Manny Rayner, David Carter, Pierrette Bouillon, Vassilis Digalakis and
Mats Wiren, editors. The Spoken Language Translator. Cambridge
University Press. 2000. ISBN 0-521-77077-7 Price $70 (hardback).
xviii+337 pages

With the turning of the century, it has become increasingly clear that
core language technologies, such as statistical parsing or tagging,
have largely reached a plateau in performance. For example, new papers
on parsing tend to deal with peripheral concerns, such as porting to
new domains or else highlighting new machine learning techniques. When
improvements in parsing performance as reported, they tend to be very
modest indeed. On the basis of such observations, one could conclude
that many language engineering tasks are solved problems. However, as
anyone who has actually used these `converged' technologies will say,
there is still room for improvement.

Now, there are a number of possible reasons for this (apparent)
convergence (and hence impasse). With only so much expensively
annotated training material, there is a ceiling on what existing
statistical systems can achieve, and we have by-and-large reached that
limit. Yet another possible reason for a slow-down in performance
improvement is that our understanding of how language works is still
very rudimentary, and what little we do know about language is further
restricted by the need to fit models to the available data.

It therefore seems timely that we revisit those largely rule-based
systems that are currently unfashionable and see whether we can
salvage anything from them. In particular, clever use of the knowledge
present within those systems, augmented with probabilistic components,
seems to be our main hope of pushing forwards progress in language
technology. In order to make this next step, we need to understand
what people used to do in the days before statistics took over. Enter
this book, which can be seen as a very accessible account of a mainly
rule-based system for translating spoken language.

The book is divided into four parts: the natural language processing
aspects of the system, a discussion of the linguistics of the task,
the speech processing techniques used and finally evaluation. In
total, there are $21$ chapters. Since there are so many chapters, I
will instead tackle the book in terms of the four parts.

The introduction to the book summarises the system (which largely
consists of technology developed for the earlier Core Language Engine
coupled with the SRI DECIPHER speech recognition system). A nice
aspect of this chapter are the numerous examples used to illustrate
the translation process. The authors seem to be aware that, when they
were developing the system, events were overtaking them elsewhere:
they have an explicit defence of using rule-based grammars.  The
evaluation of their system presented here, and expanded in the final
part, is sober and well-balanced.  

The first part of the book firstly explains how Quasi Logical Forms
(QLFs) can be used as the transfer formalism used in their machine
translation process. Numerous examples aide the reader in
understanding what is happening. There is then a description of their
application of Explanation-based Learning (EBL) to grammar
specialisation. EBL is a version of speed-up learning, where the idea
is to transform a theory into another version whereby some of the
intermediate steps are collapsed into macros. Within the context of
grammars, it can speed-up parsing as few rules to be applied. Since
their grammar-based system, just like any other system, has to deal
with ambiguities, a chapter discusses how they select amongst
alternative parses. Their approach is somewhat ad-hoc, being based
around `discriminants' (predicates for making choices at points of
ambiguity). With hindsight, a log-linear approach, which is now the
standard way of modelling parse selection within feature-based
grammars, would have been more appropriate. Creating annotated
training material or adding new linguistic rules is always a pressing
concern, and so there is a discussion of tools for helping create
training material for the discriminants and for adding new lexical
items to the system. The first part concludes with material on
morphology and on creating training material for the task.


Part two, which deals with the linguistic resources used, focuses
largely on the coverage of English, French and Swedish manually
written grammars. It also considers relationships between these
grammars. Since the grammars do have to cover real-world language,
they deal with aspects of language not normally found in linguistic
textbooks, such as dates. These chapters only sketch the grammars, and
the interested reader wanting more details will need to revisit
earlier descriptions of the grammar used in the Core Language Engine.
One missing part here is an estimate of how long it took to develop
the English grammar. Whilst developing grammars for related languages
may be quick given the existence of such a previously constructed
grammar (as the authors argue in a chapter on bootstrapping one
grammar from another one) this ignores the start-up cost.

Part three, which deals with the speech recognition aspects of the
project, starts with a review of the basic technologies used (HMMs).
It discusses the important problem of making the system operate in
near real-time. and then goes on to consider language modelling within
the context of MT. One interesting section discusses the possibility
of having a manually-written grammar generate strings with which to
train a statistical language model. There is also a brief mention of
using a language model for treating speech containing code-switching
(utterances containing a mixture of languages). As with the previous
part, there is a discussion of the porting problem (this time, that of
porting a speech recogniser to a new language). The problem of
adapting to new dialects is considered. This part ends with a general
chapter considering various odds-and-ends of their system.

The final part tackled system evaluation. As then as now, this
continues to be a vexed problem. For example, it is hard to have an
objective definition of what it means for translation to be poor.
Here, humans are used to make judgements regarding the degree of
acceptability of the translation. A large number of tables summarises
various lesion studies (results after crippling the full system in
various ways). A minor quibble is that because the tables are spread
over a number of pages, it requires some work to make comparisons.
Because the authors carry out full-system evaluation, they are able to
observe whether component-level errors multiply, or else whether
errors in one module can be compensated by some other module. This
therefore tests the hypothesis that the overall error rate of the
system can be computed on the basis of the error rates of the
individual components. Their findings suggests that at times, errors
made by one component can be cancelled by another component. This
means that the error rate of the overall system would be overestimated
if it were computed as the product of the individual error rates.

The book ends with reflective comments on the success (or otherwise)
of their system.  Refreshingly, they do not claim stellar performance,
nor that automatic speech recognition has been solved by their work.
Instead, they suggest that their book should be seen as being a
well-documented account of a large speech and language project.  This
is indeed the case.

In summary, the book is very well written and structured. There are
many lessons here for subsequent generations of speech and language
researchers. Whilst the technology described may no longer be
state-of-the-art (for example, statistical machine translation has
probably replaced transfer-based translation), many of the problems
discussed -- such as porting to a new domain, or creating new
annotated training material -- will always be present. The book itself
is not quite self-contained, and would best be read in conjunction
with the earlier Core Language Engine text. That aside, it would be a
good primer for anyone wishing to develop a serious speech or language
processing system.

Miles Osborne
University of Edinburgh
Institute for Communicating and Collaborative Systems
2 Buccleuch Place
Edinburgh EH8 9LW
UK

miles@inf.ed.ac.uk