said, there is no real alternative for NLP  researchers who need semantic data for more than a toy subset of English vocabulary.  But does that mean that WordNet is an interim stopgap solution, or is it an approximation to what one would ideally hope to have available?

 

In view of the huge quantities of time and resources that have been devoted by traditional dictionary publishers over many decades to the task of documenting the English lexicon in detail, it seems surprising that a database constructed manually by academics with no access to a dictionary-publisher’s archive could be a serious contender as the leading tool in this domain.  Traditional printed dictionaries and machine-usable lexical databases have different functions, but one would have thought it far easier to achieve the latter by adapting the former rather than by building from scratch, so that Oxford University Press or another publishing house would years ago have made the efforts of this academic team redundant.  That has not happened; and the fact that the Princeton team’s work remains unrivalled seems to me a remarkable tribute to their achievements.  I shall have occasion, below, to make a number of criticisms of WordNet, but my critical remarks should be interpreted in the context of that fundamental point.

 

There is a good reason why George Miller and his team developed their database from scratch rather than basing it on an existing dictionary: they planned it not as a service to NLP research but as a testbed for psycholinguistic theories about mental processing of vocabulary in humans (cf. p. 4 of the editor’s Introduction).  A point which emerges clearly from this book is that the eventual WordNet user base, consisting largely of computer scientists interested in developing practical language-engineering applications, is almost disjoint from the cognitive science clientele for whom the resource was intended.  (George Miller notes on p. 43 that “WordNet has been largely ignored by psycholinguists”.)  This again makes it all the more remarkable that no rivals to WordNet have been produced by organizations more directly involved with information technology, and it explains why the structure of WordNet is not always ideally adapted to current users’ purposes.

 

The structure of WordNet, in brief, is a set of separate networks for different parts of speech, in each of which the nodes of the network correspond either to individual words or to “synsets” (sets of approximate synonyms), and the links between nodes correspond to semantic relationships.  Prominent among the relationships encoded is hyponymy.  Thus, nouns are represented by means of networks for 25 broad classes of noun such as “animal”, “event”, “food”, “shape”; each network consists in large part of a tree structure whose root node corresponds to the general concept, and in which paths leading down from the root traverse nodes representing increasingly specific concepts.  For instance, horse is coded as a hyponym of mammal, and mammal as a hyponym of the top-level concept animal (there are intermediate nodes in both cases).  Hyponymy is not the only relationship coded – the noun networks also encode e.g. relationships of antonymy, and meronymy (part-whole); and the subsets of the networks limited to hyponymy relationships are not perfectly tree-structured (some concepts are treated as hyponyms of alternative concepts which are incommensurable with one another, for instance killing is a hyponym both of termination and of felony).

 

An additional complication is that the network of hyponymy relationships between nouns apparently requires some nodes which correspond to no single items of English vocabulary; for instance, the nouns offender and libertine are treated not as immediate hyponyms of person but as hyponyms of the phrase bad person, which is itself a hyponym of person.  According to Christiane Fellbaum (p. 8), without the bad person node the words offender and libertine “would be (infelicitous) sisters to adventurer, lover, and worker”.  This point perhaps needs a little more spelling out than it receives in the book.

 

The networks for parts of speech other than nouns are based on different relationships.  For verbs, the main relationship is described as lexical entailment: for instance, snore lexically entails sleep because the proposition he is snoring entails the proposition he is sleeping.  (Hyponymy in nouns might be regarded as a special case of entailment.)  The network for adjectives is organized very differently from those for other parts of speech, because the hyponymy concept is felt to be irrelevant to adjectives.  (According to Katherine Miller, p. 48, “it is not clear what it would mean to say that one adjective ‘is a kind of’ some other adjective”.  Is scarlet not a kind of red?  But perhaps this type of relationship is less pervasive among adjectives than among other words.)  For adjectives, the main relationship coded is antonymy, and this reflects the psychological interests of the project initiators: when subjects are asked to produce a word associated with a stimulus word, the response to an adjective stimulus is commonly its antonym.

 

Some of the earlier chapters in the book discuss the organization of the database in terms of what is displayed on screen when a word is looked up.  This seems a peripheral issue: WordNet is not really a tool for manual investigation of individual words (a traditional dictionary is better for that).  But later chapters discuss experiments which exploit WordNet to develop automatic systems for executing various NLP tasks that require semantic data, for instance information retrieval, or spelling correction in cases where a mistyped word coincides with the correct spelling of another word.  These experiments are all quite preliminary – at this stage, they could hardly be otherwise; but they do give the reader a clear impression of what WordNet has to offer.  (The experimenters also make many suggestions about ways in which the resource might be improved.)

 

The fact that WordNet has not appealed to psychologists seems unsurprising.  To be taken seriously as a scientifically valid model of how the language actually functions semantically for its speakers, it would need to embody some awareness of philosophical logic, a topic which is never mentioned in this volume.  This gap is admittedly not unique to the WordNet project.  It is a long-standing puzzle to me how completely the insights of “ordinary language philosophy” have been overlooked in recent decades, after forming one of the dominant academic paradigms from the Second World War until about 1975.  Anyone who engaged in abstract thought about language in those years found himself confronted at all turns by the principle, expounded in England by Ludwig Wittgenstein and in America by Willard Quine, that there is no real difference between “dictionary knowledge” and “encyclopaedia knowledge” (a distinction whose reality is assumed again and again by the writers in the volume under review).  As Wittgenstein put it, “what to-day counts as an observed concomitant of a phenomenon will to-morrow be used to define it” (Wittgenstein 1958: §79).  This principle has never, I believe, been seriously challenged, just forgotten.  It seems to imply both that an enterprise like WordNet is ultimately doomed as a model of psychological reality, because semantic relationships between words are too fluid and mutable to be pinned down in the way it aims to do, and also that the temporary relationships which do obtain are not restricted to any limited set of relation-types such as hyponymy and meronymy: the fact that one statement entails another, and hence helps to define the meanings of the words involved, tells us nothing about the grammatical or logical structures of the statements.  Within the WordNet community, this issue is apparently known as the “tennis problem” (the meanings of tennis, ball, racquet, net, etc. are perceived to be connected but are not related by hyponymy or other relationships recognized by WordNet); it is not a problem which the WordNet approach could hope to solve.  In fact, with respect to philosophical logic the system is so naive that it evidently (e.g. p. 134) recognizes no distinction between the species/genus relationship, as in horse/animal, and the individual/universal relationship, as in Shakespeare/author, treating both indifferently as cases of “hyponymy”.

 

But these obstacles to accepting the network system as a scientifically valid model of speakers’ psychology are not reasons for dismissing the usefulness of the system in practical language-engineering contexts.  After all, traditional paper dictionaries do an excellent job of defining the meanings of words, despite Wittgenstein.  They use rules of thumb rather than scientific theories to decide where to draw boundaries between “dictionary” and “encyclopaedia” knowledge.  Not all the properties mentioned in traditional dictionary definitions can be reduced to the WordNet relations of hyponymy, meronymy, etc.; but perhaps a useful proportion of them can be.  For NLP purposes, systems that work fairly well for much of the time are much better than no systems.

 

Even at this banausic, practical-approximation-to-God’s-truth level, the book under review does raise doubts in my mind about some aspects of WordNet, but these relate to possible faults which in principle could be remedied.

 

In the first place, some of the relationships between particular words which are quoted as examples seem to assume rather specialized registers of discourse.  For instance, George Miller identifies beak as a hyponym of jaw (p. 38), and seems to imply that fish is a hyponym of animal (p. 32).  It may be that, to an evolutionary biologist, a bird’s beak is a kind of jaw, but I do not think an average speaker of English understands the words that way – the example came as a surprise to me when I read it.  Likewise, in everyday usage the word fish contrasts with animal rather than entailing it.

 

In cases like these the response might be that, if the average speaker does not count a beak as a jaw, biology has shown that he ought to.  Whether this is a satisfactory response will presumably depend on the nature of the applications for which WordNet is used.

 

More worrying are cases where the claims made about English seem just wrong, for any discourse register.  The book contains examples relating both to meaning and to usage.

 

In the case of meaning relationships, for example, Katherine Miller (p. 50) uses weighty v. weightless to exemplify antonymy, and airy v. heavy to exemplify “conceptual opposites that are not lexically paired”.  Neither of these examples works for any style of English I know.  Weighty is an abstract term referring to significance, real or purported; weightless refers exclusively to physical lack of gravitational effects, commonly in connexion with space travel.  Airy has different meanings in, say, “airy room” and “airy comment”; in the former case, a rough opposite might be stuffy, in the latter perhaps considered or the like, but heavy seems well off target in either case.

 

In the area of grammar and usage, things get worse.  Shari Landes et al. quote, as their paradigm example of erroneous part-of-speech tagging, the classification of “the adverb beyond in the phrase rich beyond her wildest dreams” as a preposition (pp. 209-10).  There is some looseness in the way that different schools of linguists fit the traditional part-of-speech names to usage in awkward cases, but, in any version of grammar I have encountered, beyond in this use is a preposition – it is not an awkward case.  Again, chapter 6, by Karen Kohl et al., discusses aspects of the WordNet verb networks which reflect the clause patterns in which different verbs occur.  The authors contrast clauses which are “ungrammatical”, such as the witch turned him from a prince (i.e. changed him from a prince into something else), with clauses which allegedly are grammatical, such as the spaceship revolves around the earth.  This latter example is used repeatedly in the chapter; but to my mind (and I have confirmed this with one other English native speaker) it too is strange English.  A satellite can “orbit” or “circle” the Earth, or “travel” or “move” round the Earth, but it cannot “revolve (a)round” the Earth.

 

Obviously, in a system on as large a scale as WordNet that was built up from nothing over a few years, there are bound to be many errors waiting to be eliminated.  But it is a little disappointing that the researchers responsible for developing the system are not sensitive enough to the language to avoid such cases when choosing illustrative examples.  Quite often in this book one encounters passages which read as if modern English is a remote, imperfectly-understood phenomenon, rather than the everyday medium of communication of almost all these writers.

 

So, in summary, on the evidence of the book I suspect that WordNet may be more a stopgap than an ideal solution to the need for a machine-usable dictionary of English word meanings.  But, as stopgaps go, it is clearly a very rich and comprehensive one.  Its creators deserve admiration for what they have achieved over so short a period.

 

Finally, I should like to make a plea that relates not to the content of WordNet but to the language used in discussing it.  The writers frequently need to refer to the converse of “hyponym” – the relation which animal bears to horse.  The editor’s Introduction uses the term “superordinate”, and Sanda Harabagiu and Dan Moldovan at one point use the term “subsumer”, but otherwise all these writers repeatedly use the neologism “hypernym”.  This is a barbarous formation: the Greek root for “name” is onym-, not nym-, so obviously the form ought to be “hyperonym”.  It is also thoroughly impractical: “hyponym” and “hypernym” are opposites with identical pronunciations.  Sir John Lyons’s standard work Semantics argues that even “hyperonymy”, although current, is too similar to “hyponymy” to avoid confusion, and he advocates the term “superordination” (Lyons 1977: vol i, p. 291).  But “hyperonym” is nowhere near as confusing as “hypernym”.  Can we hope that this horrible word will be banished from future lexical discourse?

 

 

References

 

Lyons, J.  1977.  Semantics (2 vols.).  Cambridge University Press.

Mitton, R.  1986.  “A partial dictionary of English in computer-usable form”.  Literary and Linguistic Computing 1.214-15.

Wittgenstein, L.  1958.  Philosophical Investigations, 2nd ed.  Blackwell (Oxford).