Book review

Analogical modeling

Royal Skousen, Deryle Lonsdale and Dilworth B. Parkinson (eds). 2002. Analogical Modeling. An exemplar-based approach to language. Amsterdam: John Benjamins.

Reviewed by Remi van Trijp, Sony Computer Science Laboratory Paris.

Analogy has a long history in research on natural language processing and human cognition in general, but its definition and relevance have always been a big point of debate. Generally speaking, analogy means that speakers of a language reuse previously learned examples and experiences directly rather than using declarative rules. Its often vague definition (e.g. de Saussure, 1972) has prompted mainstream linguistics to prefer more abstract and clearer defined rules over analogical processing. Analogical Modeling, however, tackles this problem and offers a fully operationalized and implemented version of analogy. This book succeeds extremely well in providing the reader with a tutorial on analogical modeling (AM) and the state-of-the art of the field, and is especially interesting for computational linguists.

Part I: The basics of Analogical Modeling

The first part of the book contains two introductory chapters by Royal Skousen. Skousen is the main driving force behind AM and this book can best be seen as an extension of his previous work (see Skousen 1989, 1992). He shows the underlying algorithm of AM in a tutorial example in which AM has to predict the pronunciation of the letter c in English. In Chapter 2, he discusses the main points of debate in present-day research on AM, such as variable weighting and the heavy computational cost of the algorithm.

Part II: Psycholinguistic evidence for Analogical Modeling

The second part of Analogical Modeling considers whether AM is a cognitively realistic model of natural language processing. This part contains one chapter contributed by Steven Chandler, who reviews psycholinguistic evidence for and against various non-rule-based approaches to language. Chandler offers the reader a good understanding of the place that AM takes among psychological models of category learning. He then discusses experiments in which AM is compared to other procedural approaches of language, such as nearest neighbor and connectionist models.

Part III: Application to specific languages

The next part contains two chapters in which AM is put to test on actual language data. In Chapter 4, Doug Wulf applies AM to the problem of the German plural. The algorithm performs remarkably well at the task and categorizes the German plural with an adult-like accuracy. In the second case study, contributed by Anton Rytting, the AM algorithm has to predict the /k/-Ø alternation in the Turkish nominal system. Even though AM does not reach adult-like accuracy, it performs as well as previous experiments with rule-based grammars.

Part IV: Comparing Analogical Modeling with TiMBL

The fourth part is perhaps the most interesting one for researchers from the field of Machine Learning. Here, Analogical Modeling is compared to TiMBL (Tilburg Memory-Based Learner; see Daelemans and Van den Bosch, 2005). TiMBL is a k-nearest neighbour classifier that starts from the similar philosophy of reuse of examples and past experience, but has a different implementation. Three comparisons between AM and TiMBL are presented by David Eddington (Chapter 6), Walter Daelemans (Chapter 7), and Andrea Krott, Robert Schreuder, and R. Harald Baayen (Chapter 8).

Most of the experiments discussed in this part show that, all in all, TiMBL and AM score equally well on any given task. In one of the experiments (on predicting German plurals), TiMBL even statistically outperforms AM. As Daelemans argues, this poses a serious challenge for AM: from a language engineering point of view, there is no reason to use the computationally costly AM instead of the far more simple Memory-Based learner.

But what about cognitive realism? David Eddington argues in his contribution that AM is a cognitive more plausible model because it makes the same kinds of errors as children do when learning their language. He investigates this hypothesis through experiments on Dutch stress assignment, Spanish gender assignment, diminutive formation and stress assignment. The results, however, do not support his claims: in the first three experiments, there is no significant difference between the kinds of errors made by AM and TiMBL. Only in the last experiment, AM captures the developmental stage of children better than TiMBL, but only after manually manipulating the weight of certain features in the dataset.

Part V: Extending Analogical Modeling

Next, four papers address possible extensions of AM. The first one, by Antal Van den Bosch, tackles the question how a k-NN learner could be redefined in terms of instance families of same class-behavior (also see Daelemans and Van den Bosch, 2005). In the second paper, Mike Mudrow compares AM with a localist connectionist model (SimNet) and argues how one model can complement the other. James Myers then investigates the compatibility between AM and Optimality Theory because of the growing interest from Optimality Theory in the use of exemplars. Finally, Christer Johansson discusses the possibility of emergent linguistic categories through AM rather than given and fixed categories.

Part VI: Quantum computing and the exponential explosion + Appendix

In this last part, Skousen tackles the question of exponential explosion of possible analogical categories, an inherent problem of the costly algorithm of AM. This problem also makes certain aspects of AM computationally intractable, whereas this is no problem for other exemplar-based learners. The solution - according to Skousen - lies in quantum computing.

Evaluation

Analogical Modeling is an interesting read for anyone interested in alternative ways of processing and learning language, in particular for researchers with a more computational background. It is especially appealing because of the tutorial for interested researchers, the general theoretical background of the approach and the strong emphasis on empirical data and the willingness to compare different approaches in the experiments.

References

Daelemans, Walter, and Antal Van den Bosch. 2005. Memory-Based Language Processing. Cambridge: CUP.

De Saussure, Ferdinand. 1972. Cours de Linguistique Génerale. Edition critique préparée par Tulio de Mauro. Paris: Payot.

Skousen, Royal. 1989. Analogical modeling of language. Dordrecht: Kluwer Academic Press.

Skousen, Royal. 1992. Analogy and structure. Dordrecht: Kluwer Academic Press.

Links

© 2002-present ICLA; all rights reserved.
Site admin information and contact