Ruslan Mitkov, ed., The Oxford Handbook of Computational Linguistics.  Oxford University Press, 2003.  xx + 784 pp.

 

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

 

 

After an introduction by Martin Kay, this book surveys the field of computational linguistics in a series of 38 articles, averaging about twenty pages each, on topics ranging from theoretical underpinnings to data resources and practical applications.  Reading the list of authors gives a sense that the editor asked the leading people in different areas to write, and usually they agreed:  the level of star quality is high.  The authorship is also wide-ranging, geographically and otherwise.  Thirty per cent of the (co-)authors work in industry rather than academe; 60% of them work in Britain or the USA, but eight continental European nations together with Canada and Japan are also represented.

            The volume is entitled a ‘Handbook’ (rather than ‘Encyclopaedia’), implying that it is intended as much for use by newcomers aiming to read themselves into the subject (or to one of its subfields) as for finding answers to specific, focused questions.  But it contains author and subject indices and a 34-page glossary, together with many cross-references inserted into the articles by the editor, so the latter kind of use is also realistic.  Martin Kay expresses the hope that the volume will go a long way towards meeting the so-far unmet need for a ‘comprehensive reference work’ on computational linguistics.  Since someone who reviews a sizeable reference book like this cannot reasonable be expected to read the whole from beginning to end, I hit on the idea of assessing the book by checking the coverage of a random assortment of topics which someone consulting such a work might want to look up.

            I picked ten test items, aiming for a mixture of concepts and proper names, and of theoretical and empirically-oriented concerns.  Alphabetically, my list ran:

 

anonymization

deterministic parsing

Fidditch (Donald Hindle’s well-known parser, used in classic studies such as Hindle and Rooth 1993)

Generalized Phrase Structure Grammar

Hidden Markov Models

interlingua

kappa statistic

London-Lund Corpus

Swiss German (and overlapping dependencies)

W.A. Woods (and ATNs)

 

How did the Oxford Handbook fare on this test?

            Anonymization was possibly an unfair choice:  my own view is that developing an ethically-acceptable approach to anonymizing individual references in electronic corpora of speech or unpublished writing is an important issue for computational linguistics, and that current practice is seriously inadequate (see e.g. Sampson 2000: sec. 4.1; Rock 2001), but to date this opinion is not widely shared.  So it was no surprise that anonymization does not feature in the Handbook index.

            Deterministic parsing does not feature under that name, either; but Bob Carpenter’s chapter on ‘Complexity’ does include a very brief allusion to ‘Marcus parsing’ and to Marcus (1980).

            Fidditch is not mentioned, although two chapters each contain a reference to other work by Hindle.

            Generalized Phrase Structure Grammar is given seven lines in Ronald Kaplan’s chapter on ‘Syntax’, though his treatment, as well as brief, is strikingly unsympathetic.  For Kaplan, it seems, the most important thing about GPSG is that it cannot handle cross-serial dependencies.

            Hidden Markov Models are covered at several points, including a three-page section in Christer Samuelsson’s chapter on ‘Statistical methods’.

            I had thought that the term interlingua, for an approach to machine translation which transposes one natural language into another via some (usually artificial) universal intermediate language, might be an unreasonable test, as being mainly of historical interest.  But the Handbook covers machine translation in two chapters – one by John Hutchins (‘General overview’) and another by Harold Somers (‘Latest developments’), and both discuss interlinguas.  Contrary to my impression that the technique had not been taken seriously since a Dutch research group some twenty years ago tried using a form of Esperanto as an interlingua, Somers states that the interlingua method has recently been revived:  he allots it a full page.

            The kappa statistic (see e.g. Carletta 1996) is crucial for corpus annotation work in connexion with the issue of inter-annotator reliability, but it is not indexed.  This is a failing of the index rather than the body of the book:  I did spot a reference to kappa while glancing through the volume, before drawing up a systematic list of test items, but I have not been able to find it again.

            Incidentally, gaps in the indices are not the only way in which the editing of the volume is less helpful than it might be.  The book would have been more useful as a reference work if the chapter bibliographies had been merged into one general listing, which would probably have comprised a fairly comprehensive bibliography of the significant publications in the field; instead, they are printed separately with the successive chapters.

            The London-Lund Corpus – the first electronic corpus of spontaneous English speech ever compiled, and still unique to my knowledge in containing detailed annotations of intonation for hundreds of thousands of words of speech – is not mentioned (or, at least, not indexed).  Nor are the Edinburgh Map Task Corpus, which some may see as making London-Lund obsolete by including digitized speech waves tied to the transcribed words, or the Lancaster Spoken English Corpus, which allows the reliability of manual intonation annotation to be assessed through inclusion of two analysts’ independent annotations of some passages.  Even the terms ‘intonation’ and ‘suprasegmental’ do not appear in the index.  I found these gaps quite surprising, considering that making synthesized speech acceptable to hearers seems to be one of the leading potential applications of computational linguistics in commercial terms.

            Swiss German is not indexed,  and nor is the term ‘overlapping dependencies’; but we saw in connexion with GPSG that Ronald Kaplan does allude to his work with Joan Bresnan and others on what they call ‘cross-serial dependencies’ (though this phrase is not indexed either).  My impression was that Shieber (1985) would have been a more central reference on this issue than Bresnan et al. (1982), but I am no expert.

            Finally, there are abundant references to W.A. Woods, including more than a page on augmented transition networks in Kaplan’s ‘Syntax’ chapter.

            Out of ten test items, then, we have three hits, three cases (four with the kappa statistic) where something is included but the coverage is either hard to locate or otherwise unsatisfactory, and three or four misses.

            Ten out of ten on a test like this would surely be too much to expect from a single-volume reference work.  Readers must judge for themselves whether the actual score is satisfactory.  I find it slightly disappointing.  Others, though, may perhaps feel that my list of test items was unduly biased towards my particular interests, and that a lower score is forgivable for that reason.

            Whether or not this is the ideal reference work on computational linguistics, so far as I am aware it has no close rival to date.

 

 

Bresnan, J., Kaplan, R., Peters, S., and Zaenen, A. (1982).  Cross-serial Dependencies in Dutch, Linguistic Inquiry 13: 613–35.

Carletta, J. (1996).  Assessing Agreement on Classification Tasks: the Kappa Statistic, Computational Linguistics 22: 249–54.

Hindle, D., and Rooth, M. (1993).  Structural Ambiguity and Lexical Relations, Computational Linguistics 19: 103–20.

Marcus, M.P. (1980).  A Theory of Syntactic Recognition for Natural Language.  The MIT Press, London (1980).

Rock, F. (2001).  Policy and Practice in the Anonymisation of Linguistic Data, International Journal of Corpus Linguistics 6: 1–26.

Sampson, G.R. (2000).  CHRISTINE Corpus, Stage I: Documentation (Release 2).  www.grsampson.net/ChrisDoc.html

Shieber, S.M. (1985).  Evidence Against the Context-freeness of Natural Language, Linguistics and Philosophy 8: 333–43.