A Multi-Path Architecture for Machine Translation of  
English Text into American Sign Language Animation 
 
 
Matt Huenerfauth 
Computer and Information Science 
University of Pennsylvania 
Philadelphia, PA 19104 
matthewh@seas.upenn.edu 
 
 
Abstract 
The translation of English text into American 
Sign Language (ASL) animation tests the 
limits of traditional MT architectural designs.  
A new semantic representation is proposed 
that uses virtual reality 3D scene modeling 
software to produce spatially complex ASL 
phenomena called “classifier predicates.”  The 
model acts as an interlingua within a new 
multi-pathway MT architecture design that 
also incorporates transfer and direct 
approaches into a single system. 
1 Introduction and Motivation 
American Sign Language (ASL) is a visual/spatial 
natural language used primarily by the half million Deaf 
individuals in the U.S. and Canada.  ASL has a distinct 
grammar, vocabulary, and structure from English, and 
its visual modality allows it to use linguistic phenomena 
not seen in spoken languages (Liddell, 2003; Neidle et 
al., 2000).  English-to-ASL translation is as complex as 
translation between pairs of written languages, and in 
fact, the difference in modality (from a written/spoken 
to a visual/spatial manually performed system) adds 
new complexities to the traditional MT problem. 
Building an English-to-ASL MT system is important 
because although Deaf students in the U.S. and Canada 
are taught written English, the difficulties in acquiring a 
spoken language for students with hearing impairments 
prevents most Deaf U.S. high school graduates from 
reading above a fourth-grade level (students age 18 and 
older reading text at a typical 10-year-old level) (Holt, 
1991).  Unfortunately, many Deaf accessibility aids 
(e.g. television closed captioning or teletype telephone 
services) assume that the viewer has strong English 
literacy skills.  Since many of these individuals are 
fluent in ASL despite their difficulty reading English, an 
ASL MT system could make more information and 
services accessible in situations where English 
captioning text is above the reading level of the viewer 
or a live English-to-ASL interpreter is unavailable. 
Researchers in graphics and human figure modeling 
have built animated models of the human body that are 
articulate enough to perform ASL that native signers 
can understand (Wideman and Sims 1998).  Most 
animation systems use a basic instruction set to control 
the character’s movements; so, an MT system would 
need to analyze an English text input and produce a 
“script” in this instruction set specifying how the 
character should perform the ASL translation output.  
The MT task is conceived of as translation from English 
text into this script because ASL has no written form. 
While linguists use various ASL glosses, all were 
designed to facilitate linguistic study, not to serve as a 
natural writing system, and so they omit certain details. 
Since there is no ASL orthography used by the Deaf 
community, there are no natural sources of ASL 
corpora.  To collect a corpus for statistical MT research, 
a movement annotation standard must be developed, 
ASL performances videotaped, and finally the videos 
manually transcribed – a slow and expensive process 
(Niedle, 2000).  Motion-capture glove technology may 
seem like a solution to this problem, but this type of 
data cannot easily be synthesized into novel and fluent 
ASL animations.  The difficulty in obtaining large 
corpora of ASL is why statistical approaches to the 
English-to-ASL MT problem are not currently practical. 
2 ASL Linguistic Issues 
As opposed to spoken/written languages, ASL relies on 
the multiple simultaneous channels of handshape, hand 
location, palm orientation, hand/arm movement, facial 
expressions, and other non-manual signals to convey 
meaning.  To express additional meaning, ASL may 
modify aspects of the manual performance of a sign 
(handshape, timing, motion path, repetition, etc.), 
perform an additional grammatical facial expression, or 
systematically use the areas of space around the signer. 
ASL signers use the space around them for several 
grammatical, discourse, and descriptive purposes.  
During a conversation, an entity under discussion 
(whether concrete or abstract) can be “positioned” at a 
point in the signing space.  Subsequent pronominal 
reference to this entity can be made by pointing to this 
location, and some verb signs will move toward or away 
from these points to indicate their arguments.  
Generally, the locations chosen for this pronominal use 
of the signing space are not topologically meaningful; 
that is, one imaginary entity being positioned to the left 
of another in the signing space doesn’t necessarily 
indicate the entity is left of the other in the real world. 
Other ASL expressions are more complex in their 
use of space and position invisible objects around the 
signer to topologically indicate the arrangement of 
entities in a 3D scene being discussed.  Special ASL 
constructions called “classifier predicates” allow signers 
to use their hands to represent an entity in the space in 
front of them and to position, move, trace, or re-orient 
this imaginary object in order to indicate the location, 
movement, shape, or other properties of some 
corresponding real world entity under discussion.  A 
classifier predicate generally consists of the hand in one 
of a closed set of semantically meaningful shapes as it 
moves in a 3D path through space in front of the signer. 
For example, the sentence “the car drove down the 
bumpy road past the cat” could be expressed in ASL 
using two classifier predicates.  First, a signer would 
move a hand in a “bent V” handshape (index and middle 
fingers extended and bent) forward and downward to a 
point in space in front of his or her torso where an 
imaginary miniature cat could be envisioned.  Next, a 
hand in a “3” handshape (thumb, index, middle fingers 
extended) could trace a path in space past the “cat” in an 
up-and-down fashion as if it were a car bouncing along 
a bumpy road.  Generally, “bent V” handshapes are 
used for animals, and “3” handshapes, for vehicles. 
The ability of classifier predicates to topologically 
represent a three-dimensional scene make them 
particularly difficult to generate using traditional 
computational linguistic methods and models.  To 
produce this pair of classifier predicates, there must be a 
spatial model of how the scene is arranged including the 
locations of the cat, the road, and the car.  A path for the 
car must be chosen with beginning/ending positions, 
and the hand must be articulated to indicate the contour 
of the path (e.g. bumpy, hilly, twisty).  The proximity of 
the road to the cat, the plane of the ground, and the 
curve of the road must be selected.  Other properties of 
the objects must be known: (1) cats generally sit on the 
ground and (2) cars usually travel along the ground on 
roads.  The successful translation of the English text 
into these classifier predicates used a great deal of 
semantic analysis, spatial knowledge, and reasoning. 
3 ASL MT Architectural Designs 
There is an architectural spectrum along which most 
MT systems can be classified; loosely they are grouped 
into three basic designs: direct, transfer, or interlingua 
(Dorr et al., 1998).  Direct systems process individual 
words of the source language text; translation is 
achieved without performing any syntactic analysis.  
Transfer systems do analyze the input text to some 
syntactic or semantic level, and then a set of “transfer” 
rules produce a corresponding syntactic or semantic 
structure in the target language. Finally, a generation 
component converts this structure into a target-language 
text.  Interlingual systems take this analysis of the input 
text one step further: the source is analyzed and 
semantically processed to produce a typically language-
independent semantic representation called an 
“interlingua,” and then a generation component 
produces the target-language surface form from there.  
These design choices are often pictured as a pyramid, as 
in Figure 1, adapted from a figure in (Dorr et al., 1998). 
Generally, in the absence of statistical or case-based 
information, the higher up the pyramid that the source 
text is analyzed, the more complex and subtle are the 
divergences the system can handle.  In particular, at the 
interlingual level, a knowledge base can supplement the 
linguistic information, producing translations that use 
world knowledge and that may convey more 
information than was present in the source text (devoid 
of context).  However, any of the approaches can 
produce a correct translation for certain inputs since not 
all sentences require such sophisticated analysis to be 
translated – some exhibit little translation divergence.  
Another trend as one goes up the MT pyramid is that the 
Figure 1: Pyramid of MT Architecture Designs. 
amount of domain specific development work that must 
be performed increases dramatically.  While direct 
systems may only require a bilingual lexicon, transfer 
systems also require analysis and transfer rules. 
Interlingual systems require interlingual representations 
and sometimes domain specific knowledge bases. 
Non-statistical direct approaches to English-to-ASL 
MT generally produce simple translations that are often 
little more than word-to-sign dictionary look-ups.  With 
the addition of some basic sentence reordering 
heuristics, such systems can occasionally produce 
acceptable output on simple English inputs or on those 
English-ASL sentence pairs that have similar word 
order.1  Since no syntactic analysis is performed, there is 
no chance that an input sentence will be outside the 
linguistic coverage of the system; so, the translation 
process will always produce some output.  Even if an 
English word is not in the translation lexicon, manual 
fingerspelling can be used to express the word. 
Transfer MT designs address most of the linguistic 
shortcomings of direct systems but do require additional 
linguistic resources to be developed.  There have been a 
few transfer-based English-to-ASL systems built 
(Huenerfauth, 2003), and several have had success in 
particular aspects of the MT task, like expressing 
adverbials (Zhao et al., 2000) or representing ASL 
phonological information (Speers, 2001; Sáfár and 
Marshall, 2001).  These systems show promise that a 
transfer approach could someday handle most ASL 
sentences that do not require complex or topological use 
of the signing space.  As the “bumpy road” example 
illustrates, generating classifier predicates would require 
more than a simple syntactic or semantic analysis – 
spatial analogy, scene visualization, and/or some degree 
of iconicity seem to be involved.2  
For this reason, ASL transfer systems merely omit 
classifier predicates from their coverage; however, 
many English concepts lack a fluent ASL translation 
without them.  Further, these predicates are common in 
ASL; signers generally produce a classifier predicate at 
least once per minute (once per 100 signs) (Morford and 
MacFarlane, 2003).  So, systems that cannot produce 
classifier predicates are not a viable long-term solution 
to the English-to-ASL MT problem.  To supply the 
semantic understanding, spatial reasoning, and world 
knowledge that classifier predicate generation demands, 
an interlingual approach (one with deeper semantic 
analysis and 3D spatial representations) is required. 
                                                           
1 Direct systems more readily convert English text into a 
signing system like Signed Exact English, a manually coded 
form of English, not a distinct natural language, like ASL. 
2  Linguists debate whether classifier predicates are 
paralinguistic iconic gestures, non-spatial polymorphemic 
constructions, or compositional yet spatially-aware 
expressions (Liddell, 2003), but transfer approaches to MT 
seem ill-suited to producing classifier predicates in any case. 
4 A Multi-Path MT Architecture 
While an interlingual approach to the classifier 
predicate translation task sounds useful, there is a 
problem. It’s hard to built a true interlingual system for 
anything but a carefully limited domain; building the 
linguistic and knowledge resources needed for 
interlingual translation on less restricted texts can entail 
too much overhead to be practical.  What is special 
about the MT problem for ASL – and the reason why 
interlingual translation may be possible – is that we can 
characterize and identify the “hard” input sentences, the 
ones that require classifier predicates for translation.  
These are spatially descriptive English input texts, those 
generally containing: spatial verbs describing locations, 
orientations, or movements; spatial prepositions or 
adverbials with concrete or animate entities; or lexical 
items related to other common topics or genres in which 
classifier predicates are typically used.  Such genres 
(e.g. vehicle motion or furniture arrangement in a room) 
could be detected using the features mentioned above. 
While an interlingual approach is needed to translate 
into classifier predicates, there are a vast number of 
English input sentences for which such deep analysis 
and reasoning would not be necessary.  As we've seen 
from the direct and transfer discussion above, these 
resource-lighter approaches can often produce a correct 
translation from lexical or syntactic information alone.   
This analysis suggests a new multi-path architecture 
for an MT system – one that includes a direct, a transfer, 
and an interlingual pathway.  English input sentences 
within the implemented interlingua’s limited domain 
could follow that processing pathway, those sentences 
outside of the interlingual domain but whose syntactic 
features fall within the linguistic coverage of the 
analysis and transfer rules could use the transfer 
pathway, and all other sentences could use the direct 
pathway with its bilingual dictionary look-up. 
Limiting the domain that the transfer and interlingua 
components must handle makes the development of 
these components more manageable.  The transfer 
pathway’s analysis grammar and transfer rules would 
not have to cover every possible English sentence that it 
encounters: some sentences would simply use the direct 
translation pathway.  Limiting domains has an even 
more dramatic benefit for the interlingual pathway. 
Instead of building interlingual analysis, representation, 
and generation resources for every possible domain, the 
interlingual development can focus on the specific 
domains in which classifier predicates are used: walking 
upright figures, moving vehicles, furniture or objects 
arranged in a room, giving directions, etc.  In this way, 
the “depth” of divergence-handling power of some 
translation approaches and the “breadth” of coverage of 
others can both be part of this multi-path architecture.   
This design does more than just restrict the domains 
for which the interlingua must be implemented; it also 
reduces the ontological complexity that the entire 
interlingua must support.  The domains listed above 
share a common feature: they all discuss the movement, 
location, orientation, and physical description of entities 
in three-dimensional scenes.  Some complex 
phenomena whose handling often makes designing an 
interlingual representation quite difficult – abstract 
concepts, beliefs, intentions, quantification, etc. – do not 
need to be represented.  In a sense, this multi-path 
architecture doesn’t just limit the things that must be 
represented, but the “type” of these things as well. 
Having multiple processing pathways does not mean 
that there is necessarily a new problem of choosing 
which to use.  The system could be implemented as a 
‘fall back’ architecture in which the system could 
attempt the most complex approach (interlingual) and 
drop back to each of the simpler approaches whenever it 
lacks the proper lexical, syntactic, semantic, or 
knowledge resources to succeed for the current 
pathway.  In this way, the linguistic coverage of each of 
the levels of representation would define exactly how 
input sentences would be routed through the system.   
If the system were to use a more complex pathway 
than was necessary during translation, then, if properly 
implemented, output would be produced that could have 
been created using a simpler pathway. This is an 
acceptable, if less efficient, result.  If the system lacked 
the linguistic resources to translate a sentence using the 
sophisticated level of processing it required, then the 
output would be more English-like in structure than it 
should.  Because most Deaf users of the system would 
have had experience interacting with hearing people 
who used non-fluent English-like signing or manually 
signed forms of English, like Signed Exact English or 
Sign Supported English, then they may still find this 
overly English-like translation useful. 
5 A Spatial Interlingua for ASL MT 
When ASL signers describe a spatially complex 3D 
scene using classifier predicates, they visualize the 
elements of the scene as occupying an area of space that 
is generally within arm’s reach in front of their torso.  
So, signers have a spatial model of the scene under 
discussion that they can consider when selecting and 
generating classifier predicates to convey information.  
An automated system for creating classifier predicates 
may be able to use an analogous representation.   
One way to produce this model is to incorporate 
virtual reality 3D scene representation software into the 
MT system’s interlingual pathway.  After analyzing the 
English text, the movements of entities under discussion 
could be identified, and a 3D virtual reality model of the 
scene could be constructed and/or modified to reflect 
the information in the English text.  This spatial model 
could serve as the basis for generating the 3D and 
spatially analogous (topological) motions of the signing 
character’s hands while performing classifier predicates.   
Fortunately, a system for producing a changing 3D 
model of a scene from an English text has been built: 
the Natural Language Instructions for Dynamically 
Altering Agent Behaviors system (Bindiganavale et al., 
2000; Badler et al., 2000) (herein, “NLI”).  The system 
displays a 3D virtual reality scene and accepts English 
input text containing instructions for the characters and 
objects in the scene to follow. It updates the animation 
so that objects obey the English commands.  NLI has 
been used in military training and equipment repair 
domains and can be extended by augmenting its library 
of Parameterized Action Representations (PARs), to 
cover additional domains of English input texts. 
PARs are feature/value structures stored as a library 
of templates with slots specifying: the agent moving, the 
path/manner or translational/rotational nature of the 
motion, terminating conditions, speed/timing, and other 
motion information.  English lexicalized syntactic 
structures are associated with PARs so that the analysis 
of a text can be used to select a PAR template and fill its 
slots.  PARs serve as 3D motion primitives and are used 
as hierarchical planning operators to produce a detailed 
animation specification; so, they contain fields like 
preconditions and sub-actions used in NLI’s animation 
planning process (Badler et al., 2000).  A PAR generally 
corresponds to an English motion verb (or a set of 
related verbs); so, to extend NLI for use in an ASL 
context, additional PARs will be developed for English 
motion verbs that often produce classifier predicates. 
The MT system’s interlingual pathway will use the 
NLI software to analyze the English source text as if it 
were commands for the entities mentioned in the text.  
The NLI can create and maintain a 3D model of the 
location and motion of these entities.  The MT system, 
unlike other applications of the NLI software, does not 
care about the exact shape or appearance of the objects 
being modeled (generic box-like shapes could be used 
for each).  Instead, the location and motion paths of 
these objects in a generic 3D space are important, since 
these are used to build classifier predicates.    
The MT system would use the spatial model to 
instantiate a transparent miniature animation of these 
objects; this animation would be overlaid on an area of 
the virtual reality space in front of the torso of the 
character performing the ASL animation output.  In the 
“bumpy road” example, a small invisible object would 
be positioned in space in front of the chest of the 
signing character to represent the cat.  Next, a 3D 
animation path and location for the car (relative to the 
cat) would be chosen in front of the character’s chest. 
When objects in this “invisible world” are moved or 
reoriented to reflect information in the English text, the 
animated ASL-signing character can position its hand 
inside of the transparent (possibly moving) object to 
indicate its new location, orientation, or movement path. 
By choosing an appropriate handshape for the character, 
a classifier predicate is thus produced that conveys the 
spatial information from the English text.  Extensions of 
this design for more complex classifier predicate 
constructions are discussed in (Huenerfauth, 2004). 
This interlingual pathway design would pass along 
most of the spatial modeling and reasoning burdens to 
the NLI software, which was designed for this task.  It 
can select relative locations and motion paths for objects 
in the 3D scene based on prepositions and adverbials in 
the English input text.  It uses collision avoidance, 
physical constraints, generic and specialized motion 
primitives, and hierarchical motion planning operators 
to produce the necessary detail for a 3D animation from 
the limited information in a corresponding English text. 
The full architectural diagram is shown in Figure 2.  
This design visually resembles the pyramid in Figure 1: 
direct pathway at the bottom, transfer across the middle, 
and interlingual pathway over the top of the pyramid.  
The three paths no longer represent the design choices 
possible for different systems; they are now processing 
pathways within a single “pyramidal” architecture. 
6 Virtual Reality as Interlingua 
The 3D model produced by the NLI software serves as 
an intermediary between the English text analysis and 
the classifier predicate generation in this architecture, 
but that does not necessarily make it an interlingua.  In 
fact, the design differs from interlingual representations 
elsewhere in the MT literature significantly.  To explore 
this issue, consider a general definition of an interlingua 
as: a typically language-neutral semantic representation 
useful for MT that may incorporate knowledge sources 
beyond the basic semantics of the input text. 
First, the model represents those aspects of the input 
text’s meaning significant for translation to classifier 
predicates; thus it serves as a semantic representation 
within the 3D motion domain – albeit a non-traditional 
one due to the ontological simplicity of this domain.  
Second, this proposed architectural design has 
illustrated how this 3D scene representation is useful for 
MT.  Third, the NLI software’s ability to incorporate 
physical constraints, collision detection, and spatial 
reasoning shows how the 3D model can use knowledge 
sources beyond the original text during translation. 
So, the final determinant of this model’s interlingual 
status is its language-neutrality.  The 3D coordinates of 
objects in a virtual reality model are certainly language-
neutral. However, ASL linguists have identified 
discourse and other factors beyond the 3D scene model 
that can affect how classifier predicates are generated 
(Liddell, 2003).  If the classifier predicate generator 
needs these features, then the degree to which they are 
modeled in a language-neutral manner will affect 
whether the pathway is truly interlingual.  Until the final 
implementation of the generator is decided, it is an open 
issue as to whether this pathway is an interlingua or 
simply a spatially rich semantic transfer design. 3 
7 Discussion and Future Work 
While English-to-ASL MT motivated the multi-path 
pyramidal architecture, the design is also useful for 
other language pairs.  Merging multiple MT approaches 
in one system alleviates the traditional trade-off 
between divergence-handling power and domain 
specificity, thus making resource-intensive approaches 
(e.g. interlingual) practical for applications that require 
broad linguistic coverage.  This architecture is useful 
when a system must translate a variety of texts but 
perform deeper processing on texts within particular 
important or complex domains.  It is also useful when 
the input is usually (but not always) inside a particular 
sublanguage.  Transfer or interlingual resources can be 
developed for the domains of interest, and resource-
lighter (broader coverage) pathways can handle the rest. 
While the English-to-ASL system had no statistical 
pathways, nothing prevents their use in a multi-path 
pyramidal architecture.  Statistical approaches could be 
used to develop a direct pathway, and hand-built 
analysis and transfer rules for a subset of the source 
language could create a transfer pathway.  A developer 
could thus use a stochastic approach for most inputs but 
manually override the MT process for certain texts (that 
                                                           
3  Kipper and Palmer (2000) examined PARs as an 
interlingua for translation of motion verbs between verb-
frame and satellite-frame languages.  Unlike this system, 
they did not use PARs within a 3D scene animation; the 
PAR itself was their interlingua, not the 3D scene. 
Figure 2: Multi-Path “Pyramidal” MT Architecture. 
are important or whose translation is well understood).  
Likewise, a transfer pathway may use statistically 
induced transfer rules and parsers, and an interlingual 
pathway may be manually built for specific domains. 
While the pyramidal architecture has applications 
across many languages, the 3D scene modeling software 
has benefits specific to ASL processing.  Beyond its use 
in classifier predicate generation, the 3D model allows 
this system to address ASL phenomena that most MT 
architectures cannot.  The non-topological use of the 
signing space to store positioned objects or “tokens” 
(Liddell, 2003) for pronominal reference to entities in 
the discourse can easily be implemented in this system 
by taking advantage of the invisible overlaid 3D scene.  
The layout, management, and manipulation of these 
“tokens” is a non-trivial problem, and the richness of the 
virtual reality spatial model can facilitate their handling. 
The NLI software makes use of sophisticated human 
characters that can be part of the scenes being controlled 
by the English text.  These virtual humans possess skills 
that would make them excellent ASL signers for this 
project: they can gaze in specific directions, make facial 
expressions useful for ASL output, and point at objects 
or move their hand to locations in 3D space in a fluid 
and anatomically natural manner (Badler et al., 2000).  
If one of these virtual humans served as the signing 
character, as one did for (Zhao et al., 2000), then the 
same graphics software would control both the invisible 
world model and the ASL-signing character, thus 
simplifying the implementation of the MT system. 
Currently, this project is finishing the specification 
of the multi-path design and investigating the following 
issues: deep generation techniques for creating multiple 
interrelated classifier predicates, surface generation of 
individual classifier predicates from compositional rules 
or parameterized templates, and ASL morphological 
and syntactic representations for the transfer pathway.  
Another important issue being examined is how to 
evaluate the ASL animation output of an MT system – 
in particular one that produces classifier predicates. 
Acknowledgements 
I would like to thank my advisors Mitch Marcus and 
Martha Palmer for their guidance, discussion, and 
revisions during the preparation of this work. 
References 
R. Bindiganavale, W. Schuler, J. Allbeck, N. Badler, A. 
Joshi, and M. Palmer. 2000. “Dynamically Altering 
Agent Behaviors Using Natural Language 
Instructions.” 4th International Conference on 
Autonomous Agents.  
N. Badler, R. Bindiganavale, J. Allbeck, W. Schuler, L. 
Zhao, S. Lee, H. Shin, and M. Palmer.  2000.  
“Parameterized Action Representation and Natural 
Language Instructions for Dynamic Behavior 
Modification of Embodied Agents.” AAAI Spring 
Symposium. 
B. Dorr, P. Jordan, and J. Benoit.  1998.  “A Survey of 
Current Paradigms in Machine Translation.” 
Technical Report LAMP-TR-027, Language and 
Media Processing Lab, University of Maryland. 
J. Holt. 1991. Demographic, Stanford Achievement Test 
- 8th Edition for Deaf and Hard of Hearing Students: 
Reading Comprehension Subgroup Results. 
M. Huenerfauth. 2003. “Survey and Critique of 
American Sign Language Natural Language 
Generation and Machine Translation Systems.”  
Technical Report MS-CIS-03-32, Computer and 
Information Science, University of Pennsylvania 
M. Huenerfauth. 2004. “Spatial Representation of 
Classifier Predicates for Machine Translation into 
American Sign Language.” In Proceedings of the 
Workshop on the Representation and Processing of 
Signed Languages, 4th International Conference on 
Language Resources and Evaluation (LREC 2004). 
K. Kipper and M. Palmer. 2000. “Representation of 
Actions as an Interlingua.” In Proceedings of the 3rd 
Workshop on Applied Interlinguas, ANLP-NAACL.  
S. Liddell. 2003. Grammar, Gesture, and Meaning in 
American Sign Language.  UK: Cambridge U. Press.   
J. Morford and J. MacFarlane. “Frequency 
Characteristics of ASL.” Sign Language Studies, 3:2. 
C. Neidle. 2000. “SignStream™: A Database Tool for 
Research on Visual-Gestural Language.” American 
Sign Language Linguistic Research Project, Report 
Number 10, Boston University, Boston, MA, 2000. 
C. Neidle, J. Kegl, D. MacLaughlin, B. Bahan, and R. 
G. Lee.  2000.  The Syntax of American Sign 
Language: Functional Categories and Hierarchical 
Structure. Cambridge, MA: The MIT Press. 
É. Sáfár and I. Marshall.  2001.  “The architecture of an 
English-text-to-Sign-Languages Translation System.”  
In G. Angelova, ed., Recent Advances in Natural 
Language Processing. Tzigov Chark, Bulgaria. 
d'A. Speers.  2001.  Representation of ASL for Machine 
Translation. Ph.D. Diss., Linguistics, Georgetown U.  
C. Wideman & M. Sims. 1998. “Signing Avatars.” 
Technology & Persons with Disabilities Conference. 
L. Zhao, K. Kipper, W. Schuler, C. Vogler, N. Badler, 
and M. Palmer.  2000.  “A Machine Translation 
System from English to American Sign Language.” 
Association for Machine Translation in the Americas. 
