1. What is Computational Linguistics? Simply put, computational linguistics is the scientific study of language f rom a computational perspective. Computational linguists are interested in providing computational models of various kinds of linguistic phenomena. These models may be "knowledge-based" ("hand-crafted") or "data-driven" ("statistical" or "empirical"). Work in computational linguistics is in some cases motivated from a scientific perspective in that one is trying to provide a computational explanation for a particular linguistic or psycholinguistic phenomenon; and in other cases the motivation may be more purely technological in that one wants to provide a working component of a speech or natural language system. Indeed, the work of computational linguists is incorporated into many working systems today, including speech recognition systems, text-to-speech synthesizers, automated voice response systems, web search engines, text editors, language instruction materials, to name just a few. 2. Motivations for Computational Phonology A The Practising Phonologist is Frequently Beset by Two Problems Regarding Data and Analysis. (1) The first problem: It is difficult to maintain and access corpus of data if it is stored on paper.
(2) The second problem: A: Analysis is virtually impossible to check by hand. Automation promises to provide a solution, provided that phonological analyses can be represented on computer.These considerations motivated some of the earliest work on computational phonology. B. Motivation for Computational Phonology Comes from the Field of Speech Technology.The field of natural language processing is currently limited to a small group of languages which lack morphophonological alternations in the orthography. This is a challenge to develop a computational phonology which can be applied to the full diversity of the world's languages. Once this is done, the achievements of the field of natural language processing will have far greater applicability. C. Long-term Prospect of Having Integrated Speech and Language Systems.Phonology is a potential link between the speech technology community and the natural language processing community.However, contemporary phonology is inadequately formalized to play this mediating role. Perhaps computationalphonology will ultimately bridge the gap between these two independent areas of technological development. 3. Computational Phonology Methods A. Finite-state MethodsThe idea of employing "Finite-state Transducers" to represent the rule systems of generative phonology was proposed inthe early 1980s by Kaplan and Kay. Koskenniemi proposed an FST model where rules could refer to both surface andlexical context, but that these were the only levels of representation. Koskenniemi also proposed a high-level notationfor rules which could be compiled into transducer specifications. Antworth gives a detailed exposition of the rule notion,the transducer specifications and the compilation process. Ritchie et al. And Sproat also gives expositions of the twolevel model, while Pulman and Kepple present a two-level system incorporating a unification-based representation ofsegments.B. Connectionist MethodsThe view of computation as neural processing has gained popularity amongst some phonologists. The appeal of thismetaphor lies in the fact that it permits gradient behavior to be modeled and it comes supplied with learning techniques.One approach, based on the notion of spreading activation in a simple linear model, is that of Goldsmith and Larson.The network proposed for modeling the metrical grid is given. The model consists of a sequence of units, eachcorresponding to a syllable. Each unit has an activation level, lying in the range [-1,1]. The arcs represent inhibitoryrelationships between neighboring units. 4. Other Approaches to Computational Phonology There have been several other applications of computational techniques to phonology. A. Learning Ellison presents a model for automatic learning of phonological generalizations, exemplified for vowel harmony in a rangeof languages. As the system searches for new generalizations, each candidate generalization is evaluated as to itsrestrictiveness on a list of words. If a generalization is too restrictive, the number of exceptions will be high and this willdetract from the overall evaluation of the generalization. On the other hand, if a generalization is too unrestrictive, it willnot make useful predictions about which segment is coming next in a given context and this will increase the cost ofstoring the word list. B. Speech Church presents a chart parser for phonological parsing in speech recognition. Kornai presents a formalization ofautosegmental phonology which is designed to inform a new class of speech recognition devices called structuredMarkov models. To my knowledge this system has not been implemented. Phonologically well-informed approaches tospeech synthesis which have been implemented include Hertz and Coleman. 5. Links 1. The Proper Treatment of Optimality in Computational Phonology-An on-line paper-By LauriKarttunen http://www2.parc.com/istl/members/karttune/publications/pto/bilkent.html 2. The Association for Computational Linguistics 3. Special Interest Group on Computational Phonology (SIGPHON) http://www.cogsci.ed.ac.uk/sigphon/ 4. Goldsmith-Larson: a researcher whose work overlap between linguistics and computers. http://hum.uchicago.edu/~jagoldsm/Webpage/index.html 6. References Asher, R.E.(ed). 1994. The Encyclopedia of Language and Linguistics. Great Britain: Pergamon Press. Bird, Steven. 1995. Computational Phonology: A Constraint-Based Approach. Cambridge: Cambridge U.P. Bright, William(ed). 1992. International Encyclopedia of Linguistics. New York: Oxford U.P. Julie, Carson-Berndsen. 1997. Time Map Phonology: Finite State Models and Event Logics inSpeech Recognition. Kluwer Academic Publishers.Kornai, Andras. 1995. Formal Phonology. Garland Publishing.Scobbie, James. 1998. Attribute-value Phonology. Garland Publishing. |