Automatic Acquisition of Feature-Based Phonotactic Resources
Julie Carson-Berndsen & Robert Kelly & Moritz Neugebauer
Department of Computer Science
University College Dublin
Dublin 4, Ireland
{julie.berndsen,robert.kelly,moritz.neugebauer}@ucd.ie
Abstract
Automata induction and typed feature theory are de-
scribed in a unified framework for the automatic
acquisition of feature-based phonotactic resources.
The viability of this data-driven procedure is il-
lustrated with examples taken from a corpus of
syllable-labelled data.
1 Introduction
This paper combines two hitherto distinct areas of
research, namely automata induction and typed fea-
ture theory, for the purposes of acquiring phonotac-
tic resources for use in speech technology. In order
to illustrate the methodology a small annotated data
set for Italian has been chosen1; however, given an-
notated data, the techniques can be applied to any
language thus supporting language documentation
at the phonotactic level and eventually building up
a catalogue of reusable multilingual phonotactic re-
sources.
There are numerous ways in which phonotactic
information has been represented for use in speech
technology applications ranging from phrase struc-
ture rules to n-grams. In this paper, the feature-
based phonotactic automaton of the Time Map
model (Carson-Berndsen, 1998) is used as the rep-
resentational device. A phonotactic automaton de-
scribes all permissible sound combinations of a lan-
guage within the domain of a syllable in terms of
a finite state automaton, describing not only ac-
tual lexicalised syllables but also idiosyncratic gaps
which would be considered well-formed by a na-
tive speaker of a language. The advantage of this
representation of phonotactic constraints in the con-
text of speech recognition is that it allows out-
of-vocabulary items (new words) to be classified
as well-formed if they adhere to the constraints.
Furthermore, since the phonotactic automaton con-
strains with respect to the syllable domain, it pro-
vides a more flexible and linguistically motivated
1We use phonemically annotated data from the EUROM1
Multilingual European Speech Database.
context than n-grams which restrict their context to
a domain of fixed length (the n-1 preceding units).
A phonotactic automaton describes language-
specific constraints. Therefore, in order to develop
multilingual phonotactic resources, phonotactic au-
tomata for different languages must be produced.
Phonotactic automata for German and English have
already been constructed for the Time Map model
using manual techniques (Carson-Berndsen, 1998;
Carson-Berndsen and Walsh, 2000). Since manual
construction of phonotactic automata is time con-
suming and laborious, more recently focus has been
placed on combining manual and automatic tech-
niques in order to reduce the level of required hu-
man linguistic expertise. This will become more im-
portant when lesser-studied languages are addressed
when an expert may not always be available. The
techniques presented here are regarded as support
tools for language documentation which allow in-
ferences to be made based on generalisations found
in an annotated data set. The linguist is free to ac-
cept or reject the suggestions made by the system.
In what follows, a technique is described in
which phonotactic automata are acquired automat-
ically given annotated data for a language. While
this technique describes all forms found in the data,
acquired automata cannot be considered complete
since the data is likely to be sparse (in this paper
we illustrate this using a small data sample). How-
ever, by combining phonotactic automata with a
typed feature classification of sounds encountered
in the data, it is possible to highlight not only dis-
tributional similarities, but also phonetic similarities
which can be used to predict gaps in the represen-
tation. These can be presented to a user (at least a
native speaker of a language) who can accept or re-
ject these. Accepted forms are then integrated into
the phonotactic automaton.
2 Automatic Acquisition of Phonotactic
Automata
The approach described in this section is one alter-
native to a fully manual construction of phonotac-
                                                                  Barcelona, July 2004
                                              Association for Computations Linguistics
                       ACL Special Interest Group on Computational Phonology (SIGPHON)
                                                    Proceedings of the Workshop of the
tic automata whereby they are rapidly acquired au-
tomatically and at a low cost. Given a corpus of
well-formed syllables for the language in question,
it is assumed here that the phonotactics for the lan-
guage is implicit in the syllables themselves and can
be automatically extracted by examining each syl-
lable structure in turn. An extracted phonotactics
is assumed to describe at least the syllables in the
corpus and is also assumed to be an approximation
of the complete phonotactics for the language from
which the data was drawn. Since the phonotactics in
question are finite-state structures and the data avail-
able for acquiring phonotactics is a corpus of posi-
tive examples of well-formed syllable structures, the
approach adopted here is to apply a regular gram-
matical inference procedure which can learn from
positive data alone. The field of grammatical infer-
ence has yielded many important learnability results
for different language classes. A full discussion of
these results is beyond the scope of this paper how-
ever see Belz (2000, Chapter 3) for a concise sum-
mary and discussion and Angluin and Smith (1983)
for a survey style introduction to the field. Suffice
to say that since the formal language of well-formed
syllables in a given natural language is finite, it is
possible to learn the structure of a regular grammar
i.e. the required phonotactic automaton from posi-
tive data alone i.e. the corpus of well-formed sylla-
bles.
The choice of regular inference algorithm is in
fact arbitrary. Many algorithms have been devel-
oped which can perform this learning task. For the
purposes of this paper however the ALERGIA (Car-
rasco and Oncina, 1999) regular inference algorithm
is used. This algorithm as applied to the problem of
inferring phonotactic automata is described in detail
elsewhere (Kelly, 2004b) . Here, the workings of
the algorithm are described by example. Note that
ALERGIA in fact treats any positive data sample
as having been generated by a stochastic process.
Thus, learned automata are in fact stochastic au-
tomata i.e. automata in which both states and tran-
sitions have associated probabilities, however tradi-
tional automata can be obtained by simply ignoring
these probabilities. Table 1 shows a small subset
of the Italian data set consisting of 14 well-formed
Italian syllables each consisting of 3 segments and
transcribed using the SAMPA phonetic alphabet 2.
The ALERGIA inference algorithm takes as in-
put a sample set of positive strings S (represent-
ing well-formed syllables in this case) together with
a confidence value α and outputs a determinis-
tic stochastic automaton A which is minimal for
2http://www.phon.ucl.ac.uk/home/sampa/
/v e n/ /r a n/ /b e n/ /m e n/ /t w a/
/k a n/ /n o n/ /t o n/ /f j o/ /r a n/
/d j o/ /s t o/ /s t e/ /t s a/ /p l o/
Table 1: Training set of Italian syllables.
the language it describes. ALERGIA proceeds by
building a Prefix Tree Automaton (PTA) from the
strings in S. The PTA is a deterministic automa-
ton with a single path of state-transitions from its
unique start state for each unique prefix which oc-
curs in S. Also, the PTA has a single acceptance
path, i.e. path of state-transitions from the start
state to some final state, for each unique string in
S where an initial subset of the transitions in ac-
ceptance paths for strings are shared if those strings
have a common prefix. Thus, common prefixes are
essentially merged giving the PTA its tree shape.
The PTA for S accepts exactly the strings in S and
each state of the PTA is associated with a unique
prefix of S. ALERGIA also assigns each transi-
tion in the PTA a frequency count dependent on the
number of prefixes in S which share that transition.
Similarly, each state has an assigned frequency de-
pendent on the number of strings from S which are
accepted (or equivalently, generated) at that state.
The PTA for the Italian syllables of table 1 is shown
in figure 1. Note that final states are denoted by
double circles and the single start state is state 0. In
figure 1 final state 26 has a frequency of 2 since the
two occurrences of the syllable /r a n/ terminate at
this state. All other final states have a frequency of
1 since exactly one syllable terminates at each final
state. All other states have a frequency of 0. Simi-
larly, the transition from state 0 to state 13 has a fre-
quency of 3 since three of the syllables in the train-
ing set begin with /t/and the transitions from state 0
to 17 and from state 17 to 18 have a frequency of
2 since two syllables begin with the segment com-
bination /s t/. The frquencies associated with the
states and transitions of the PTA can be used to as-
sociate a stochstic language with each state. The
set of acceptance paths from a given state determine
the set of strings in its associated language and the
probability for a given string in the language is eas-
ily derived from the frequencies of the states and
transitions on the acceptance path for that string.
ALERGIA uses the PTA as the starting point for
constructing a canonical, i.e. minimal deterministic,
automaton for S. The canonical automaton is iden-
tified by performing an ordered search of the au-
tomata derivable from the PTA by partitioning and
merging subsets of states of the PTA. Using the stan-
Figure 1: Prefix Tree Automaton for the syllables in
table 1.
dard order on the prefixes associated with the states
of the PTA, pairs of states are subsequently exam-
ined to determine if they generate a similar stochas-
tic language within a statistical significance bound
dependent on the supplied confidence value α. If
a pair of states are deemed to statistically generate
the same language then they are merged into a sin-
gle state and the state-transitions of the automaton
are altered to reflect this merge. The canonical au-
tomaton is identified when no more state merges are
possible. Figure 2 shows the canonical automaton
derived from the PTA in figure 1.
Since automata are derived from training sets of
syllables through the use of a language indepen-
dent regular inference algorithm, the procedure de-
scribed above is generic and language independent.
However, the procedure is of course dependent on
the existence of a corpus of training syllables for
the language in question and since it is entirely data
Figure 2: Canonical Automaton for the PTA in fig-
ure 1.
driven, the quality of the resulting phonotactics will
be dependent on the quality and completeness of the
syllable corpus. Thus, firstly the corpus must have
high quality annotations. Fortunately, the need for
high quality annotations in corpora is now recog-
nised and has become an essential part of speech
technology research and we assume here that high
quality annotations are available. Secondly, if valid
sound combinations are not detailed in the train-
ing corpus then they may never be represented in
the learned phonotactics. In order to be complete
the learned automaton must model all valid sound
combinations, however. In this case, generalisation
techniques must be applied in conjunction with the
inference algorithm in order to identify and rectify
gaps in the training corpus. This ensures that the
acquired phonotactics describes as close an approx-
imation as possible to the complete phonotactics for
the language. One such approach to generalisation
which operates independently of the chosen regular
inference algorithm is described in Kelly (2004a).
An alternative technique is discussed in section 3.
Finally, note that learned automata represent the
first stage in the development of multilingual phono-
logical resources called Multilingual Time Maps
(MTMs) (Carson-Berndsen, 2002). An MTM ex-
tends the single tape model of a phonotactic au-
tomaton to a multitape transducer whereby the dif-
ferent transition tapes detail linguistic information
of varying levels of granularity and related to the
original segment label. An MTM might have in-
dividual tapes detailing the segment, the phonolog-
ical features associated with that segment, the av-
erage duration of the segment in a particular syl-
labic position etc. In particular, the segment tape of
the learned phonotactic automata can be augmented
with additional tapes detailing feature type labels
associated with the segments. These additional type
label tapes are discussed in more detail in the fol-
lowing section.
3 Phonotactic Automata and Typed
Feature Structures
Lexical knowledge representation in computational
phonology has already made extensive use of in-
heritance hierarchies to model lexical generalisa-
tions ranging from higher level prosodic categories
to the phonological segment. In contrast to the ap-
proach presented in this section, the work described
in (Cahill et al., 2000) is set in an untyped fea-
ture system using DATR to define inheritance net-
works with path-value equations (Evans and Gaz-
dar, 1996). The merits of applying a type discipline
even to untyped feature structures is considered in
Wintner and Sarkar (2002) from a general perspec-
tive and in Neugebauer (2003b) with special refer-
ence to phonological lexica.
Previous proposals to cast phonological structure
in a typed feature system can be found in Bird
and Klein (1994) and Walther (1999). However,
there are two major differences with regard to our
work. First, while types may denote sets of seg-
ments, we go beyond the idea of sets as arc labels in
finite-state automata (Bird and Ellison, 1994; Eis-
ner, 1997; van Noord and Gerdemann, 2001) which
says that boolean combinations of finitely-valued
features can be stored as a set on just one arc, rather
than being multiplied out as a disjunctive collection
of arcs. This choice has no theoretical consequences
but is merely a convenience for grammar develop-
ment (Bird and Ellison, 1994). The difference in
our approach consists in the hierarchical ordering
of types (or sets) which relates each arc label to
any other type in a given phonological typed feature
system; such type-augmented automata have been
formally defined in Neugebauer (2003c). Second,
inheritance of type constraints is assumed to gov-
ern all subsegmental feature information (Neuge-
bauer, 2003b). Since here the crucial inheritance
relationships are induced automatically, we elabo-
rate on work by Walther (1999) where a complex
hand-crafted type hierarchy for internal segmental
structure is mentioned instead of simple appropri-
ateness declarations (Bird and Klein, 1994).
The interaction of finite-state automata and typed
feature structures is depicted in figure 3. Transitions
are exhaustively defined over a set of type labels
which are characterised by a unique position in the
underlying type hierarchy. This hierarchy is key to
the compilation of well-formed segment definitions
which are achieved by unification of partial feature
structures. In the simplest case, only atomic types
appear on the arcs which means that types corre-
spond to singleton sets. This can be achieved for
a phonemically annotated corpus (just like the cor-
pus in figure 1) by replacing all occurrences of a
phoneme with its appropriate atomic type label.
Figure 3: Example of the type-augmented automa-
ton for [traf].
The semantics of the type system assumed here
are extremely simple: the denotation of a parent
type in the directed acyclic graph that constitutes a
type hierarchy is defined as the union of the deno-
tation of its children, whereas a type node without
children denotes a unique singleton set (A¨ıt-Kaci et
al., 1989). Complex type formulae – as constructed
by logical AND – are implicitly supported for the
case of intersections since the greatest lower bound
condition (Carpenter, 1992) is assumed: its the for-
mal definition (also known as meet) states that in
a bounded complete partial order that is an inheri-
tance hierarchy, two types are either incompatible or
compatible. While in the first case, type constraints
are shared, in the latter case we require them to have
a unique highest common descendant. As suggested
in the LKB system (Copestake, 2002), these types
will in our approach be generated automatically if a
type hierarchy does not conform to this condition.
These types – such as glbtype2 in figure 3 – do not
have their own local constraint description and thus
do not rely on purely linguistic motivation.
A useful application of the greatest lower bound
condition seems to be the possibility that we can re-
fer to a set of compatible types simply by reference
to their common descendant. As indicated by the
hierarchical structure which is built over type09 in
figure 3, atomic types encode maximal information
whereas non-atomic types characteristically contain
only partial information. Thus, by defining transi-
tions over types such as type34 we might elegantly
capture phonotactic generalisations over a subset of
fricative sounds. This naturally raises the question
as to how the hierarchies are actually determined;
a suitable algorithm is described below, a detailed
specification is provided in Neugebauer (2003a).
Given a set of phonological feature bundles, an
inheritance hierarchy may be generated in the fol-
lowing way. For each feature which is defined for a
linguistic object (here: a phoneme) we compute the
corresponding extent or set description. The algo-
rithm then inserts these set descriptions into a lattice
and looks them up at the same time: it asks for the
smallest description that is greater than a singleton
set o with respect to the total order ≺ used inside
the tree. Every fully specified feature structure for
a given phoneme will deliver such a singleton set,
given that no two segments have been defined using
the identical feature structure.
The algorithm can be employed to recursively
compute all set descriptions of a feature system by
starting from the smallest set description of the lat-
tice. We need the lattice structure to encode the
inheritance relationships between sets; in the con-
text of lattice computation we will refer to these sets
in terms of nodes. Every set description o has two
lists associated with it: the list o∗ of its upper nodes
and the list o∗ of its lower nodes. One node may
be shared by two different set descriptions as their
upper node. While the algorithm processes each of
those two set descriptions, their shared upper node
must be detected in order to configure the relation-
ships correctly. To this end, all set descriptions are
stored in a search tree T. Every time the algo-
rithm finds a node it searches for it in the tree T to
find previously inserted instances of that set descrip-
tion. If the description is found, the existing lists
of nodes are updated; otherwise the previously un-
known set description is entered into the tree. Figure
4 demonstrates this procedure for the feature bundle
{fricative,labiodental,voiceless}. Once the smallest
set description [labiodental] is not able to include
one of segments which are successively added to it
a new upper node is created.
To make sure that all set descriptions that are in-
serted into the tree are also considered for their up-
Figure 4: Induction of subsumption hierarchies.
per nodes, the total tree order ≺ must relate to the
partial lattice order ≤ in the following way: o1 < o2
implies o1 ≺ o2. This is how recently inserted
nodes are greater than the actual set description with
respect to ≺ and will be considered later.3
Once we have computed all set descriptions, we
finally assign types and type constraints to all nodes
in the hierarchy. Therefore, the set of feature struc-
tures which constituted the starting point of our al-
gorithm has now been computed into a data struc-
ture which supersedes the previous level of infor-
mation in terms of a type inheritance network. The
last step consists of the insertion of greatest lower
bounds thus generating a well-formed lattice. Fig-
ure 5 visualises the final type inheritance hierarchy
for an Italian corpus containing 22 phonemes, each
corresponding to a unique atomic type (type01, . . . ,
type22). While the types numbered 23 to 39 are
generated by our set-theoretic algorithm, the great-
est lower bounds (the glb-types) are required by the
formal characteristics of our type system.
Any of the non-atomic types may be used to ex-
press generalisations over sets of phonological seg-
ments since each partial feature structure subsumes
all compatible fully specified segment entries. Ad-
ditionally, non-atomic nodes may be associated with
constraints which define appropriateness declara-
tions for linguistic signs of a particular type. For
example, all segments are at least characterised with
respect to four attributes (phonation, manner, place
and phonetic symbol). The next section sketches
an application of typed feature structures addressing
3Just computing the (ordered) set descriptions turns out to
be more effective since no hierarchy has to be computed. Com-
puting set descriptions as well as their hierarchical structure
takes twice as long for the same input when the algorithm is
used. This is also due to memory usage: while the computation
of all set descriptions is only based on a single predecessor, the
integration of a lattice algorithm stores all set descriptions in a
persistent search tree.
Figure 5: Complete generated type hierarchy.
data sparseness in automatically learned phonotac-
tic automata.
4 Examples
The integrated approach utilising both automata in-
duction and typed feature theory as presented in
the previous sections requires a phonemically an-
notated corpus. Each phoneme is then mapped to a
canonical feature bundle which is based on the pho-
netic characteristics specified in the International
Phonetic Alphabet (IPA); the features used in Fig-
ure 3 serve as an example. Our set-theoretic algo-
rithm operates on these feature structures thus de-
riving a type inheritance network for the corpus in
question. Note that phonemes and features are not
corpus-specific but rather a subset of a language-
independent set of linguistic descriptions that is the
IPA. As a result, we obtain a representation of our
annotation alphabet (phoneme and feature labels)
which exclusively refers to (sets of) linguistic ob-
jects via their corresponding types. This is exempli-
fied in figure 3 for an individual sound and in figure
5 for the full corpus.
Once the complete type hierarchy has been gen-
erated the inheritance relationships described can be
used to construct more compact finite-state struc-
tures than the automata learned over the original
data set. In addition, the linguistic generalisations
described by the hierarchy can be used to address
data sparseness in the training corpus. To illustrate
this, the automata are learned over type labels rather
than segments. Since all transitions of the learned
automata will now be labelled with types the infor-
mation in the feature hierachy can be used to ex-
press generalisations. To ensure that automata are
learned over types the segments in the training data
must be replaced with type labels that correspond to
singleton sets containing only the original segment.
For example, the syllable /r a n/ in table 1 would be
replaced by /type03 type01 type10/. Note that in this
case the transitions of the learned automaton will be
labelled with type labels rather than segments.
In the first case, the type hierarchy allows more
compact automata to be constructed by examining
the set of transitions emanating from each state of
the learned automaton. If each of the transitions em-
anating from a given state s1 have the same destina-
tion state s2 then the type labels on each transition
are examined to determine if they have a common
ancestor node in the hierarchy. If a common an-
cestor exists and if no other type label is the child
of that parent other than those appearing on the set
of transitions then they can be replaced by a sin-
gle transition from s1 to s2 labelled with the parent
type. The topmost diagram of figure 6 illustrates a
small section of the learned automaton for the full
Italian data set. In this case there are two transi-
tions from state 22 to state 35, one labelled with
type15 and the other labelled with type06. Refer-
ring to the hierarchy in figure 5, a common parent of
type15 and type06 is type30 and the only children of
type30 are type15 and type06. Therefore these two
transitions can be replaced by the single transition
labelled with type30.
Note that replacements of the kind described
above serve only to produce more compact au-
tomata and do not extend the coverage of the au-
tomaton. However, it is possible to use the type
hierarchy to achieve a more complete phonotactics.
The middle diagram of figure 6 shows another small
section of the learned automaton for the full Italian
data set. Referring again to the hierarchy in figure 5,
it can be seen that type29 is a parent of type14 which
labels the transition from state 0 to state 7 and is also
a parent of type18 which labels the transition from
state 0 to state 5. Similarly, type25 is a common par-
ent of type06 and type01 which label the transitions
from state 7 to state 17 and state 5 to state 25 respec-
tively. Finally, type35 is a common parent of type10
and type14 which label the transitions from state 17
to state 35 and state 25 to state 35. If each type label
is replaced by its common parent then both transi-
tions from state 0 are labelled with type29. Also,
the paths emanating from the destination states of
these transitions (states 5 and 7) are both labelled
with type25. In this case state 5 and state 7 can be
merged into a single state. A similar state merging
can be performed for states 17 and 35 resulting in a
new automaton as shown in figure 6. This process
yields a more general phonotactics since type29 ac-
tually denotes the segment set {p,m,b} and type25
denotes the set {a,i,e,E}. Thus, the segment p has
been effectively introduced as a new onset conso-
nant cluster that can precede any vowel in the set
denoted by type25. Also, as a result of introducing
type25 the additional vowels i and E have been in-
troduced as new vowel clusters. Note however that
type35 denotes exactly the set {m,n} and so no new
coda clusters are introduced.
Figure 6: Finite-state diagrams.
5 Conclusion
An important pre-requisite for the development of
robust multilingual speech technology applications
is the availability of language resources at vary-
ing levels of granularity. This paper has presented
generic techniques for acquisition of language-
specific phonotactic resources. The techniques were
exemplified using a small data set for Italian,4 but
scale to larger data sets and can be applied to any
language. Although the induction techniques as de-
scribed here assume that data is annotated at the syl-
lable level, only very few corpora are actually an-
notated at this level; a more usual annotation is at
the phonemic level. As a result, a cyclical learn-
ing procedure has been developed which learns as
syllable annotation is being performed and uses the
phonotactic automaton developed thus far to predict
syllable boundaries for annotation support (Kelly,
2004b). The work presented in this paper repre-
sents one specific step towards the provision of fine-
grained representations for speech recognition and
4Due to space constraints this paper only in-
cludes selected examples of the acquired resources.
Additional information is publicly available at
http://muster.ucd.ie/sigphon/. This includes
the complete annotation alphabet (phoneme and feature set),
the typed feature system and complete state diagrams for all
phonotactic automata.
synthesis based on a combination of data-driven and
user-driven techniques.
Acknowledgements
This material is based upon works supported by
the Science Foundation Ireland under Grant No.
02/IN1/ I100. The opinions, findings and conclu-
sions or recommendations expressed in this mate-
rial are those of the authors and do not necessarily
reflect the views of Science Foundation Ireland.

References
A¨ıt-Kaci, Hassan, R. Boyer, P. Lincoln, and R. Nasr.
1989. Efficient Implementation of Lattice Oper-
ations. ACM Transactions on Programming Lan-
guages and Systems, 11(1):115–146.
Dana Angluin and Carl H. Smith. 1983. Inductive
inference: Theory and methods. ACM Comput-
ing Surveys, 15(3):237–269.
Anja Belz. 2000. Computational Learning of
Finite-State Models for Natural Language Pro-
cessing. Ph.D. thesis, University of Sussex.
Steven Bird and T. Mark Ellison. 1994. One–
Level Phonology. Computational Linguistics,
20(1):55–90.
Steven Bird and Ewan Klein. 1994. Phonological
Analysis in Typed Feature Systems. Computa-
tional Linguistics, 20:455–491.
Lynne Cahill, Julie Carson-Berndsen, and Gerald
Gazdar. 2000. Phonology–based Lexical Knowl-
edge Representation. In Lexicon Development
for Speech and Language Processing, pages 77–
114. Kluwer Academic Publishers, Dordrecht.
Bob Carpenter. 1992. The Logic of Typed Feature
Structures, volume 32 of Cambridge Tracts in
Theoretical Computer Science. Cambridge Uni-
versity Press, Cambridge.
Rafael C. Carrasco and Jose Oncina. 1999. Learn-
ing deterministic regular grammars from stochas-
tic samples in polynomial time. ITA, 33(1):1–19.
Julie Carson-Berndsen and Michael Walsh. 2000.
Interpreting multilinear representations in
speech. In Proceedings of the 8th Australian
Conference on Speech Science and Technology,
pages 472–477, Canberra, December.
Julie Carson-Berndsen. 1998. Time Map Phonol-
ogy: Finite State Models and Event Logics in
Speech Recognition. Kluwer Academic Publish-
ers, Dordrecht, Holland.
Julie Carson-Berndsen. 2002. Multilingual time
maps: Portable phonotactic models for speech
technology applications. In Proceedings of the
LREC 2002 Workshop on Portability Issues in
Human Language Technology.
Ann Copestake. 2002. Implementing Typed Fea-
ture Structure Grammars, volume 110 of CSLI
Lecture Notes. CSLI Publications, Center for the
Study of Language and Information.
Jason Eisner. 1997. Efficient generation in primi-
tive optimality theory. In Proceedings of the 35th
Annual Meeting of the Association for Compu-
tational Linguistics and the 8th Conference of
the European Association for Computational Lin-
guistics, Madrid.
Roger Evans and Gerald Gazdar. 1996. DATR: A
language for Lexical Knowledge Representation.
Computational Linguistics, 22(2):176–216.
Robert Kelly. 2004a. Generalisation in the auto-
matic acquisition of phonotactic resources. To
Appear in Proceedings of The University of Cam-
bridge Second Postgraduate Conference in Lan-
guage Research.
Robert Kelly. 2004b. A language independent ap-
proach to acquiring phonotactic resources for
speech recognition. In Proceedings of the 7th
Annual Colloquium for the UK Special Interest
Group for Computational Linguistics, pages 126–
133. CLUK04.
Moritz Neugebauer. 2003a. Automatic Generation
of Constraint Hierarchies. Poster presented at the
14th Meeting of Computational Linguistics in the
Netherlands, University of Antwerp.
Moritz Neugebauer. 2003b. Computational
Phonology and Typed Feature Structures. In
Proceedings of the First CamLing Postgraduate
Conference on Language Research. Cambridge.
University of Cambridge.
Moritz Neugebauer. 2003c. Subsumption in
Speech Recognition and Feature Theory. In Pro-
ceedings of the Twenty-ninth Annual Meeting of
the Berkeley Linguistics Society, University of
California at Berkeley. Berkeley Linguistics So-
ciety.
Gertjan van Noord and Dale Gerdemann. 2001. Fi-
nite State Transducers with Predicates and Iden-
tities. Grammars, 4(3):263–286.
Markus Walther. 1999. One–Level Prosodic Mor-
phology. In Marburger Arbeiten zur Linguistik,
volume 1, Philipps–Universit¨at Marburg.
Shuly Wintner and Anoop Sarkar. 2002. A note
on typing feature structures. Computational Lin-
guistics, 28(3):389–397.
