One-Level Phonology: Autosegmental 
Representations and Rules as Finite 
Automata 
Steven Bird* 
University of Edinburgh 
T. Mark Ellison* 
University of Edinburgh 
When phonological rules are regarded as declarative descriptions, it is possible to construct a model 
of phonology in which rules and representations are no longer distinguished and such procedural 
devices as rule-ordering are absent. In this paper we present a finite-state model of phonology 
in which automata are the descriptions and tapes (or strings) are the objects being described. 
This provides the formal semantics for an autosegmental phonology without structure-changing 
rules. Logical operations on the phonological domain--such as conjunction, disjunction, and 
negation--make sense since the phonological domain consists of descriptions rather than objects. 
These operations as applied to automata are the straightforward operations of intersection, union, 
and complement. If the arrow in a rewrite rule is viewed as logical implication, then a phonological 
rule can also be represented as an automaton, albeit a less restrictive automaton than would be 
required for a lexical representation. The model is then compared with the transducer models for 
autosegmental phonology of Kay (1987), Kornai (1991), and Wiebe (1992). We conclude that the 
declarative approach to phonology presents an attractive way of extending finite-state techniques 
to autosegmental phonology while remaining within the confines of regular grammar. 
1. Introduction 
The decade since the publication of Koskenniemi's dissertation (1983) and since the 
development of the KIMMO system (Karttunen 1983) has witnessed a spectacular flurry 
of activity as the linguistic and computational consequences of this work have been 
fleshed out. A considerable body of literature has grown up around TWO-LEVEL MOR- 
PHOLOGY, along with texts 1 and implementations. 2 The existence of a rule compiler 
(Koskenniemi 1985) has made it possible for the linguist to work at a conveniently 
abstract level, and analyses of several languages now exemplify the approach. Today, 
two-level morphology encompasses much of traditional segmental generative phonol- 
ogy of the SPE variety (Chomsky and Halle 1968). 3 
Although further development and application of this model is set to continue for 
some time, there is now a clear need to integrate it more closely with computational 
* University of Edinburgh, Centre for Cognitive Science, 2 Buccleuch Place, Edinburgh EH8 9LW, 
Scotland, U.K. E-mail: {steven,marke}@cogsci.ed.ac.uk 
1 (Antworth 1990; Ritchie, Russell, Black, and Pulman 1992; Sproat 1992) 
2 (Bear 1986; Antworth 1990; Schiller and Steffens 1991; Pulman and Hepple 1993) 
3 Two caveats are necessary here. SPE rules must be restricted so as not to apply to their own output 
(Johnson 1972) and there is no guarantee that the transducer encoding an SPE rule can be expressed 
using the two-level rule notation (Ritchie 1992). 
(~) 1994 Association for Computational Linguistics 
Computational Linguistics Volume 20, Number 1 
grammar frameworks on the one hand and modern nonlinear phonology on the other. 
The primary goal of this article is to show how the central tenets of autosegmental 
phonology translate into an implemented finite state model. 
The model is named ONE-LEVEL PHONOLOGY for two reasons. First, the model 
is monostratal, in that there is only one level of linguistic description. Second, the 
name is intended to contrast with models employing two levels (such as the FST 
model mentioned above) or three levels (Goldsmith 1991; Touretzky and Wheeler 
1990), or an unbounded number of levels (Chomsky and Halle 1968). The one-level 
model represents the outgrowth of three independent strands of research: (i) the finite- 
state modeling of phonology, (ii) the declarative approach to phonology, 4 and (iii) the 
automatic learning of phonological generalizations (Ellison 1992, 1993). 
The paper is organized as follows. Section 2 presents an overview of autosegmen- 
tal phonology and the temporal semantics of Bird and Klein (1990). Then we define 
state-labeled automata (Section 3.1), show their equivalence to finite state automata 
(Section 3.2), define the operations of concatenation, union, intersection, and com- 
plement (Section 3.3), and further define state-labeled transducers (Section 3.4). The 
central proposals of the paper are contained in Section 4. We show how autosegmental 
association can be interpreted in terms of the synchronization of two automata, where 
each automaton specifies an autosegmental tier. We now give a brief foretaste of this 
procedure. Suppose that we have the autosegmental diagram in (1), encoding high 
(hi) and round (rnd) autosegments. 
. +hi -hi 
I/I 
-rnd +rnd 
This diagram is encoded as the following expression, where each numeral indicates 
the number of association lines incident with its corresponding autosegment. 
+hi:l -hi:2 
-rnd: 2 +rnd: 1 
From this encoding, we can write down the following regular expression. Although 
such expressions will be opaque at this early stage of the exposition, it suffices to note 
here that each line of the expression represents a tier and the tiers are combined using 
the intersection operation (m). Moreover, the ls act as synchronization marks between 
the operands of the intersection operation. 
(+hi, 0)* (+hi, 1) (+hi, 0)* (-hi, 0)* (-hi, 1)(-hi, 0)* (-hi, 1)(-hi, 0)* 
m (-rnd, 0)* (-rnd, 1) (-rnd, 0)* (-rnd, 1) (-rnd, 0)* (+rnd, 0)* (+rnd, 1) (+rnd, 0)* 
The final step is to compute the intersection and project the first element of each tuple 
(ignoring the ls and 0s). This produces the expression: 
(+hi A -rnd) + (-hi N -rnd) + (-hi n +rnd) +. 
4 Wheeler 1981; Bird 1990; Coleman 1991; Scobbie 1991; Broe 1993; Russell 1993; Mastroianni 1993. 
56 
Steven Bird and T. Mark Ellison One-Level Phonology 
Given plausible interpretations of the high and round features, this last expression 
simplifies to i+a+o +, which describes an automaton tape (or a string) divided into 
three nonempty intervals, the first containing \[i\], the second containing \[a\], and the 
third containing \[o\]. This, we shall claim, is the intended interpretation of (1). 
After a detailed discussion of this procedure, the remainder of Section 4 is given 
over to generalizing the procedure to an arbitrary number of autosegmental charts 
(Section 4.4), an evaluation of the encoding with respect to Kornai's desiderata (Sec- 
tion 4.5), and a presentation of the encoding of autosegmental rules (Section 4.6). 
Finally, Section 5 compares our proposals with those of Kay (1987), Kornai (1991), 
and Wiebe (1992). While our model has regular grammar power and is fully imple- 
mented, these three models go beyond regular grammar power and to our knowledge 
have never been implemented. 
2. Background 
It has long been recognized that the SPE model lacks explanatory adequacy, a fact 
noted in SPE itself (Chomsky and Halle 1968, pp. 400ff). For example, it is unable to 
explain why a final devoicing rule like that in (2a) is commonplace in the languages 
of the world, whereas the rule in (2b) is unattested (Kaye 1989, p. 61). 
2. (a) 
(b) 
(c) 
(d) 
\[-sonorant\] --* \[-voice\] / --# 
\[-sonorant\] --+ \[+nasal\] / --# 
\[\] --* \[anasal\] / \[anasal\] --# 
\[\] ~ \[around\] / \[anasal\]- # 
Similarly, the nasal harmony rule in (2c) occurs frequently, while the generalization 
expressed in (2d) is as unlikely as (2b). Both SPE and the two-level model are unable 
to express the fact that some rules are commonplace while others are highly unnatural. 
Perhaps the SPE model could be rescued from these problems with additional 
stipulations. However, a more fundamental problem for the model was raised by the 
following tone language data (Leben 1973; Goldsmith 1976). At first blush, Mende 
vowels appear to manifest five tone patterns, namely high (ko), low (kph), falling 
(mbfi), rising (mb~) and rise-fall (mbS). The SPE model would predict 25 tonal patterns 
for two-syllable morphemes, but instead we find only 5. These are high-high, low-low, 
high-low, low-high, and low-falling. Similarly, for three-syllable morphemes we get 5 
patterns, not 125. Leben noticed that the one-, two-, and three-syllable morphemes 
could be put into correspondence as shown in (3), from Leben (1978). 
. H: k5 war p~l~ house hhw~m~ waistline 
L: kpa debt b~l~ trousers kpak/tll tripod chair 
HL: mbfi owl ngflh dog f61amh junction 
LH: mbh rice fhnd6 cotton nd/~vfil~i sling 
LHL: rob8 companion nyhh~ woman nikili groundnut 
Goldsmith (1976) devised a graphical notation that made the above correspon- 
dence clearer still. We display several examples of his notation in (4). Here, the H 
indicates high tone, while L indicates low tone. (The association lines are assumed to 
be incident with the vowels.) 
57 
Computational Linguistics Volume 20, Number 1 
. k6 1 ~ 1£ h~i w~i mfi 
I/ \/ 
H H H 
kpa b~ lP kpa kh li 
E/ \l/ 
L L L 
mbfi ngi 13 f6 1~ mh 
\ I I I I/ 
H L H L H L 
mb~ fh nd6 ndh vti lfi 
L H L H L H 
L 
mb~ nyh ha ni kl Ii 
/1\ I\ I I 
H L L H L L H L 
Observe that each row of the table has the same tone pattern. Only the synchro- 
nization varies. In diagrams like the ones in (4), units such as H and L are termed 
AUTONOMOUS SEGMENTS, or AUTOSEGMENTS, and a linear sequence of autoseg- 
ments is called a TIER. The synchronization markers are called ASSOCIATION LINES. 
A pair of tiers linked by some association lines is called a CHART. A chart is called 
WELL-FORMED if the following conditions hold (Goldsmith 1976, p. 27). 
5. Well-Formedness Condition: 
(a) 
(b) 
All vowels are associated with at least one tone; 
all tones are associated with at least one vowel. 
Association lines do not cross. 
The reader can ascertain that the above charts are well-formed according to (5). How- 
ever, (5) is insufficiently restrictive on its own, and a further stipulation is required. 
, Association Convention: Only the rightmost member of a tier can be 
associated to more than one member of another tier. 
58 
Steven Bird and T. Mark Ellison One-Level Phonology 
When (5) and (6) are combined, we achieve the effect of one-to-one left-to-right 
association, where multiple association (or SPREADING) occurs only at the right-hand 
end. Observe also that the charts in (4) do not contain adjacent identical tones. For 
example, there is no HH tone melody. This is expressed by a principle attributable to 
Leben (1973). 
. Obligatory Contour Principle: At the melodic level of the grammar, any 
two adjacent \[autosegments\] must be distinct. Thus HHL is not a 
possible melodic pattern; it automatically simplifies to HL. 
Although nonlinear models like autosegmental phonology represent a major ad- 
vance on the linear model of SPE in the area of explanatory adequacy, it has sometimes 
been pointed out (e.g., Bird and Ladd 199l) that the formal explicitness of the SPE 
model has not been matched by these more recent proposals. Before we can begin 
to compute with autosegmental representations and rules, they need to be given a 
formal semantics. Our starting point here is the temporal semantics of Bird and Klein 
(1990), based on Sagey's (1988) model, which has gained widespread acceptance in au- 
tosegmental phonology. Under this temporal semantics, phonological properties are 
attached to intervals that are related using precedence (an asymmetric, transitive re- 
lation) and overlap (a reflexive, symmetric relation). 
Bird (1990) showed how a phonological description language can be modeled by 
such event structures, where the precedence relation models the linear ordering of 
tiers and the overlap relation models association lines. In this paper, we shall pro- 
vide an automaton-based semantics for precedence and overlap, thus arriving at a 
computational semantics for the autosegmental notation. 
3. State-Labeled Automata 
In this section we give definitions for a new device called a state-labeled finite automa- 
ton, and then we define various useful operations on these automata. (Some readers 
may prefer to skip Section 3 on a first reading.) 
3.1 Definitions 
Definition 1 
A STATE-LABELED NONDETERMINISTIC FINITE AUTOMATON (SFA) is a septuple 
(V~ ~, ~, 6~ S, F, e) where 
V is a finite set, the set of STATES, 
~. is a finite set, the ALPHABET, 
C V x ~, is the LABELING RELATION (states are labeled with subsets of the 
alphabet), 5 
6 c V x V is the TRANSITION RELATION, 
S C V is the set of START STATES, and 
F c V is the set of FINAL STATES. 
e is a Boolean flag that is true iff the null string A is accepted, and false otherwise. 
5 Without loss of generality we have chosen to label states with segments (subsets of ~), rather than with strings (subsets of G*). 
59 
Computational Linguistics Volume 20, Number 1 
Before describing the execution of an SFA, we need to define COMPATIBILITY. 
Definition 2 
We say that a state v is COMPATIBLE with an input cr E G if (v, c~ / E )~. 
At each step in the execution of an SFA, a subset of the states is active while the 
remainder are inactive. An SFA begins operation with those start states active that are 
compatible with the first symbol in the input string. If, at a certain step in processing, a 
subset T of states is active, then at the next step, the subset T t of states reachable from T 
and compatible with the next input become active. This operation is formalized below, 
following the approach taken by Partee et al. (1990). First we define a SITUATION to 
be the processing status of an SFA. 
Definition 3 
A SITUATION of an SFA, A, is a triple (x, T,y I where T c V is the set of currently 
active states, and x and y are the portions of the input string to the left and right of 
the reading head, respectively. 
As an SFA operates, it moves through a sequence of these situations. Now F-A is defined 
as a successor relation on situations for an automaton A. 
Definition 4 
Let Ix, T, y I and Ix', T', y' I be two situations. Then Ix, T, y I }-"A IX'~ T', y'I iff there is a 
E G such that 
(i) 
(ii) 
(iii) 
y = cry ~ and x ~ = xG 
for each Y E T I there is a t E T such that (t, t' I E 6, and 
(t', cr / E )~ for each t' E T'. 
The first condition in the definition concerns or, the tape symbol being scanned. This 
symbol is the first in the string y and the last in the string x'. The second condition 
concerns the transition relation, requiring that the new situation must be reachable 
from the previous situation. The third condition is a check that ¢ is in the label set of 
each currently active state. We define t-~ to be the transitive closure of ~-a. Now we 
can specify the conditions under which an SFA accepts a string, where A is the empty 
string. 
Definition 5 
Let A = (V,G,A,6, S,F,e) be an SFA and let w E G*. A ACCEPTS w iff either e is true 
and w = A, or (or, {s}, a) t-~ (era, F', A/, for some (s, cr / E ;~, s E S, a E ~* and FMF' ~ O, 
and where w = era. 
If no string can cause an SFA to have more than one active state at any processing step, 
then we say that the SFA is DETERMINISTIC. In order to signify that an SFA accepts 
the empty string A (i.e., if e is true), we shall include a special state, labeled with a 
distinguished symbol 0, which is marked both initial and final. 
60 
Steven Bird and T. Mark Ellison One-Level Phonology 
3.2 Relationship to FSAs 
The equivalence of SFAs and arc-labeled finite-state automata (FSAs) follows from the 
equivalence of Mealy and Moore machines. 6 Although SFAs are no more expressive 
than FSAs, there are good linguistic reasons for wishing to use them. The primary 
difference between the two devices lies in the relative ease with which particular gen- 
eralizations can be expressed. As an illustration, we shall consider the automaton that 
prohibits two adjacent occurrences of any given symbol in a string, a constraint known 
in autosegmental phonology as the Obligatory Contour Principle. Here is the FSA ver- 
sion, using the graphical conventions for representing FSAs adopted by Hopcroft and 
Ullman (1979). Bullet marks a state, circled states are final, and states with incoming 
arrow heads are initial. 
d a -- 
The notation we use for state-labeled automata is different. Because states carry labels, 
there is no need to use the contentless bullet symbol to mark a state. Instead, the label 
itself marks the state. Initial and final states are indicated by arrowheads and circles, 
as for FSAs. Arcs are unlabeled. 
The SFA that does not admit adjacent occurrences of any symbol in the alphabet 
is: 
Notice that the number of labels required by the SFA is considerably smaller (5 labels) 
than the number required in the FSA (12 labels), while the number of transitions is 
the same for both. On the other hand, the SFA has an extra state, labelled with 0 to 
show that the automaton accepts the null string (see the final clause of Definition 1). 
The difference between the two devices becomes even clearer when we consider 
the representations for a* (zero or more as) and a + (one or more as). First, here are a* 
and a + expressed as FSAs. 
4c a a* : a + 
: >'e a ~" 
Observe that the specification of the Kleene star requires one state, while Kleene plus 
requires two. If we use SFAs instead, we find the reverse: Kleene star requires two 
6 See Hopcroft and Ullman (1979, p. 44) for a discussion of this equivalence. An FSA is a Mealy machine 
that ignores its input, while an SFA is a Moore machine that ignores its input. 
61 
Computational Linguistics Volume 20, Number 1 
states, while Kleene plus only requires one. 
a*: >(~ >~ a+: ~~-~ 
As we shall see in Section 4, the semantics of phonological representations requires 
frequent use of the Kleene plus and little use of the Kleene star. The intuition behind 
this is simple. Recall from Section 2 that phonological entities such as distinctive fea- 
tures are considered to be descriptions of phonetic events that may extend over an 
interval of time. As we saw in the case of tone, the defining feature of an autosegment 
may be spread across several segments. Crucially, however, an autosegment must be 
present at at least one point, and so it makes sense to view phonological entities--such 
as segments and autosegments--in terms of the Kleene plus rather than the Kleene 
star. 
It might be reasoned that our interval interpretation of segments is better pictured 
with arc-labeled devices. After all, the states resemble points while the arcs resem- 
ble extended intervals. Furthermore, it may be tempting to use a single arc between 
two temporally distant points to show the spreading of an autosegment coarticulated 
with two consecutive autosegments on another tier. For example, (8) shows a labial 
autosegment, L, bridging two instants also bridged by a nasal segment, N, and stop, S. 
. 
>"o- N>o S > 
However, this representation is flawed: the semantics assigned to coterminous paths 
contradicts the standard interpretation of FSAs. The automaton pictured above would 
normally be interpreted as either a nasal followed by a stop or by a single labial 
articulation. Crucially, it could not be interpreted as necessarily both a nasal followed 
by a stop and a labial articulation. 
While viewing states as instants of time and arcs as intervals offers some iconicity, 
it is also misleading. It is just as natural--and more in keeping with the logical founda- 
tions presented in Section 2--to employ the states for temporally extended intervals, 
and the arcs for the relationship of immediate precedence. 
3.3 Basic Operations 
In this section we define the operations of concatenation, union, intersection, and 
complement on SFAs. These operations correspond naturally to the operations on the 
languages accepted by the automata, as indicated in the following table. 
operation 
concatenation 
union 
intersection 
complement 
Kleene plus 
Kleene star 
automata languages 
AB 
AUB 
A\[TB 
-d 
A + 
A* 
(ab\]a L(A), b E L(B)} 
U 
L(A) n r(B) 
L(A) 
L(A) + 
L(A)* 
62 
Steven Bird and T. Mark Ellison One-Level Phonology 
The Kleene plus operation, which, when applied to a language L gives another L +, 
contains the concatenation of one or more strings from L. The Kleene star operation 
takes L to {A} U L + and is written L*. 
Recall that the structure (~*; U, N, --, 0, G*), containing languages over an alphabet 
Y. together with the standard set operations, is a Boolean algebra (Partee et al. 1990, 
p. 297ff). Similarly, if A is the set of SFAs, then (A; U,  , 4, 3_, T) is also a Boolean 
algebra, where 3_ is the empty automaton (i.e., L(3_) = 0) and T is the automaton that 
accepts G* (i.e., L(T) = y~$k).7 
The concatenation of two SFAs A and B, written AB, has an arrow linking each 
final state of the first SFA to each initial state of the second. The states that are initial 
or final in AB depend on whether A or B accepts the empty string A, as specified in 
the following table. 
A 
A (t L(A) 
A E L(A) 
A ~ L(A) 
A C L(A) 
B 
A EL(B) 
A L(B) 
A L(B) 
A C L(B) 
AB 
A (d L(AB) 
A ~ L(AB) 
A (d L(AB) 
A E L(AB) 
AB initials 
A initials 
A & B initials 
A initials 
A & B initials 
AB finals 
B finals 
B finals 
A & B finals 
A & B finals 
Suppose we wished to recognize the language AB where A -- (12+3) + and B = 
(45+6) +. The SFAs describing A and B are the following. 
~- 2 ~ ~- A: 1 ~ B: 4 \! 
® 
Linking final states of A to the initial states of B gives the concatenation. 
AB : 1 4 \l \l 
3" ® 
Since neither A nor B accepts A, the initial state of AB is the initial state of A, and the 
final state of AB is the final state of B. 
The union (or disjunction) on SFAs is similar, in many ways, to the concatena- 
tion operation. The difference is that rather than executing in sequence, the automata 
operate in parallel. The union of two automata A and B, written A U B, accepts the 
string s iff either A or B, or both, accept s. The union of A and B is expressed dia- 
7 of course, these algebras are different, for there are many languages that cannot be defined by SFAs. 
63 
Computational Linguistics Volume 20, Number 1 
grammatically by placing the diagram for A alongside the diagram for B. The union 
(12+3) + U (45+6) + is drawn as: 
AUB: 1 
® 
~ 4 ~5 ~ 
® 
The intersection of two SFAs A and B, written A n B, accepts a string s iff both A 
and B accept s. Consider the intersection of the automata that recognize the languages 
(12+3) + and (1+23+) + . These two automata are: 
A: 1 ~ B: >2 
Two states are compatible if and only if their label sets have a nonempty intersection. 
The intersection of these automata is formed by taking all pairs of states that are 
compatible and linking these with arcs whenever both projections of the pairs are 
linked. The intersection of the above automata is shown below. 
AnB: 1 72 
® 
This automaton recognizes the language (123) + . 
The complement of an automaton A, written A, accepts a string s iff A rejects s. 
One way of forming the complement involves the following steps. First, the automaton 
must be DETERMINIZED (Hopcroft and Ullman 1979, pp. 22ff). The next step is to form 
the COMPLETION. A complete automaton is one that has a transition from every state 
for each element of G. The final step is to mark all final states nonfinal, and all nonfinal 
states final. So if S is the set of states and F is the set of final states, then S\F is the set 
of final states in the complement. 
3.4 Transducers 
In the previous three sections, we defined state-labeled automata and some operations 
that can be used to combine them. We also saw that these automata are equivalent 
64 
Steven Bird and T. Mark Ellison One-Level Phonology 
to traditional, arcqabeled automata. Just as we can define arc-labeled automata called 
finite-state transducers (FSTs), we can define state-labeled transducers (SFFs). 
An (epsilon free) state-labeled transducer is just an SFA with a special alphabet. 
Instead of labeling each state with a subset of a single alphabet, we label them with 
subsets of the product of two alphabets. The strings accepted are sequences of pairs 
consisting of one letter from each alphabet. A transducer can be used as a translator: 
it takes as input one half of the label on a state, and simultaneously writes as output 
the other half. All output strings generated by a path from initial to final states are 
translations of the input string recognized by the same path. Since the SFT is also an 
SFA, intersection is defined for SFTs. 
The reader may wonder whether there is any distinction between one-level phonol- 
ogy and two-level phonology if SFAs and SFTs are formally identical. There is an im- 
portant distinction to be drawn, however. First, most two-level models employ FSTs 
with epsilons, which are more powerful devices than FSAs. Second, in the one-level 
model, representations and rules are interpreted as automata. In contrast, the two- 
level model employs strings for representations and automata for rules. Finally, in 
one-level phonology surface forms and generalizations about them are stated directly 
in a hierarchical lexicon akin to that of head-driven phrase structure grammar (HPSG) 
(Pollard and Sag 1987), rather than being mediated through a transducer (Bird and 
Klein, in press). 
4. Association and Synchronization 
In this section we present the automaton-based semantics for autosegmental phonol- 
ogy. 
4.1 The Representation of Autosegments and Tiers 
Recall that an autosegment denotes a possibly extended interval. In terms of automata, 
this means that an autosegment must allow multiple copies of its defining property. 
This is expressed as follows. 
This state of affairs fits well with our intuitive understanding that a pair of adjacent 
intervals in which some property holds is indistinguishable from a single interval---the 
union of the first two intervals--during which that same property holds. Furthermore, 
such a claim connects with the Obligatory Contour Principle (7). 8 
Unfortunately, however, this definition of autosegment is inadequate. Suppose we 
have a nasal segment N that is lexically unspecified for its place of articulation. In a 
language with the nasals m and n, the intention is that this segment denotes intervals 
such as the following: 
Imlm\]m\[mlml In\[n\]n\]n~ 
However, there is nothing to stop N from denoting the following interval: 
Imrnrnlm\[nJmf 
8 A consequence of this approach is that it circumvents some potential problems caused by our not 
employing epsilons (cf. Section 3.1). If an autosegment alternates with zero, we do not employ ¢ for the zero alternant but permit surrounding autosegments to extend to 'fill in the gap.' 
65 
Computational Linguistics Volume 20, Number 1 
Here, it is clear that N is behaving like a variable. If N is instantiated, then the 
entire interval must remain homogeneous. So the representation of this N is not (9a) 
but (9b). 
9. (a) 
(b) 
Some autosegments, however, appear to lack this homogeneity property. For ex- 
ample, the Turkish word pe~eleri contains several front vowels. An analysis of such 
words that employs the principles of vowel harmony posits a +front autosegment that 
is associated with every vowel of the word. Observe that this autosegment is heteroge- 
neous, as it includes in its temporal extent both e and i. Accordingly, its interpretation 
will follow the scheme of (9a) above. Briefly continuing in this vein, we can conceive 
of a range of different kinds of segment, using varying numbers of states and varying 
numbers of labels per state. Four options are described below: 
Simple Segments. These capture the ordinary kind of segment and consist of a 
single state labeled with a singleton set. In general, when we employ a 
symbol like b it will be interpreted as a simple segment. 
Homogeneous Segments. These represent slots (like N) and members of tem- 
plates (like CVCCVC), and consist of more than one state. Each state is 
labeled with a singleton set. An example of a homogeneous segment is 
found in (9b). 
Heterogeneous Segments. These represent spreading autosegments, like +high 
(and b in Section 5.4). The automata have a single state which is labeled 
with a nonsingleton set. An example of a heterogeneous segment is found 
in (9a). 
Hybrid Segments. These represent spreading autosegments that have greek letter 
variables, like c~place or ~high. An example of a hybrid segment for odront 
is given in example (10). 
10. 
Recall that an autosegmental tier is just a linear ordering of autosegments. There- 
fore, if P, Q and R are segments (of any of the four kinds specified above), then a tier 
P-Q-R is formed by simply concatenating P, Q, and R together. 
Now we have seen the interpretation of autosegments and tiers. The synchronization 
of tiers is controlled by association lines. The next section discusses the interpretation 
of these lines. 
66 
Steven Bird and T. Mark Ellison One-Level Phonology 
4.2 The Interpretation of Association 
In Section 2 we presented an interpretation of association based on temporal overlap. 
Now we must find a way of simulating this temporal structure using automata. Let 
us begin by considering the simplest possible autosegmental diagram. 
11 .... A ... 
• • • B • • • 
Since each autosegment denotes an interval and the two intervals must overlap, we 
would like to interpret the above diagram as describing any of the following strings, 
among others: 9 
IA A A " " {AIAIA AIAI " " A 
• • B B B • • B B • B B B 
What kind of automaton will give us the required behavior? The clue is that a pair 
of intervals overlap if and only if they share a point in common (Bird and Klein 1990, 
p• 36)• Note that in each of the above diagrams, the third interval contains an instance 
of both A and B, and the existence of this interval was both a necessary and sufficient 
requirement for the association line to have its overlap interpretation. So we need an 
automaton for each autosegment that captures the two required properties: (i) deno- 
tation of an extended period, and (ii) existence of a special point. The "automaton" 
required for the autosegment A is given in (12). 
12. 
This automaton accepts any string of one or more As, and requires that there is some 
A that is coincident with an autosegment somewhere else. The line extending from 
the middle state informally indicates that this state is simultaneous with a state in 
another automaton. Now we must create a similar automaton for the autosegment B. 
Recall from Section 4.1 that we need to construct tiers for A and B by concatenating 
the autosegments to the left and right (elided in (11))• Finally, the automata are put 
together as shown below: 
• •• ~- + •.. 
o.• -}- + •.. 
9 Note that in using such diagrams we are attempting to isolate the effects of two automata: those cells containing both A and B should be understood as containing A N B. 
67 
Computational Linguistics Volume 20, Number 1 
This unusual kind of automaton represents a halfway house between autosegmental 
diagrams and SFAs; we shall call it a SYNCHRONIZED SFA. Such automata have a 
simple interpretation: both component automata run in parallel, but the second state 
of the A automaton is active iff the second state of the B automaton is simultaneously 
active. In a sense, we have converted an autosegmental model involving intervals 
and overlap into a simpler model involving atomic periods and simultaneity. Now 
consider autosegmental diagram (13). 
13. A B 
\/ 
C 
Note that there is no additional material on either side of the group of three autoseg- 
ments (cf. (11)). Therefore, we assume that both tiers are descriptions of a complete 
utterance and so must begin and end simultaneously. 1° In this diagram, C has two 
associations. The automaton we need for C is more complex than before, since the 
interval that C denotes must have two special points, one for overlap with A and the 
other for overlap with B. The required automaton for C is given in (14). 
14. 
I 
Automaton (14) is the concatenation of two copies of an automaton like (12). Putting 
the automata for A, B, and C together gives the synchronized automaton in (15). 
15. 
A~A ~ A~-~ ~~--------~ 
The behavior of synchronized SFAs can be simulated by an ordinary SFA, as we show 
in the next section. 
4.3 Simulating Synchronized Automata 
To do this simulation, we need to do away with the synchronization lines connecting 
the automata. Starting with (15), we add indices to each state: a 0 to unsynchronized 
10 It is a trivial matter to do away with this restriction, since we can always add a completely unspecified 
autosegment to the start and end of each tier, thereby permitting slippage between the substantive 
material on each tier. 
68 
Steven Bird and T. Mark Ellison One-Level Phonology 
states and a 1 to synchronized states, and then erase the lines. 
( 
>(A~>{A, 1} , (A~/B 
Observe that each of these state labels is actually a pair. These automata are defined 
over the product alphabet G x {0, 1}. The intersection of these automata is as follows: 
~ANC'0) 7 \[ANC' 1} 4.. ~" (ANC'0} 7 (BN~' 4,. 
The function of the indices was to rule out certain states in the intersection. Now that 
the indices have served their purpose, we can erase them and further simplify the 
automaton: 
This, then, is the semantics assigned to (13). Since we no longer require the graphical 
notation for synchronization, it will be convenient to represent SFAs using the notation 
of regular expressions over ordered pairs. In order to do this it is useful to employ 
some macros. An autosegment A is represented as s(A) =def /A~ "/q-. Bullet (e) is used 
here as a context-dependent wildcard, indicating an alphabet. (In this definition of 
s(A), the bullet indicates the alphabet {0,1}; in the definition of a(n) below it indicates 
the alphabet G.) The n association lines incident to an autosegment are expressed as 
follows: 
a(n) =def /'~ 0/* (/'~ 11 /', 0/*) n 
Observe that a(m + n) = a(m) +a(n). Finally, we combine these two macros into a third 
macro thus: 
A:n =def s(a) ha(n) 
This states that A is an autosegment with n associations. Now we shall illustrate the 
workings of these definitions. Consider diagram (13) again, reproduced below: 
69 
Computational Linguistics Volume 20, Number 1 
13. A B 
\/ 
C 
A singly associated autosegment, such as A, is written down as the following formula: 
A:I = s(A) Ma(1) 
-- (A,.) + M (.,0)*(-,1)(-,0)* 
= (A, 0)* (A, 1) (A,0)* 
We can now simply write down a formula for (13): 11 (A:I+ B:I) M C:2. This expression 
evaluates to the following: 
(ANC,0/* (AnC,1) (A nC,0)* (B n C,0)* (B n C,1) (B n C,0)* 
Now that the indices have served their purpose, we would like to ignore them by pro- 
jecting the ordered pairs onto their first element. The result of evaluating this projection 
for our current example is (An C)+(B n C) +, which is the intended interpretation of 
diagram (13). We shall adopt the following notational convention: if D is the encoding 
of a diagram then \[D\] is the projection of the encoding that ignores the indices. 
As another example, consider an autosegmental diagram consisting of two tiers, 
each with two autosegments, and two association lines between the tiers, as shown 
in (16). 
16. A B 
C D 
The expression for (16) is (A : 1 B : 1) M (C : 1 D : 1). This evaluates to the follow- 
ing expression under the projection: (A N C)+((A N D) + t3 (B n C) + u ~)(B N D) +. The 
corresponding automaton for (16) is given in (17). 
17. 
Cc\, / 
11 It is important to notice that the numerals in this expression are not the same as the indices that occur 
as the second member of pairs like (A, 1). The former represent degree of association, while the latter function as synchronization marks in an automaton. 
70 
Steven Bird and T. Mark Ellison One-Level Phonology 
Automaton (17) will accept the following sequences, among others: 
A B IAIBrBI IArArB r C D C C D C D D 
Note that all of these examples have A overlapping C and B overlapping D. The 
second and third examples have an extra cell, for B overlapping C and A overlapping 
D, respectively. This range of possibilities is compatible with (16), as it requires an A-C 
overlap and a B-D overlap and optionally permits an A-D overlap or a B-C overlap 
(but not both). 
We have now seen an automaton-based interpretation of an autosegmental chart. 
Next we consider how the above regular expressions defined over ordered pairs can 
be generalized to representations consisting of more than one chart. 
4.4 Multiple Charts 
It is straightforward to generalize the interpretation procedure for single charts to one 
for representations with an arbitrary number of charts. Recall that for one chart we 
needed to employ ordered pairs. In general, for n charts we must employ ordered 
n + 1-tuples. The construction will be demonstrated using a diagram attributable to 
Pulleyblank (1986, p. 13) involving three tiers and three charts. 12 
18. A B C 
2 E F 
Now, since there are three charts in (18) we must employ 4-tuples. We adopt the 
following abbreviatory conventions: 
s(a) 7--de f (/:/, ,, ,, ,)+ 
a12(n) =def (',0,',')* ((,, 1,,,,)(,,0,,,,)*) n 
a23(n) =def (',',0,')* ((,,,, 1,,)<',',O,o)*) n 
a13(H) =def (-,-, .,0)* ((', ", ", 15 (',-,., 0)*)" 
A:p:q:r =def s(A) nat2(p)na23(q)na13(r) 
We can now write down the expression for (18) as follows: 
19. (A:I:0:0 B:0:0:I C:1:0:0)~(E:2:0:0 F:0:1:0)~D:0:1:1 
The first three terms of this expression correspond to tier 1 of (18). The first term of 
the expression concerns the autosegment A and its association line on chart 1-2. The 
second term concerns B and its line on chart 1-3. Notice that lines AE and BD are in 
12 Pulleyblank observes that the temporal interpretation of this diagram is ill-defined if association is 
assumed to be transitive, However, since overlap is not a transitive relation we do not have this 
problem. 
71 
Computational Linguistics Volume 20, Number 1 
different charts in (18) and so the count of association lines is in a different position in 
the 4-tuple for A and B. The third term concerns C and its line in chart 1-2. The fourth 
and fifth terms of the expression concern the E-F tier. Since E is doubly associated in 
chart 1-2, a 2 is used for E's degree of association. D also has two associations, but they 
are in different charts. Expanding and projecting this expression gives the following: 
(AnDnE)+(BnDnE)+(CnDnE)+(CnDnF) + 
So the shortest string that satisfies the constraints expressed in diagram (18) is the 
following: 
A B C C 
D D D D 
E E E F 
Recall that the encoding of the three-tier diagram (18) was given as expression 
(19). In such expressions it is difficult to identify tiers. Therefore, in the remainder of 
this paper we shall write expressions like (19) in the following format: 
20. tier1 A:I:0:0 B:0:0:I C:1:0:0 
tier2 E:2:0:0 F:0:I:0 
tier 3 D:0:1:1 
In general, if D is a diagram, then C(D) is its encoding in the format exemplified 
in (20). 
4.5 Evaluating the Encoding 
Now, we evaluate our encoding with respect to Kornai's desiderata: computability, 
compositionality, invertibility, and iconicity (Kornai 1991). 
Computability: The number of terms in the encoding is equal to the number of 
autosegments, and each term has a fixed size} 3 Therefore, the encoding 
can be computed in linear time. 
Compositionality: If D1 and D 2 are two autosegmental diagrams then E(D1D2) = 
g(D1)E(D2), where concatenation of encodings is done in a tier-wise man- 
nen Thus the encoding is compositional. 
Invertibility: A representation can be reconstructed from its encoding. 
Iconicity: If an autosegment in a diagram is changed, the effect on the encoding 
is local, since only one term is altered. However, if an association line 
is added or removed, two terms must be altered. Although these terms 
may not be adjacent, we believe the encoding is more iconic than Kornai's 
triple code (see Section 5.2), in which changes can affect an unbounded 
amount of material. 
In addition to these properties, our encoding can be used directly as a finite-state 
recognizer of surface forms, simply by forming the intersection of the n tier encodings 
13 Observe that when we expand these macros the resulting expression has s + 2a terms, where s is the 
number of autosegments and a is the number of associations. 
72 
Steven Bird and T. Mark Ellison One-Level Phonology 
and projecting the first elements of the tuples. Note that if we revert to the encoding in 
(19), where the tier encodings are combined into a linear expression using intersection, 
then compositionality is lost. Thus, the encoding is either linear or compositional, but 
not both. Unfortunately, this is the best that we can hope for; Wiebe (1992) has shown 
that a compositional linear encoding does not exist. 
4.6 Phonological Rules 
Phonologists typically encode their descriptive generalizations in terms of RULES. 
Often these rules are interpreted as processes that manipulate representations. In the 
one-level approach they are interpreted as a logical implication between two descrip- 
tions, which simplifies to a single description given the Boolean operations presented 
in Section 3.3. 
As our first example, consider the phenomenon of homorganic nasal assimilation, 
whereby nasals agree in place of articulation with the following consonant. An SPE- 
style rule for this is given in (21a), the corresponding logical implication in (21b). 
21. (a) 
(b) 
\[+nasal\] --* \[c~place\] / __ \[+cons,~place\] 
\[+nasal\]\[+cons\] --+ \[c~place\]\[~place\] 
Thus, the sequences mb and nd are allowed, while md and nb are ruled out. Let 
N -- {m,n}, S = {b,d}, L = {m,b}, and A = {n,d}. The required constraint can be 
expressed as NS --* LL U AA. However, in order to make this rule apply to a whole 
string (rather than just the first NS sequence it comes across), we must express it in 
the following format. 14 
22. -~(o*(NS F1LL U AA)o*) 
This states that it is not possible to find anywhere a nasal-stop cluster (NS) that is 
not made up of two labials (LL) or two alveolars (AA). We can simplify the above 
expression to .*(mA).* f\] .*(nL),*. 
Now consider a general rule of the form SD ~ SC. Since SD and SC pertain to 
parts of a string rather than a whole string, we have to ensure that the rule applies to 
all substrings of an 'input' string S. We do this as follows: 
23. Vs c s, SD(s) ~ SC(s) 
Vs C_ S, ~SD(s) V SC(s) 
~s c s, SD(s) A -~SC(s) 
-4,*(SD n SC),*) 
This is how we arrived at (22) from (21b). 
Autosegmental rules can also be expressed in this framework. Consider again the 
case of assimilation. The following diagram is the autosegmental rule corresponding 
to the SPE rule in (21). Here, ~place is a hybrid autosegment (see Section 4.1) ranging 
over places of articulation. 
14 Note that we employ a '9" sign or an overline to represent complement, depending upon which is most convenient. 
73 
Computational Linguistics Volume 20, Number 1 
24. N C 
\ 
\ 
c~place 
This rule states that wherever an NC sequence can be found, if the C is associated 
with an L, then the N is also associated with L. We can express this rule in the more 
familiar rewrite notation: 
N C --* N C 
\ 
~place c~place 
We can give an automaton-based semantics to this rule. In order to do this, we must 
employ two independent charts between the two tiers, one for the structural description 
and one for the structural change. We can represent this as follows, where 'sd' refers to 
lines in the structural description chart, and 'sc' refers to lines in the structural change 
chart. 
N C 
sd 
~place 
Now we can write down the formulas for the structural description and the structural 
change independently and combine them into the rule format of (23). 
25 N0 c .0*  .00" N01 c10.00"1 / .. 
e:0* ~place:l e:0* U-~ e:0:0* c~place:l:l e:0:0* sc sd 
In (25) the 'sd' and 'sc' subscripts on the brackets refer to projection functions that 
ignore the structural description and structural change charts, respectively. So, in eval- 
uating (25) we intersect the two tiers of the structural change part of the rule, and then 
delete the second index of each tuple. The complement of this automaton is then inter- 
sected with the structural description part of the rule and the first index of each tuple 
is then deleted. The final step is to add the e* wildcards and form the complement 
again. The result is an automaton that rejects any nonhomorganic NC clusters. 
A more complex example of a phonological rule, this time concerning vowel har- 
mony, will now be discussed. Since the advent of autosegmental phonology, vowel 
harmony has been analyzed as the spreading of autosegments from left to right. Here 
we show how such a rule can be translated into a regular constraint on surface forms. 
Turkish exhibits two orthogonal types of vowel harmony: one requiring that con- 
secutive vowels agree in fronting, the other requiring that consecutive vowels agree 
in rounding unless the second vowel is low. As the rounding harmony is the more 
complex of the two, we will take it as our example. To avoid the complications of 
fronting harmony, we will only consider examples involving back vowels. 
74 
Steven Bird and T. Mark Ellison One-Level Phonology 
Turkish has eight vowels. Four of them (a e i i) are unrounded, and four (o 6 u 
ii) are rounded. Turkish is an agglutinating language, and the vowels in many affixes 
depend on the final vowel in the root of the word. Compare, for instance, three of the 
cases of the words son end and adam man displayed in (26). 
26. case 
nominative 
accusative 15 
dative 
son adam 
son adam 
sonu adam1 
sona adama 
When the rounded root vowel is followed by a high vowel in a suffix, this vowel must 
agree with the root in rounding. Low vowels in harmonising suffixes are, however, 
always unrounded. 
The notation of autosegmental phonology makes it easy to state this generalization 
as a rule. The rule spreads a rounding autosegment onto the next vowel if (and only 
if) the next vowel is high. 
27. +rnd 
\ 
\ 
V V 
\[+ hi\] 
This rule only applies on the vowel tier, skipping over consonants, and the in- 
terpretation given to autosegments on this tier must reflect this. We use the idea of 
heterogeneous autosegments (see Section 4.1) to differentiate ordinary segments from 
autosegments on restricted tiers such as the vowel tier. A vowel on this tier may denote 
not only a vowel but also consonants interspersed within this vowel as well. Whereas 
a segmental a corresponds to an automaton with a single state having a singleton label 
(this is what we have called a SIMPLE segment; see Section 4.1), an a on the vowel 
tier corresponds to a single state automaton whose state accepts not only a but any 
consonant as well (a HETEROGENEOUS segment). 
Given that the vowel tier uses this kind of heterogeneous representation, we can 
combine the charts of our rule into the regular expression in (28). 
28  rndl "°* "°°*  rndll "°°* J J /.* e:0* V:0 +hi:l e:0* e:0:0* V:0:I +hi:l:0 e:0:0* sc sd 
Using the procedure of Section 4.6, we can translate this into a regular expression that 
15 More precisely, the accusative case is used only for definite objects. 
75 
Computational Linguistics Volume 20, Number 1 
implements rounding harmony as a constraint on surface forms. After simplification, 
this automaton is: 
@ 
Each of the states accepts the specified class of vowels and any consonant. Informally, 
this automaton will accept any sequence of vowels except those in which a round 
vowel is followed by an unrounded high vowel. This is precisely the intended effect 
of the autosegmental version of the harmony rule (27). This concludes our discussion 
of vowel harmony. 
A comment is in order here about why two charts were required for the encoding 
of autosegmental rules. After all, our use of an input and an output chart appears to be 
a procedural device, and we have eschewed these from the outset. However, note that 
it is not possible for our structural description and structural change to be encoded 
on the same chart: the structural change has an association line not present in the 
structural description, and so they are incompatible. Of course, the overlaps described 
by the structural description and structural change are mutually compatible; it is only 
the associations that are not. Once the structural change has been computed, we can 
throw away the 'sc' chart, and once the structural description and structural change 
have been combined, we can also throw away the 'sd' chart. In this way, the rule 
functions only as a filter on surface forms, and there is no way for two separate rules 
to communicate via these rule-internal charts. 
This approach permits us to interpret any nondestructive autosegmental rule. 16 
Those rules involving deletion of autosegments or association lines must be approached 
in a completely different way. Rather than deleting an element in a particular context, 
we set up an alternation with zero, following Bloomfield (1926). See Bird and Klein 
(in press) and Bird (in press) for detailed examples of this approach to deletion. 
This concludes our discussion of the automaton-based semantics for autosegmen- 
tal representations and rules. In the next section we review some other attempts to 
treat autosegmental phonology using finite-state techniques. 
5. Other Finite-State Approaches to Nonlinear Phonology 
5.1 Kay (1987) 
The earliest treatment of autosegmental phonology in a finite-state setting is attributable 
to Kay (1987). It is tailored to the framework of nonconcatenative morphology de- 
veloped by McCarthy (1981) in which consonants and vowels are segregated onto 
different tiers, as shown for the Arabic verb stem kattab in (29). 
16 Applied to a destructive autosegmental rule, the interpretation determines the restriction imposed on 
surface forms by the rule, were it the last rule in a derivation. 
76 
Steven Bird and T. Mark Ellison One-Level Phonology 
29. a perfective active 
/\ 
C V C C V C causative 
I/ 
k t b write 
This verb stem contains three morphemes. Together, they mean 'caused to write.' The 
challenge posed by Arabic morphology is to come up with a simple account of the 
interleaving of the morphemes, relating the forms a, CVCCVC, and ktb to the stem 
kattab. 
Kay's solution, which we sketch here, involves the use of a kind of transducer 
that reads four-tuples (rather than pairs like a normal FST). This transducer scans 
four strings, one for each of the three tiers in (29) and one for the corresponding 
surface form. In this way, Kay has identified tiers with tapes. The association relation 
is encoded in the way the transducer scans these four tapes. 
Kay specifies a transducer by providing a set of FRAMES. In effect, each frame 
specifies an association between a CV-tier slot and a melodic unit, such as a k or an 
a. This melodic unit appears on the current surface tape cell. Each frame is a four- 
tuple whose components correspond to (i) the consonantal root, (ii) the CV-tier, (iii) 
the vocalic melody, and (iv) the surface tape. A simple frame is the following: 
k : C :c: k 
This frame specifies that when a k is being read from the consonantal tape and a C is 
being read from the CV tape, then there is an empty transition on the vocalism tape 
and the surface tape must have a k. There is a similar frame for each consonant. A 
frame for the vowel a is c:V:a:a. 
More work is necessary in order to capture the idea of autosegmental spreading. 
Kay modifies the transducer model so that tape symbols can be inspected without 
the reading head being advanced. Three notational devices manipulate the way this 
revised model behaves. Square brackets around one of the components of a frame 
causes the corresponding tape cell to be scanned without advancing the read head. 
Braces around a frame component behave in the same way, but only if they are scan- 
ning the final symbol on a tape. (This is required to prevent the spreading of i, which 
will not be discussed here.) Finally, the symbol G is used instead of C for geminate 
consonants. We shall see how these devices operate using a worked example, in which 
the surface form kattab is derived from the three lexical forms of diagram (29). 
Display (30) gives the initial configuration. The box shows the collection of symbols 
currently being scanned on the four tapes. To the far right is the appropriate frame. 
An empty pair of brackets is equivalent to ~. 
30. 
t b k 
V G C V C C 
\[\] 
k 
77 
Computational Linguistics Volume 20, Number 1 
The first CV tape symbol is a C and so the current consonant k is written onto the 
surface tape and the read heads are advanced. Notice that the read head for the vowel 
tape is left on the first cell. This is because the frame specified that there be no transition 
on this tape. 
31. 
k b II 
C Va~ a G C V C V 
{a} 
k a 
In (31), the CV tape symbol is a V, and so the a is copied to the surface tape. Since 
this a is given in braces in the frame, there is no movement of the read head. 
32. 
/"-"x 
k I t I b It\] 
C V a~ t C V C G \[\] 
k a t 
Having reached configuration (32), we read a G, which indicates a geminate. The t 
is written to the surface tape and the read head for the consonant tape stays where 
it is. The brackets around t in the frame mean that there is a nondeterministic choice 
between moving the read head and leaving it in the same position. However, the G 
symbol requires that the head stays put. Next we get a C on the CV tape. 
33. 
/'--'x 
k / t / b It\] 
C V G a~ t V C C \[\] 
k a t t 
After (33), the consonant tape read head advances 17 and the CV tape head moves on 
to deal with another vowel (34). 
17 Note that there is nondeterminism hidden here again. As it happens, if the consonant tape does not 
advance then the b will never be used (i.e., we will get *kattat) and the transducer will fail, because of the requirement that all read heads be at the end of their input for a successful completion. 
78 
Steven Bird and T. Mark Ellison One-Level Phonology 
34. 
k t \[J 
C V G C ~a a C V 
{a} 
k a t t a 
Finally, the consonant b is transferred to the surface tape. 
35. 
k t /b\] \[b\] 
C V G C V ~ C 
\[\] 
k a t t a b 
The result on the surface tape is kattab as required. Of course, the transducer also 
works for recognition. We could specify a surface tape and leave one or more of the 
lexical tapes unspecified. 
Kay's system is ingenious and we have not been able to demonstrate all of its 
capabilities in this small space. Nevertheless, we believe it suffers from a number of 
problems. The first concerns the form katab (form I) and the (corresponding) reflexive 
form ktatab (form VIII) with an infixed t. Indeed, more than half of the forms that 
Kay cites have infixes, although these are not explicitly analyzed in the model. Two 
conservative extensions to the model that encompass this infixation are (i) to insert 
infixes into the CV tape directly (so form VIII is CtVCVC) or (ii) to introduce another 
CV tape symbol A (for affix), which directs the transducer to read from a fifth tape. 
A second set of problems concerns the appropriateness of Kay's model for non- 
linear phonology more generally. First, notational devices like G move the frame 'lan- 
guage' away from what it is supposed to represent, namely autosegmental structures. 
No longer is gemination represented by association (perhaps derived by a spread- 
ing rule), but by a special marking on the skeleton. Second, the model builds in the 
assumption that each morpheme appears on a separate autosegmental tier. However, 
most applications of autosegmental phonology employ morphemes with phonological 
information arrayed on more than one tier (e.g., Clements and Ford 1979). Similarly, 
the modeling of subsegmental feature geometry of the kind advocated by Clements 
(1985) and others also involves a single morpheme having material on several tiers. 
Third, the model breaks down in the area of MORPHEMIC SEGREGATION. Since there 
is no principled upper bound on the number of morphemes that may be overlaid 
in the way McCarthy advocates for Arabic, there is similarly no principled upper 
bound on the number of tapes Kay's transducer would require. The assumption that 
each morpheme defines its own set of tiers, implicit in early work (McCarthy 1981) 
but explicit in more recent work (McCarthy 1989), is incompatible with a fixed upper 
bound on the number of tapes. Finally, using Kay's model for recognition would lead 
to much nondeterminism in positing G symbols, brackets, and braces. For example, 
in processing kattab the lexical tapes kttb, CVCCVC, and aa could be generated. The 
model generates all possible violations of the Obligatory Contour Principle. 
79 
Computational Linguistics Volume 20, Number 1 
5.2 Kornai (1991) 
Kornai (1991) has developed a linear encoding of autosegmental representations that 
allows the two-level transducer model to be applied to autosegmental phonology. 
We shall present Kornai's central innovation and describe a few of its strengths and 
weaknesses. 
As we saw in Section 4.5, Kornai presents four criteria under which an autoseg- 
mental encoding should be judged. An encoding should be easily computable; ideally 
by finite automata. It should also be invertible. An encoding should be iconic; minimally 
changing the input should minimally change the output. Finally, it should be compo- 
sitional, in the sense that the concatenation of the encodings of A1 and A2 ought to be 
the same as the encoding of the concatenation of A1 and A2. Kornai demonstrates that 
an optimal encoding under these criteria does not exist, and he sets about defining an 
encoding that is claimed to be near-optimal. 
We shall only cover one of the codes he considers, namely, the TRIPLE CODE. 
This code represents two tiers of autosegments and a chart between them as a linear 
description. There are four keywords in the code that are interpreted as instructions to 
a device that is scanning two tiers (left to right) and drawing association lines between 
certain pairs of segments. These keywords are as follows: 
0 : leave the current segments unassociated and advance the read head on each 
tier; 
1 : draw an association line between the current segments on each tier and ad- 
vance the read heads on each tier; 
t : override the advance instruction on the bottom tier, i.e., only advance the read 
head on the top tier; and 
b : only advance the read head on the bottom tier; retain the same segment on 
the top tier. 
Each 0 or 1 is flanked by a statement of the current segments on the two tiers. A 
number flanked by two segments forms the TRIPLE that gives the code its name. 
Where this code could give a number of different representations of the same 
autosegmental structure, Kornai (1991) restricts the encoding to operating in the same 
manner as the association convention of Goldsmith (1976). Association, or the lack 
of association, is marked left-to-right in a one-to-one fashion until one tier is devoid 
of new autosegments. From that point, only one tier advances until all remaining 
autosegments are represented in the linearization. 
As examples, let us encode two charts, the first completely devoid of associations. 
To show the pairs of current autosegments through the steps of the encoding, we link 
them with dotted lines indexed by the count of the step in the derivation. 
A 
I 
a 
B C 
2 3 
b c 
D 
i -.'-. 6 4i5 • 
d e 
The first four code steps give the following encoding (separating triples with full 
stops): 
AOa.BOb. COc.DOd 
80 
Steven Bird and T. Mark Ellison One-Level Phonology 
For the remaining two steps, we must spread (in virtual associations, not real ones) the 
final segment of the top tier, remembering to record the fact that only the read head 
on the bottom tier advances. The total encoding of this chart is thus the following: 
AOa.BOb.COc.DOd.b.DOe.b.DOf 
As a second example, let us fill the same chart with some associations. 
A B C D 
// J\ 
a b c d e f 
The encoding for this chart is the following: 
Ala.t.Bla.Clb.b.Clc.Dld.b.Dle.b.DOf 
This code can be measured against the four criteria given above. The code is definitely 
computable; we have just given an algorithm for constructing the code for an arbitrary 
chart. Likewise' the code is invertible: given any code, it is clearly possible to reproduce 
the original chart. Similarly, given the chart decoded from any encoding, it is likewise 
possible to reproduce the encoding. 
The triple code is, however, neither iconic nor compositional. Consider the two 
autosegmental representations given below. 
A B C A B C \ 
a b c a b c 
The triple code for the first representation is A0a. BOb. C0c. For the second representa- 
tion, the triple code is the sequence: 
AOa.b.AOb.b.A1c.t.BOc.t.COc 
As we can see, a minor change in one part of the autosegmental representation has 
resulted in major changes in much of the triple-coded representation. Thus the code 
is not iconic. 
Similarly, if we concatenate the two representations shown below 
A B C D 
we get 
a b c d 
A B C D 
a b c d 
81 
Computational Linguistics Volume 20, Number 1 
If we concatenate the two encodings, we get AOa.b.AOb.b.AOc.BOd.t.COd.t.DOd, which is 
not the same as the encoding of the concatenation: Ala.Blb.Clc.Dld. So the triple code 
satisfies only two of the desiderata Kornai gives for linearizations of autosegmental 
representations. As mentioned above, Wiebe (1992) has proven that linear encodings 
cannot get much better than this: no linear encoding can be both invertible and com- 
positional. 
Under the triple code, a lot of phonological processes are not finite-state. For 
example, consider the process of putting morphemic tiers together. In the case of the 
Arabic example we saw in (29), we need to combine the root morpheme (with empty 
template and vocalic tier) ktb and the template (with empty vocalic and consonantal 
tiers) CVCCVC together in such a way that we end up with the encoding in (38). 
In autosegmental representations, this is achieved by concatenation, and the "before" 
and "after" diagrams appear in (36) and (37). 
36. C V C C V C 
37. 
k t b 
C V C C V C 
k t b 
Concatenation of codes does not bring the same result. Display (38) shows the two 
concatenands, written with 0 as place holder for the empty tier. TM 
38. C00.V00.C00.C00.V00.C00 + 00k.00t.00b 
C0k.V0t.C0b.t.C0b.t.V0b.t.C0b 
To obtain the correct result, we need a more complex operation than concatenation-- 
one that cannot be performed by any finite-state device. 
If the ktb is read before the CVCCVC is read, it is necessary to have sufficient state 
information to be able to store k, t, and b, which in an alphabet of about 28 consonants 
involves at least log2(283) ~ 15 extra bits of information in each state, compared to 
an otherwise equivalent transducer. Suppose now that we wish to apply the well- 
formedness condition (5), the association convention (6), and some other rules to (38) 
to achieve the pattern of association between the two tiers given in (29), repeated 
below. We are not interested here in the detail of how a transducer might perform this 
operation, only whether a transducer can perform it. 
39. C V C C V C 
/ I 
k t b 
18 As far as we have been able to find, Kornai (1991) does not discuss how to code a representation that 
contains an empty tier. Here, we have filled empty tiers with a "place holder' segment, and then used 
the triple code. 
82 
Steven Bird and T. Mark Ellison One-Level Phonology 
The encoding of (39) is Clk.V0t.t.Clt.t.Clt.V0b.t.Clb. Observe that the t is met in the 
second triple of the right-hand side of (38) and it must be 'stored' until the fourth 
triple in the output. In a form like tada.hraj, two consonants must be stored in this 
way, leading again to an explosion of state information. 19 If the template consisted 
of n CV syllables, and the consonant tier consisted of n consonants, then the asso- 
ciation mechanism would need to store at least L(n - 1)/2J autosegments. This is, 
in principle, unbounded, and so the association mechanism for coded autosegmental 
representations is not finite state. 
In general there is no limit on the amount of information that needs to be pre- 
served as state information in the transducer. Consequently, using the triple code, 
certain phonological processes cannot be modeled using a finite state transducer. In- 
deed, this is true of any linear code (Wiebe 1992). It is true that adding each individual 
association--that is, finding the next position to associate, making the association and 
shuffling the lower tier along--can be performed by a finite-state transducer. But no 
principled bound on the length of the derivation can be made, and the quantity of 
memory required is an increasing function of the length of the derivation. Conse- 
quently, the step from unassociated to associated form cannot be made by a single 
FST. 
This type of problem with manipulating coded representations is not limited to 
Arabic morphology. In fact, the problem arises whenever there is an autosegment that 
has a restricted set of 'landing sites' (such as the consonants or the high vowels), a very 
common occurrence. Another example is that of accent systems: Kornai and K~lman 
(1988, p. 185), citing Goldsmith (1982), state the Basic Tone Melody Association rule 
for Ci-Ruri, a Bantu language of Tanzania, as follows: 
40. Basic Tone Melody Association Rule: 
Associate the accented element of the Basic Tone Melody to the accented 
element of a word. 
In general, the accented element of a word may be an unbounded distance from the 
start and end of a word. Goldsmith (1982, p. 53) gives the following example: 
41. na a kamir* e I milked 
L H* L 
As before, performing the association using the triple code requires a multiplication 
in state information, which in the general case cannot be bounded. True, this example 
only requires the automaton to store H; but Goldsmith (1982, p. 55) gives examples 
where there are two stresses in the word, and, therefore, the association automaton 
needs to store more tones (see Wiebe 1992, pp. 109-110 for the details). 
Next, consider the problem" of generalizing from one chart to an arbitrary number 
19 Wiebe (1992, p. 112) points out that the final consonant does not need to be stored, as it is already associated with the final C position. 
83 
Computational Linguistics Volume 20, Number 1 
of charts. Kornai (1991, p. 70) gives the following example: 
d e f 
I/ 
g h i 
J k 1 
Each chart is encoded as a string and then these two strings are linked by association 
lines. Each line indicates that the segments it connects are actually the same segment 
(in the original diagram). Here is the first step of the encoding of the above diagram: 2° 
dlgtelgfOhbfli 
\I I I\ 
g13 thlkiOkb± i1 
Next, this encoding can itself be encoded into a string 77 characters long. Kornai 
himself admits (p. 72) that his encoding is impractical for autosegmental structures 
having more than two charts, n Kornai discusses another encoding, which involves 
transducers that can read three (or more) tapes simultaneously. However, this comes in 
for the same criticism leveled at Kay's system above, namely that there is no principled 
upper bound on the number of tiers due to such considerations as complex feature 
geometry and morphemic segregation. 
5.3 Wiebe {1992) 
Wiebe (1992) has recently devised an encoding of autosegmental diagrams that over- 
comes many of the problems of Kornai's triple code. It is called the MULTI-LINEAR 
CODE. Consider again diagram (29), reproduced below. 
29. a perfective active 
/\ 
C V C C V C causative 
r~ 
k t b write 
This diagram contains two charts. The upper chart, which connects the vowel to two 
V slots will be referred to as chart 1, and the other chart as chart 2. Since there is a total 
ordering on the associations in a given chart (Goldsmith 1976, p. 29), it is sufficient to 
record how many associations each autosegment has on each chart, without specifying 
where these associations lead. The multi-linear code for (29) is displayed in (42). 
20 Here the dots separating the triples have been omitted to enable more convenient location of 
association lines. 21 The encoding procedure would not terminate for Pulleyblank's example (18). This is because the first 
application of the encoding would take the three-tiered structure and produce another (more 
complicated) three-tiered structure. In fact, this is just a special case of a more general problem, for the 
encoding fails for any tier structure containing a cycle, such as the one proposed by Clements (1991). 
84 
Steven Bird and T. Mark Ellison One-Level Phonology 
42. all 
C2VlC2C2VlC2 
k2t22b2 
Here, a numeral n following an autosegment A indicates that A has an association 
on chart n. These numerals can be stacked up; the first line specifies that a has two 
associations on chart 1. A given tier can participate in two charts; the second line of 
(42) has Cs associated on chart 2 and Vs associated on chart 1. Wiebe shows how 
the multi-linear code satisfies the criteria for computability, invertibility, iconicity, and 
compositionality. 
Wiebe's encoding has some similarities to our revised encoding presented in Sec- 
tion 4.5. Here is our encoding of (29): 
43. tier 1 a:2:0:0 
tier2 C:0:1:0 V:I:0:0 C:0:1:0 C:0:1:0 V:I:0:0 C:0:1:0 
tier3 k:0:l:0 t:0:2:0 b:0:l:0 
Suppose that a:p:q:r is an arbitrary 4-tuple of the kind in (43). We use position in the 
tuple to specify which chart an association line is in, and use numerals to specify the 
number of association lines. However, Wiebe uses numerals to specify the chart and 
repetitions to specify the number of lines. There is a mapping between terms like all 
in Wiebe's encoding and a:2:0:0 in our encoding: 
f(a:p:q:r)=alP2q3 r 
If we apply f to each term in (43), the result is as follows. 
all 
C2VIC2C2VIC2 
k2t22b2 
Now it should be clear that there is an isomorphism between the two encodings. 
Wiebe goes on to show how his encodings can be processed by new devices 
called MULTI-TAPE SFAs and SFTs, 22 where each tape corresponds to a row in the 
multi-linear code. The devices are used for checking well-formedness constraints and 
applying (possibly destructive) autosegmental rules. Wiebe also demonstrates that 
these devices are more powerful than FSTs without epsilon transitions, claiming that 
they can recognize some (strictly) context-sensitive languages. He argues that this extra 
computational power is crucially required for processing autosegmental analyses with 
feature- or structure-modifying rules. 
The read heads can scan n-tuples separated by arbitrary distances, 
and each head reads one co-ordinate of the n-tuple under it .... It 
is precisely this ability to scan different parts of an input word at 
the same time that is so important in modelling autosegmental rules. 
Association lines can associate segments in any part of one tier to 
segments in any part of the facing tier. In order for any computational 
22 Wiebe borrows the terms SFA and SFT from an earlier version of this paper (Bird and Ellison 1992). 
85 
Computational Linguistics Volume 20, Number 1 
device to efficiently process autosegmental representations, it must be 
able to scan two associated segments from widely separated parts of the 
representation at the same time (Wiebe 1992, pp. 95-96, emphasis added). 
While this is a reasonable statement regarding any model like Kornai's, it does 
not apply to our one-level model since the notion 'widely separated' is meaningless 
in this context. Two terms in a multi-linear code are widely separated if they differ 
significantly as to their distance from the left- or right-hand end of the encoding. Thus, 
the longer and more slanted the line that associates the autosegments, the greater their 
separation. Although this is true of the ink on a page, our semantics pays no attention 
to the angle and length of association lines. If a pair of autosegments is associated 
then they are ipso facto proximate. The temporal extent of autosegments expands or 
shrinks to accommodate these temporal constraints imposed by association lines. The 
intensional character of our approach gives rise to a flexibility of interpretation that 
obviates the need for more powerful devices of the kind advocated by Wiebe. 
Now that we have reviewed some details of other finite-state approaches to au- 
tosegmental phonology, we present a one-level analysis of an aspect of Arabic verb 
morphology. 
5.4 A One-Level Analysis 
As an alternative to either the elaboration of the transducer or the reduction of the 
nonlinear representation to a code, we present a partial analysis of the Arabic verb in 
terms of SEAs. 23 Rather than trying to generate one representation from another, we 
construct a description of the individual Arabic word by taking the intersection of the 
SFAs describing each of the morphemes. The forms we analyze here, like those given 
in McCarthy (1981), do not include inflections and also assume that the following 
suffix is consonant-initial} 4 We shall derive the stem kattab from three morphemes, 
specifying the form, the root, and the voice/mood. 
The form I SFA (44) generalizes over all verbs: it accepts/generates all correct 
verbs of this form, as well as a number of nonsense verbs, and factors out information 
specific to the individual words. 
44. C:1 V:0 C:1 V:0 C:1 
The indices mark associations that will link this tier to the root. 
Another tier is the consonantal tier. It contains heterogeneous autosegments cor- 
responding to consonants. These symbols are heterogeneous because they allow not 
only the corresponding consonant, but also any of the vowels. So the f autosegment on 
the consonant-tier denotes any sequence of the segments f, a, i, or u. The two uses of f 
are disambiguated by reference to the tiers on which they occur. Under this definition, 
autosegments on the consonant tier can spread over vowels to the next consonantal 
position. 
The root SFA generalizes over all stems constructed from the same root. The en- 
coding on the consonant tier describing the root fql, to do, appears in (45). 
23 The reader interested in other finite-state models of Arabic phonology is directed to the work of 
Narayanan and Hashem (1993), Beesley, Buckwater, and Newton (1989), and Beesley (1990). These have 
not been discussed at length here, because they do not seek to implement autosegmental phonology. 
24 Biliteral roots in forms I, III, IV, V, VI, VII, VIII, X, XI, XII, and XIV, triliteral roots in forms IX and XI, 
and quadriliteral roots in form QIV reorder the consonant-vowel sequence in the final syllable of the 
root if followed by a vowel-initial inflection. Where possible, the Arabic verb stem will metathesize or 
delete short vowels to create a geminate with root consonants (McCarthy 1981, pp. 197f). 
86 
Steven Bird and T. Mark Ellison One-Level Phonology 
45. .:0" f:l (.:0" ~:1) + 1:1 
The wildcards are necessary for affix consonants, such as the n- prefix of form VII or 
the -t- infix of form VIII. 
It might be argued that fixing the associations (by the indices) in the specification 
of the morpheme is redundant--that the associations should be supplied by rule. 
However, 3 of the 15 triliteral forms of the Arabic verb, namely forms II, V and XII, 
violate the association convention, multiply associating central autosegments rather 
than peripheral ones. For these forms there seems little choice but to lexically specify 
the associations or else introduce an ad hoc, morphologically triggered, rule (McCarthy 
1981, p. 392, rule 24). In our notation, we can specify these forms as: 
46. II C:1 V:0 C:1 C:1 V:0 C:1 
V t:l V:0 C:1 V:0 C:1 C:1 V:0 C:1 
XII C:1 C:1 V:0 w:0 C:1 V:0 C:1 
The intersections of the three forms in (46) with the root automaton (45) are: 
47. II f:l V:0 ~:1 ~:1 V:0 1:1 
V t:l V:0 f:l V:0 ~:1 ~:1 V:0 1:1 
XII f:l ~:1 V:0 w:0 ~:1 V:0 1:1 
The vocalism for the active perfect verb (except form I) is very simple to express: it 
is a sequence of as interleaved with consonants. On a vowel plane, where each au- 
tosegment is heterogeneous, accepting a vowel and any of the consonants, this aspect 
may be expressed by the single autosegment a, which corresponds to the automaton 
in (48). 
48. 
Forming the intersection of this SFA with the intersections of roots and forms 
shown in (47) gives the following forms. 
49. II f:l a:0 q:l ~:1 a:0 1:1 
V t:l a:0 f:l a:0 ~:1 ~:1 a:0 1:1 
XII f:l ~:1 a:0 w:0 q:l a:0 1:1 
This brief analysis shows the ease with which a computable analysis of nonconcate- 
native morphology can be constructed in the SFA formalism. While it may not cover 
all possible generalizations about Arabic verbal structure, the important point here is 
that it captures generalizations that are always surface-true of the phonology of the 
separate morphemes that come together to build up the verb. 
5.5 Automata versus Linearizations 
We have seen that Kornai (1991) finds it necessary to choose between the imposition 
of restrictions on autosegmental phonology and the loss of finite stateness in the trans- 
duction relationship. As it turns out, the one-level approach does not suffer from this 
problem. In this section we explain why. 
87 
Computational Linguistics Volume 20, Number 1 
Note that the natural processes by which finite-state automata are combined, and 
therefore by which regular languages are manipulated, are not themselves regular. To 
see why this is so, suppose we have two regular expressions describing the first form 
and the root of the Arabic verb to write: 
50. CVCVC 
k (e* t) + e* b 
The intersection is the following regular expression: 
51. kVtVb 
The associations fixing the incidence of k with the first consonant slot, t with the third, 
and b with the final, are made by the intersection operation. The question arises as to 
how we can construct the associations if the same operation for Kornai's system is not 
regular. The operation we have applied here--intersection---cannot be performed by a 
regular transducer. This does not invalidate our claim to regularity. What is regular in 
our theory is each individual description and generalization about phonological data. 
That is, the descriptions we use are all regular descriptions of phonological objects. 
What is not regular in one-level phonology is the relationship between different 
formats of the same description. There is no finite-state transducer that will form the 
product of two regular expressions. Multilevel analyses necessarily seek to capture 
relationships between different descriptions, and like the product operation, these 
relationships often cannot be captured by finite-state transducers. 
6. Conclusions 
The starting point of this paper was the distinction between descriptions and objects. 
Multidimensional phonological structures were taken to be descriptions of classes 
of phonetic objects, following Wheeler (1981), Bird and Klein (1990), Pierrehumbert 
(1990), Bird (1990), and Coleman (1992). Multiple tiers could be put together not by 
a clever encoding but by the simple operation of intersection, which corresponds to 
logical conjunction. Furthermore, this move of intensionalizing phonology enabled us 
to provide a straightforward formal basis for adding logical negation and disjunction 
to our representations. 
One important consequence of this work is that there are now good prospects for 
the incorporation of nonlinear phonology into constraint-based grammar formalisms 
such as HPSG (Pollard and Sag 1987). Such a move gives rise to a novel view of the 
relationship between phonology and the other modules of grammar, as some initial 
investigation has already demonstrated (Bird 1992; Bird and Klein in press). Mak- 
ing surface generalizations the only goal of analysis makes the machine learning of 
analyses simpler (Ellison forthcoming). The automaton semantics for autosegmental 
representations and rules gives us a mechanical way of comparing the empirical claims 
made by a range of autosegmental and segmental accounts of natural language phe- 
nomena. Finally, to the extent that phonologists are becoming increasingly committed 
to a declarative, constraint-based view of their domain, we believe that the model 
proposed here is well suited to their computational needs. 
Acknowledgments 
This research is funded by the U.K. Science 
and Engineering Research Council, under 
grant GR/G-22084. Computational Phonology: 
88 
Steven Bird and T. Mark Ellison One-Level Phonology 
A Constraint-Based Approach. We are grateful 
to John Coleman, Mark Johnson, Andr~is 
Kornai, Ewan Klein, Henry Thompson, 
Markus Walther, Bruce Wiebe, and two 
anonymous reviewers for comments on this 
work. We are also grateful to the students 
who attended our course at the Fifth 
European Summer School on Logic, Language 
and Information (Lisbon, August 1993), and 
gave valuable feedback on this work. The 
authors take equal responsibility for the 
material presented here. 
References 
Antworth, E. (1990). PC-KIMMO: A 
Two-Level Processor for Morphological 
Analysis. Summer Institute of Linguistics. 
Bear, J. (1986). "A morphological recognizer 
with syntactic and phonological rules." In 
Proceedings, the 11th International Conference 
on Computational Linguistics, 272-276. 
Beesley, K. (1990). "Finite-state descriptions 
of Arabic morphology." In Proceedings, 
Second Conference on Bilingual Computing in 
Arabic and English. 
Beesley, K.; Buckwater, T.; and Newton, S. 
(1989). "Two-level finite state analysis of 
Arabic." In Proceedings, First Conference on 
Bilingual Computing in Arabic and English. 
Bird, S. (1990). Constraint-based phonology. 
Doctoral dissertation, University of 
Edinburgh. 
Bird, S. (1992). "Finite-state phonology in 
HPSG." In Proceedings, Fifteenth 
International Conference on Computational 
Linguistics, 74-80. 
Bird, S. (in press). Computational Phonology: 
A Constraint-Based Approach. Studies in 
Natural Language Processing. Cambridge 
University Press. 
Bird, S., and Ellison, T. M. (1992). "One 
level phonology: autosegmental 
representations and rules as finite-state 
automata." RP 51, University of 
Edinburgh, Centre for Cognitive Science. 
Bird, S., and Klein, E. (1990). "Phonological 
events." Journal of Linguistics 26, 33-56. 
Bird, S., and Klein, E. (in press). 
'.'Phonological analysis in typed feature 
systems." Computational Linguistics, 20. 
Bird, S., and Ladd, D. R. (1991). "Presenting 
autosegmental phonology." Journal of 
Linguistics, 27, 193-210. 
Bloomfield, L. (1926). "A set of postulates 
for the science of language." Language, 2, 
153-164. Reprinted in Readings in 
Linguistics I: The Development of Descriptive 
Linguistics in America 1925-56, edited by 
Martin Joos, 26-31. 
Broe, M. (1993). Specification Theory: The 
treatment of redundancy in generative 
phonology. Doctoral dissertation, 
University of Edinburgh. 
Chomsky, N., and Halle, M. (1968). The 
Sound Pattern of English. Harper and Row. 
Clements, G. N. (1985). "The geometry of 
phonological features." In Phonology 
Yearbook 2, edited by C. Ewen and 
J. Anderson, 225-252. Cambridge 
University Press. 
Clements, G. N. (1991). "Place of articulation 
in consonants and vowels: a unified 
theory." In Working Papers of the Cornell 
Phonetics Laboratory, Number 5, edited by 
G. N. Clements and E. Hume, 77-123. 
Phonetics Laboratory, CorneU University. 
Clements, G. N., and Ford, K. C. (1979). 
"Kikuyu tone shift and its synchronic 
consequences." Linguistic Inquiry, 10, 
179-210. 
Coleman, J. S. (1991). Phonological 
representations--their names, forms and 
powers. Doctoral dissertation, University 
of York. 
Coleman, J. S. (1992). "The phonetic 
interpretation of headed phonological 
structures containing overlapping 
constituents." Phonology, 9, 1-44. 
Ellison, T. M. (1992). "Discovering vowel 
harmony." In Background and Experiments 
in Machine Learning of Natural Language, 
edited by W. Daelemans and D. Powers, 
131-136. ITK. 
Ellison, T. M. (1993). The machine learning of 
phonological structure. Doctoral 
dissertation, University of Western 
Australia. 
EUison, T. M. (forthcoming). "The iterative 
learning of phonological constraints." 
Computational Linguistics. 
Goldsmith, J. A. (1976). Autosegmental 
phonology. Doctoral dissertation, 
Massachusetts Institute of Technology. 
Goldsmith, J. (1982). "Accent systems." In 
The Structure of Phonological Representations, 
edited by H. van der Hulst and N. Smith, 
47--63. Foris. 
Goldsmith, J. (1991). "Phonology as an 
intelligent system." In Bridges Between 
Psychology and Linguistics, edited by D. J. 
Napoli and J. Kegl, 247-267. Lawrence 
Erlbaum. 
Hopcroft, J., and Ullman, J. D. (1979). 
Introduction to Automata Theory, Languages 
and Computation. Addison-Wesley. 
Johnson, C. D. (1972). Formal Aspects of 
Phonological Description. Mouton. 
Karttunen, L. (1983). "KIMMO: A general 
morphological processor." Texas Linguistic 
Forum, 22, 163-186. 
Kay, M. (1987). "Nonconcatenative 
89 
Computational Linguistics Volume 20, Number 1 
finite-state morphology." In Proceedings, 
Third Meeting of the European Chapter of the 
Association for Computational Linguistics, 
2-10. 
Kaye, J. (1989). Phonology: A Cognitive View. 
Erlbaum. 
Kornai, A. (1991). Formal Phonology. Doctoral 
dissertation, Stanford University. 
Kornai, A., and K~ilman, L. (1988). 
"Hungarian sentence intonation." In 
Autosegmental Studies on Pitch Accent, 
edited by H. van der Hulst and N. Smith, 
183-195. Foris. 
Koskenniemi, K. (1983). Two-level morphology: 
A general computational model for word-form 
recognition and production. Doctoral 
dissertation, University of Helsinki. 
Koskenniemi, K. (1985). "Compilation of 
automata from morphological two-level 
rules." In Papers from the Fifth Scandinavian 
Conference of Computational Linguistics, 
University of Helsinki, 143-149. 
Leben, W. R. (1973). Suprasegmental 
phonology. Doctoral dissertation, 
Massachusetts Institute of Technology. 
Leben, W. R. (1978). "The representation of 
tone." In Tone--A Linguistic Survey, edited 
by V. A. Fromkin, 177-219. Academic 
Press. 
Mastroianni, M. (1993). "Attribute logic 
phonology." Technical Report CMU-LCL 
93-4, Carnegie Mellon University. 
McCarthy, J. (1981). "A prosodic theory of 
non-concatenative morphology." 
Linguistic Inquiry, 12, 373-418. 
McCarthy, J. (1989). "Linear order in 
phonological representation." Linguistic 
Inquiry, 20(1), 71-99. 
Narayanan, A., and Hashem, L. (1993). "On 
abstract finite-state morphology." In 
Proceedings, Sixth Conference of the European 
Chapter of the Association for Computational 
Linguistics, 297-304. 
Partee, B.; ter Meulen, A.; and Wall, R. E. 
(1990). Mathematical Methods in Linguistics. 
Studies in Linguistics and Philosophy. 
Kluwer. 
Pierrehumbert, J. (1990). "Phonological and 
phonetic representation." Journal of 
Phonetics, 18, 375-394. 
Pollard, C., and Sag, I. (1987). 
Information-Based Syntax and Semantics, 
Volume 13 of CSLI Lecture Notes. Stanford: 
Center for the Study of Language and 
Information. 
Pulleyblank, D. (1986). Tone in Lexical 
Phonology. Studies in Natural Language 
and Linguistic Theory. Reidel. 
Pulman, S. G., and Hepple, M. R. (1993). "A 
feature-based formalism for two level 
phonology: a description and 
implementation." Computer Speech and 
Language, 7, 333-358. 
Ritchie, G. D. (1992). "Languages generated 
by two-level morphological rules." 
Computational Linguistics, 18(1), 41-59. 
Ritchie, G. D.; Russell, G. J.; Black, A. W.; 
and Pulman, S. G. (1992). Computational 
Morphology: Practical Mechanisms for the 
English Lexicon. MIT Press. 
Russell, K. (1993). A constraint-based approach 
to phonology. Doctoral dissertation, 
University of Southern California. 
Sagey, E. (1988). "On the ill-formedness of 
crossing association lines." Linguistic 
Inquiry, 19, 109-118. 
Schiller, A., and Steffens, P. (1991). 
"Morphological processing in the 
two-level paradigm." In Text 
Understanding in LILOG, edited by 
O. Herzog and C.-R. Rollinger, 112-126. 
Springer-Verlag. 
Scobbie, J. (1991). Attribute-value phonology. 
Doctoral dissertation, University of 
Edinburgh. 
Sproat, R. (1992). Morphology and 
Computation. Natural Language 
Processing. MIT Press. 
Touretzky, D. S., and Wheeler, D. W. (1990). 
"A computational basis for phonology." 
In Advances in Neural Information Processing 
Systems 2: The Collected Papers of the 1989 
IEEE Conference on Neural Information 
Processing Systems, edited by 
D. S. Touretzky. Morgan Kaufmann. 
Wheeler, D. (1981). Aspects of a categorial 
theory of phonology. Doctoral dissertation, 
Unversity of Massachusetts, Amherst, 
MA. 
Wiebe, B. (1992). Modelling autosegmental 
phonology with multi-tape finite state 
transducers. Master's dissertation, Simoh 
Fraser University. 
90 
