...symbol.

When presented with a speech sound, a speaker naturally classifies it into one of the phonemes in his language. See [7] for details.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...currents.

The introduction of phonemes provides the benefit of a digital abstraction: noise immunity by nonlinear amplification.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...process.

For example, the voicing feature refers to the state of the vocal cords. If a phoneme (e.g., [z]) is pronounced with vibration of the vocal cords, the phoneme is said to be [+voice]. On the contrary, an unvoiced phoneme (e.g., [s]) is said to be [-voice]. The plus indicates the presence of voicing, while the minus indicates its absence.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...features

The only phonological assumption we make is that phonemes are represented by distinctive features. Our mechanistic model acquires constraints encoded in bit patterns. The bit patterns can be interpreted as SPE-like rules (developed in The Sound Pattern of English [3]) or as constraints that arise in the optimality framework [14].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...acquired

Modern phonology postulates more elaborate representation devices such as multiple tiers and metrical grids. See [5]. These devices describe phonological phenomena that we do not address.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...mechanisms

For example, the phonemes may be extracted from the acoustic waveform using techniques such as hidden Markov models [15] on other features, such as cepstrum coefficients. We don't know how this works. Our only concern here is that the resulting phonemes are represented in terms of some set of distinctive features similar to the SPE set.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...noted.

Constraint elements provide a mechanism of computation where there is no preferred direction of flow of information. For example, a mathematical model of a mechanical structure might include the information that the deflection d of a metal rod is related to the force F on the rod, the length L of the rod, the cross-sectional area A, and the elastic modulus E via the equation

Such an equation is not one-directional. Given any four of the quantities, we can use it to compute the fifth. See [18, 17] for more information about constraint programming.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...procedure.

The learning procedure is explained in much greater detail in [21]

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...descriptions.

Our generalization algorithm differs from the version space algorithm [10] in two respects. First, our algorithm does not maintain all the most general and most specific generalizations consistent with the current set of examples. Second, our algorithm handles disjunctive generalizations and noise.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...covered.

The generalization algorithm has a dual, the specialization algorithm, which is used to refine competing rule-classifiers that assert inconsistent values. Specialization is an incremental general-to-specific search. It aims to avoid negative examples while retaining most of the positive examples. The algorithm uses the same beam search and goodness function. It repeatedly shrinks cubes by lowering don't cares, at most two bits at a time, starting from the most recently heard phoneme.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...generalizations.

Winston [20] emphasized the usefulness of near misses in his ARCH learning program. In our program, the near misses are not supplied by a teacher or given in the input. They are generated internally.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...features.

The strident feature refers to noisy fricatives and affricates. In English there are eight stridents: [s,z,f,v,ch,j,sh,zh].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...non-continuant.

A phoneme is a non-continuant or a stop if the passage of air through the month is stopped completely for a brief period. [b,d,g,p,t,k] and the nasals [m,n] are examples of stops.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...words.

Initially we planned to use a corpus of several thousand most frequent words. But it soon became apparent that the learner can do extremely well even with a few dozen words.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...width

The beam search width is set to 2.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...presentation.

The intermediate behavior of the learner, however, is more sensitive to the order of presentation.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...non-coronal

The coronal feature refers to phonemes articulated by raising the tongue toward the alveolar ridge and the hard palate.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...first.

There is some evidence to support this prediction. It is possible that Berko's observation on the add-[I.z] rule for plural formation is not entirely robust because the same Berko subjects perform quite well in adding [I.z] to form the third person singular verbs and possessives. Of course, we cannot at this stage rule out Berko's interpretation. It might be the case that the plural formation involves mechanisms more complicated than the addition of the [s] or [z] or [I.z] ending. For instance, Pinker and Prince [12] suggests that, instead of the three rules for adding plural endings, one might have a combination of different morphological and phonological rules that produce the same pronunciation. In their account, there is a morphological rule that adds [z] to a stem. The phonetic content of the ``stem + [z]'' is then modified by two competing phonological rules. The first rule, devoicing, changes the terminal [z] to [s] under certain contexts. The second rule, vowel insertion, inserts the vowel [I] between two word-final adjacent consonants that sound too similar. According to this account, the difficulty in producing plurals like ``tasses'' may be correlated with the child's additional effort in acquiring the vowel insertion rule.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...[dc.[-tense,-strident],t,z]

The tense feature refers to phonemes produced with considerable muscular effort and a long duration.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...question.

Debate in the context of a specific problem-learning English past tense verbs-is documented in [16, 12, 11, 13, 6].

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

...system.

By ``connectionist system'' we mean a system constructed out of myriads of densely interconnected circuits where the behavior is determined by the numerical strengths of the connections.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.