Design of a Hybrid Deterministic Parser 
Kanaan A. Faisal 
Information ,and Computer Science Department 
King Fahd University of Petroleum and Minerals T 
Dhahran 31261 
Kingdom of Saudi Aa'abia 
Start C. Kwasny 
Center for Intelligent Computer Systems :I 
Department of Computer Science 
Washington University 
St. Louis, MO 63130-4899 
U.S.A. 
1. Introduction 
A deterministic parser is under development which 
represents a departure from traditional deterministic 
parsers in that it combines both symbolic and connec- 
tionist components. The connectionist component is 
trained either from patterns derived from the rules of a 
deterministic grammar. ~The development and evolu- 
tion of such a hybrid architecture has lead to a parser 
which is superior to any known deterministic parser. 
Experiments are described and powerful training tech- 
niques are demonstrated that permit decision-making 
by the connectionist component in the parsing process. 
This approach has permitted some simplifications to the 
rules of other deterministic parsers, including the elimi- 
nation of rule packets and priorities. Furthermore, 
parsing is performed more robustly and with more 
tolerance for error. Data are presented which show 
how a connectionist (neural) network trained with 
linguistic rules can parse both expected (grammatical) 
sentences as well as some novel (ungrammatical or lex- 
ically ambiguous) sentences. 
2. Determinism and Parsing 
The determinism hypothesis which forms the basis for 
PARSIFAL (Marcus, 1980) imposes important restric- 
tions on NatLu'al Language Processing. It states (p.ll) 
that 
"Natural Language can be parsed by a 
mechanism that operates 'strictly 
deterministically' in that it does not simulate a 
nondeterministic machine..." 
If we accept this hypothesis, it must follow that pro- 
cessing need not depend in any fundamental way on 
backtracking. As a further consequence, no partial 
structures are produced during parsing which fail to 
become part of the final structure. PARSIFAL was the 
first of a number of systems to demonstrate how deter- 
ministic parsing of Natural Language can be performed 
using a rule-based grammar. Extensions to PARSIFAL 
have been researched independently including the pars- 
ing of ungrammatical sentences in PARAGRAM 
(Charniak, 1983), the resolution of lexical ambiguities 
in ROBIE (Milne, 1986), and the acquiring of syntactic 
rules from examples in LPARSIFAL (Berwick, 1985). 
t The first author gratefldly ackmowledge the support of Khlg 
Fahd University of Petroleum and Minerals. 
Traditional deterministic parsers process input sen- 
tences primarily left-to-right. Determinism is accom- 
plished by permitting a lookahead of up to three consti- 
tuents with a constituent buffer designated for that pur- 
pose. To permit embedded structures, a stack is also 
part of the architecture. Rules are partitioned into rule 
packets which dynamically become active or inactive 
during parsing, but are usually associated with the 
current (top-level) node of the structure being built. A 
single processing step consists of selecting a rule that 
can fire from an active rule packet, firing the rule, and 
performing its action. Conflicts are resolved within 
packets from the static ordering (priority) of rules. The 
action effects changes to the stack and buffer. After a 
series of processing steps, a termination rule fires and 
processing ends. The final structure is left on top of the 
stack. 
3. Hybrid Deterministic Parsing 
Our parser takes the approach of deterministic parsing 
and combines it with connectionism. McClelland and 
Kawamoto (1986, p.317) first suggested the combina- 
tion of these ideas. Deterministic parsing provides a 
setting in which no backtracking occurs while connec- 
tionism provides generalization and robustness. Our 
goal is to combine the two in a way that enhances their 
advantages and minimizes their faults. In simple terms, 
the rules of the deterministic parser are replaced by a 
network which is trained from training sequences 
derived from the grammar rules. The network embo- 
dies the decision-making component of the parser and 
maps a state of the parser to an action. Actions are per- 
formed in the traditional way by symbolically manipu- 
lating the stack and buffer contents. 
Parsing experiments are conducted to determine the 
effectiveness of training by attempting to process 
ungrammatical and lexically ambiguous sentence 
forms. The performance of our hybrid parser depends 
on the extent and nature of the training. Once trained, 
the network is efficient, both in terms of representation 
and execution. 
3.1. Design 
Some sm~l modifications to deterministic grammar 
rules arc necessary to insure the suitability of each rule 
for use with our "winner-take-all" network. Many of 
these changes are simplifications that have been 
~: The sponsors of the Center are McDonnell Douglas 
Coq~oration and Southwestern Bell Telephone Company. 
1 11 
Rule Main-verb 
IF: 
THEN: 
in packet parse-vp 
priority: 10 
The first element in buffer is a verb 
Rule Create._VP 
IF: current node is S node 
Attached is AUX node 
first is a verb 
DEACTIVATE packet parse-vp THEN: 
if the active node is a major sentence 
then ACTIVATE packet ss-final 
else if the active node is a secondary sentence 
then ACTIVATE emb-s-final. 
CREATE a VP node. 
ATTACH a VP node to the S. 
ATTACH the first element in the buffer to 
the active node as verb. 
ACTIVATE the clause level packet cpool THEN~ 
if verb is labeled passive 
then ACTIVATE the packet passive 
and RUN the grammar rule passive next. 
CREATE VP node 
Rule Main_verb 
IF: current node is VP node 
Attached is AUX node 
first is a verb 
ATTACH as MVB 
Figure I: PARSIFAL and Hybrid Parser Rules Compared 
proposed by others and are not essential to the success 
of our approach. All of these changes are made 
without substantially altering the capabilities 
represented in the original grammar rules. Changes 
include: elimination of the packet system; removal of 
attention-shifting rules; removal of rule priorities; 
reduction of lookahead to two positions instead of 
three; and revision of the rules so that a single action is 
performed by each. 
As an example, consider part of one sample grammar 
rule from PARSIFAL and its reformulation in the 
hybrid parser. Figure 1 shows the two styles side-by- 
side. Rule actions are in capital letters; rule names are 
in bold. In the PARSIFAL rule, a priority number is 
given explicitly and the rule contains multiple actions 
and conditionals similar to a programming language. It 
explicitly activates and deactivates rule packets, exe- 
cutes rules, creates new phrase structure nodes, and 
tests for complex properties of the elements in the 
buffer. 
Rules in the hybrid parser eliminate many of these 
details without substantially changing the capabilities 
of the grammar. In the figure, two of several rules 
derived from the Main-verb rule are shown. In the 
first rule, a new VP active node is created on the stack 
and in the second rule the verb is attached as a main 
verb to the active node (VP) on top of the stack. With 
the elimination of rule packeting, no priorities nor 
explicit packet activations/deactivations are required. 
While this mechanism is precisely what is required for 
efficient design of a symbolic parser, priorities are at 
the essence of what is learned when training the con- 
nectionist component of the hybrid. Actions such as 
creating and attaching or selecting the argument struc- 
ture of the verb are carried out symbolically in the 
hybrid parser. Also, a symbolic lexicon is consulted to 
determine the properties of words. When a predicate 
such as a verb is encountered, the requirements or 
expectations for its arguments are made part of the 
features of the active VP node, thus affecting which 
actions will be executed later on. 
3.2. Evolutionary Steps from PARSIFAL 
Elimination of the packet system. In PARSIFAL, rules 
are organized into packets. Only those rules in an 
active packet are considered while processing. Often, 
more than one packet is active. For example, the 
packet CPOOL, or clause level packet, is always active. 
Since the hybrid parser has no packets, every rule is 
considered in parallel with the situation dictating which 
action should be taken. 
Removal of attention-shifting rules. PARSIFAL relies 
on attention-shifting rules to transparently build certain 
constituents, particularly NPs, which begin in the 
second buffer position. For example, in the sentence 
taken from Marcus: Have the students who missed the 
exam taken the makeup today?, the subject-aux inver- 
sion mechanism (switch) must be deliberately post- 
poned until the NP starting in the second position is 
analyzed as a complete constituent. Only then can the 
inversion take place. PARSIFAL solves this problem 
by temporarily shifting buffer positions so that the 
parser is viewing the buffer beginning in the second 
position. The second lefunost complete constituent (the 
NP) is then reduced before the first element constituent. 
We follow the lead of Berwick (1985) and others in our 
treatment of such cases by using the parse stack as a 
"movement stack" and stack the postponed item. Two 
actions, PUSH and DROP, are suitable for this purpose. 
In the example above, the end of the noun phrase, the 
students, can not be determined without applying the 
rules to the embedded clause. When complete, the NP 
is dropped into the buffer and the auxilim'y verb can be 
re-inserted into the buffer allowing the inversion can 
take place. Note that at no point is the "monotonic" 
property of determinism violated by undoing previous 
actions. 
Removal of rule priorities. In PARSIFAL, rules are 
ordered by priority. In the hybrid parser, rules have no 
priority. They compete with each other and the most 
relevant rule, based on training, wins the competition. 
Only one action, corresponding to the Iiring of one 
single-action rule, will be performed on each process- 
ing step. The current active node and its attachments 
12 2 
along with the contents of the two buffer cells is the 
basis for this decision. The rules are coded in such a 
way that every rule has a unique left-hand side and is 
thus relevant to situations most similar to its left-hand 
side pattern. 
Restriction of grammar rule format. The format of 
grammar rules in the hybrid parser is different from 
PARSIFAL in two ways. First, grammar rules are for- 
bidden to have more than a single action which is per- 
formed on the first buffer cell only; and second, rule 
patterns are defined to uniformly mention items in both 
buffer cells. 
Grammar actions. The repertoire of rule actions is 
slightly different in the hybrid parser. Actions such as 
ACTIVATE and DEACTIVATE have been removed. 
The basic actions are: 
a) ATTACH as <node>: The first item in the buffer is 
attached through an intermediate descriptive 
<node> to the current active node. 
b) CREATE <type>: Generates a new node of type 
<type> and pushes it onto the parse stack as the 
current active node. 
c) DROP: Pops a node or an item off the top of tile 
stack and inserts it into the buffer in the first buffer 
position. The previous contents of the buffer is 
shifted back by one position. 
d) INSERT <item>: Inserts the designated item into 
the buffer in the first buffer position. The previous 
contents of the buffer is shifted back by one posi- 
tion. In the general form, only a small number of 
designated lexical items (you, to, be, wh-marker) 
can be inserted. The special form INSERT TRACE 
inserts an (unbounded) NP trace. 
e) LABEL <feature>: Adds designated feature to the 
first buffer item. 
f) PUSH: Pushes an item onto the stack for temporary 
storage whenever the parse stack is used as a move- 
ment stack. 
g) SWITCH: Exchanges the items in the first and 
second buffer positions. 
These are the only actions the grammar rules can per- 
form. The buffer is manageA symbolically and if a 
position is vacated, an item is taken from the input 
stream to fill the position. The connectionist com- 
ponent can only examine the current active node, its 
immediate attachments, and the features of the first two 
buffer items. Once a node is attached to its parent, it 
can never again be examined. 
3.3. The Grammar 
The hybrid parser is capable of successfully processing 
a wide; variety of sentence forms such as simple 
declarative sentences, passives, imperatives, yes-no 
questions, wh-questions, wh-clauses, and other embed- 
ted sentences. The grammar to be learned by the sub- 
symbolic system, which has 73 rules, can be separated 
into base phrase structure rules and transformational- 
type rules. The base structure system can be further 
broken down into rules for NPs, VPs, auxiliaries, main 
sentence, PPs, and embedded sentences. 
Transformational rules fall into two groups: simple 
local transformations (like subject-aux inversion) and 
major movement rules like wh movement. In general, 
for each type of phrase, creation of the phrase (creating 
a new node on the active node stack) and completion of 
the phrase (dropping it into the buffer) is carried out by 
a separate grammar rule action. 
The rules for analyzing verb phrases discriminate 
among verbs that take different kinds of complements. 
For example, verbs that take a wh complement are 
discriminated from ones that take a that complement. 
Verbs like want that take either a missing or lexical 
subject in embedded sentential complements are 
separated from verbs like try or believe that do not take 
a lexical subject. Verbs that take one NP object are dis- 
tinguished from ones that take two NP objects through 
lexical features. 
4. Architecture of the Hybrid Parser 
The hybrid parser is composed of a connectionist net- 
work trained using backward propagation (Werbos 
1974; Rumelhart et al, 1986) from rule templates which 
,are derived from the deterministic grammar. Rule tem- 
plates are intermediate between symbolic rules and the 
training patterns required by the network. Each rule 
template typically represents a large number of pat- 
terns. They serve to relate situations that occur during 
parsing with the action deemed appropriate for that 
situation. Actions in the hybrid parser are performed 
symbolically on traditional data structures which are 
also maintained symbolically. 
As Figure 2 illustrates, the hybrid parser is organized 
into a symbolic component and a connectionist com- 
ponent. The latter component is implemented as a 
numeric simulation of an adaptive neural network. The 
symbolic and connectionist components cooperate in a 
tightly coupled manner since there are proven advan- 
tages to this type of organization (Kitzmiller and 
Kowalik, 1987). For the hybrid parser, the advantages 
are performance and robustness. 
The symbolic component manages the input sentence 
and the flow of constituents into the lookahead buffer, 
coding them as required for the input level of the net- 
work in the connectionist component. On the return 
side, it evaluates the activations of the output units, 
decides which action to perform, and performs that 
action, potentially modifying the stack and buffer in the 
process. The responsibility of the connectionist com- 
ponent, therefore, is to examine the contents of the 
buffer and stack and yield a preference for a specific 
action. These preferences are garnered from many 
iterations of back-propagation learning with instances 
of the rule templates. Learning itself occurs off-line 
and is a time-consuming process, but once learned the 
processing times for file system are excellent. Compu- 
tations need only flow in one direction in the network. 
The feed-forward multiplication of weights and compu- 
tation of activation levels for individual units produce 
the pattern of activation on the output level. Activation 
of output units is interpreted in a winner-take-all 
manner, with the highest activated unit determining the 
action to be taken. 
3 13 
CONNECTIONIST SYMBOLIC 
Coded Actions 
? 
Coded Stack & Buffer 
I 
/ -....... 
Buffer 
I Joh  Isho  l 
Stack 
<2 
r 
i 
i 
i 
i 
i 
i 
i 
k 
have scheduled the meeting. 
Np~~ VP 
MVB NP 
Figure 2: System Overview 
In tile set of experiments described here, the network 
has a three-layer architecture, as illustrated, with 66 
input units, 40 hidden units, and 40 output units. Each 
input pattern consists of two feature vectors from the 
buffer items and one vector from the stack. The first 
vector activates 26 input units and the second vector 
activates 12 input units in a pattern vector representing 
a word or constituent of the sentence. The stack vector 
activates 28 units representing the current node on the 
stack and its attachments. One hidden layer has proven 
sufficient in all of these experiments. The output layer 
permits the choice of one out of 40 possible actions that 
can be performed on a single iteration of processing. 
During sentence processing, the network is presented 
with encodings of the buffer and the top of the stack. 
What the model actually sees as input is not the raw 
sentence but a coded representation of each word in the 
sentence in a form that could be produced by a simple 
lexicon, although such a lexicon is not part of the 
model in its present form. The network produces the 
action to be taken which is then performed. If the 
action creates a vacancy in the buffer and if more of the 
sentence is left to be processed then the next sentence 
component is moved into the buffer. The process then 
repeats until a stop action is performed, usually when 
the buffer becomes empty. Iteration over the input 
stream is achieved in this fashion. 
Figure 2 illustrates the nature of the processing, 
although it shows a composite of the initial and final 
states of the parser. When a sentence form like "John 
should have scheduled the meeting" appears in the 
input stream, the first two constituents fill the buffer. 
These contents along with the contents of the top of the 
stack and its attachments are encoded and presented to 
the network. Coding is based on a simple scheme in 
which those features of the buffer and stack that are 
actually tested by grammar rules are represented (see 
Faisal, 1990). The network, in turn, produces a single 
action. Specification of the action by the network is 
done by activating one of the output units more than the 
others thus determining the winner (called "winner- 
take-all"). This action is then executed symbolically, 
yielding changes in the buffer and stack. The process 
repeats until a stop action is performed at which time 
the resultant parse structure is left on top of the stack as 
shown. 
4.1. Learning a Grammar 
Training of the hybrid parser proceeds by presenting 
patterns to the network and teaching it to respond with 
an appropriate action. The input patterns represent 
encodings of the buffer positions and the top of the 
slick from the deterministic parser. The output of the 
network contains a series of units representing actions 
to be performed during processing and judged in a 
winner-take-all fashion. Network convergence is 
observed once the network can achieve a perfect score 
on the training patterns themselves and the error meas- 
ure has decreased to an acceptable level (set as a 
parameter). Once the network is trained, the weights 
are stored in a file so that sentences can be parsed. A 
sentence is parsed by iteratively presenting the network 
with coded inputs and performing the action specified 
by the network. 
Our neural network simulator features a logistic func- 
tion that computes values in the range of -1 to +1. 
Each grammar rule is coded as a training template 
which is a list of feature values. In general, each con- 
stituent is represented by an ordered feature vector in 
which one or more values is ON(+1) for features of the 
form and all other values are either OFF(-1) or DO 
NOT CARE (?). A rule template is inslintiated by ran- 
domly changing ? to +1 or -1. Thus, each template 
can be instantiated to give many training patterns and 
each training epoch is slightly different. It is obviously 
impossible to test the performance of all these cases, so 
for the purpose of judging convergence, a zero is sub- 
stituted for each ? in the rule template to provide test- 
ing patterns. For more discussion of the training pro- 
cess, see Faisal and Kwasny (1990). 
14 4 
TABLE 1 
Examples of Grammatical Sentences 
Sentence Form 
(1) Scheduled a meeting for Monday. 
(2) John has scheduled the meeting for Monday. 
(3) The meeting seems to have been scheduled for Monday. 
(4) The jar seems broken. 
(5) I persuaded John to do it. 
(6) I saw him do it. 
(7) Ma131 wants John to have a party. 
(8) Mary wants to have a party. 
(9) What will the man put in the comer? 
(10) What will the man put the book in? 
(11) Who (lid John see? 
(12) Who broke the jar? 
(13) Who is carrying the baby? 
(14) What is the baby carrying? 
(15) What did Bob give Mary? 
(16) The man who wanted to meet Mary has disappeared. 
(17) The: man who hit Mary with a book has disappeared. 
(18) The man whom Mary hit with a book has disappeared. 
(19) I told that boy that boys should do it. 
(20) That mouse that the cat chased had squeaked. 
(21) I told Sue you would schedule the meeting. 
(22) I told the girl that you would schedule the meeting. 
(23) John is scheduling the meeting for Monday. 
5. Performance 
For testing purposes, several sentences are ceded that 
would parse correctly by the rules of the deterministic 
parser. Additionally, severed mildly ungrammatical and 
lexical ambiguous sentences are coded to determine if 
the network would generalize in any useful way. Most 
of these examples were drawn from work cited earlier 
by Chamiak and Milne. The objective is to discover 
exactly how syntactic context can aid in resolving such 
problems. In previous work, a simpler (23-rule) gram- 
mar was tested with similar results (Kwasny and 
Faisal,1989). 
5.1. Parsing Grammatical Sentences 
Experimentation with grammatical sentences confirms 
that indeed the rules from the grammar have been 
learned sufficiently to parse sentences. When training 
with the rule templates, testing for convergence is pos- 
sible by changing each ? to a zero value. Here the per- 
formance of the hybrid parser is examined with actual 
sentences and the claim that the parser simulates both 
PARSIFAL and LPARSIFAL is substantiated. 
Gramrnatical sentences, by our definition, are those 
which parse correctly in the rule-based grammar from 
which the training set is derived. Table 1 shows several 
examples of grammatical sentences which are parsed 
successfully. Parse trees are developed which are 
identical with ones produced by other deterministic 
parsing systems. 
5.2. Parsing Ungrammatical Sentences 
Capabilities described above only duplicate what can 
be done rather comfortably symbolically. Of course, 
the feedforward network in the hybrid parser allows 
very fast decision-making due to the nature of the 
model. But what other features does the model pos- 
sess? Importantly, how robust is the processing? As a 
TABLE 2 
Ungrammatical vs. Grammatical Sentences 
Sentence Form Strength 
(la) *John have should scheduled the meeting for Monday. 14.4 
(lb) John should have scheduled the meeting for Monday. 56.9 
(2a) *Ilas John schedule the meeting for Monday? 32.3 
(2b) Itas John scheduled the meeting for Monday? 36.8 
(3a) *John is schedule the meeting for Monday. 9.5 
(3b) John is scheduling the meeting for Monday. 54.7 
(4a) *John is is scheduling the meeting for Monday. 7.2 
(4b) John is scheduling the meeting for Monday. 54.7 
(5a) *The boy did hitting Jack. 14.8 
(5b) The boy did hit Jack. 137.7 
(6a) *'llae meeting is been scheduled for Monday. 559.6 
(6b) The meeting has been scheduled for Monday. 565.5 
symbolic model, PARAGRAM extends PARSIFAL to 
handle ungrammatical sentences. This is accomplished 
by considering all rules in parallel and scoring each test 
performed on the left-hand side of a rule according to 
predefined weights. The rule with the best score fires. 
In this way, processing will always have some rule to 
fire. Reported experimentation with PARAGRAM 
shows this to be an effective method of extending the 
inherent capabilities of the grammar. 
To demonstrate its generalization capabilities, the 
hybrid parser is tested with several exmnples of 
ungrammatical sentences. Its performance is strictly 
dependent upon its training experiences since no relax- 
ation rules (Kwasny and Sondheimer, 1981), meta-rules 
(Weischedel and Sondheimer, 1983), or other special 
mechanisms were added to the original grammar rules 
to handle ungrammatical cases. In Table 2, ungram- 
matical sentences used in testing are shown along with 
their strengths. These strengths are computed as the 
reciprocal of the average error per processing step for 
each sentence and reflect the certainty with which indi- 
vidual actions for building structures are being selected. 
Although there is no real meaning in the values of these 
numbers, they are a useful means of comparison. 
These examples produce reasonable structures when 
presented to our system. Note that overall average 
strength is lower for ungrammatical sentences when 
compared to similar grammatical ones. 
In sentence (la), for example, the structure produced 
was identical to that produced while parsing sentence 
(lb). The only difference is that the two auxiliary 
verbs, have and should, were reversed in the parse tree. 
Sentence (2a) contains a disagreement between the 
auxiliary has and the main verb schedule and yet the 
comparable grammatical sentence (2b) parsed identi- 
cally. Sentences (3a) and (4a) parse comparable to sen- 
tence (3b). Sentence (5a) is processed as if it were pro- 
gressive tense ('The boy is hitting Jack'). In 
PARAGRAM, a nonsensical parse structure is pro- 
duced for this sentence, as reported by Charniak (p. 
137). It can be compared with sentence (5b), but there 
is not one clear choice for how the sentence should 
appear if grammatical. The problems with using a 
syntax-based approach to handling ungrammatical sen- 
tences are well-known (see, for example, Kwasny, 
1980). 
5 15 
TABLE 3 
Lexically Ambiguous vs. Unambiguous Sentences 
Sentence Form Strength 
(la) <Will> John schedule the meeting for Monday? 5.0 
(lb) Will(aux) John schedule the meeting for Monday? 57.46 
(2a) Tom <will> hit Mary. 29.8 
(2b) Tom will(aux) hit Mary. 125.8 
(3a) Tom <hit> Mary. 13.6 
(3b) Tom hit(v) Mary. 29.5 
(4a) The <will> gave the money to Mary. 16.6 
(4b) The will(noun) gave the money to Mary. 61.9 
(5a) They <can> fish(np). 20.6 
(5b) They can(v) fish(np). 30.0 
(6a) They can(aux) <fish>. 2.9 
(6b) They can(aux) fish(v). 6.3 
5.3. Lexical Ambiguity 
As a further test of the generalization properties of the 
hybrid parser, sentences containing lexically ambigu- 
ous words are tested. Some of these sentences are 
shown in Table 3. Of course, ROBIE takes a symbolic 
approach in extending PARSIFAL to address these 
issues by requiring additional rules and lexical features. 
Note that in the deterministic approach, it is essential 
for lexical items to be properly disambiguated or back- 
tracking will be required. 
In testing the hybrid parser, normal sentences are 
presented, except that selected words are coded ambi- 
guously (here indicated by angle brackets < > around 
the word). Sentences containing words followed by 
parentheses are presented to the hybrid parser unambi- 
guously, even though these words have ambiguous 
uses. The lexical choices are shown in parentheses. In 
the cases shown, the lexically ambiguous words were 
correx:tly interpreted and reasonable structures resulted, 
although lower strengths were observed. The hybrid 
parser utilizes syntactic context to resolve these ambi- 
guities and automatically works to relate novel situa- 
tions to training cases through the generalization capa- 
bility of the network. As before, no additional rules or 
mechanisms are required to provide this capability. 
Sentence (la) contains the word will coded ambigu- 
ously as an NP and an auxiliary, modal verb. In the 
context of the sentence, it is clearly being used as a 
modal auxiliary and the parser treats it that way as (lb). 
A similar result was obtained for sentence (2a) which 
parses as (2b). In sentence (3a), hit is coded to be 
ambiguous between an NP (as in a baseball hit) and a 
verb. The network correctly identifies it as the main 
verb of the sentence as in sentence (3b). Sentence (4a) 
is constructed as for sentence (4b). Sentence (5a) 
presents can ambiguously as an auxiliary, modal, and 
main verb, while fish is presented uniquely as an NP. 
Can is processed as the main verb of the sentence and 
results in the same structure as sentence (5b). Like- 
wise, sentence (6a), which contains fish coded ambigu- 
ously as a verb/NP and can coded uniquely as an auxi- 
liary verb, produces the same structure as sentence 
(6b). In the cases shown, the lexically ambiguous 
words were disambiguated and reasonable structures 
resulted. Note that the overall average strengths were 
lower than comparable grammatical sentences dis- 
cussed, as expected. 
6. Summary 
Our hybrid deterministic parser is based on a deter- 
ministic grammar modified slightly from that found in 
traditional systems. Our grammar is derived from one 
used by Marcus, but with much inspiration from the 
work of Milne, Berwick, and Chamiak. The rules of 
the grammar are utilized in training a connectionist 
component. The result is a hybrid system which exhi- 
bits characteristics from several well-known extensions 
of the basic deterministic parser. In particular, some 
ungrammatical and lexically ambiguous inputs can be 
successfully processed although no special provisions 
are made for them. These extended properties come 
essentially for free due to the coupling of a symbolic 
component with connectionism. 

References 

Berwick, R.C. 1985. The Acquisition of Syntactic 
Knowledge. MIT Press, Cambridge, MA. 

Charniak, E. 1983. "A Parser with Something for Every- 
one." In Parsing Natural Language, M. King, ed. Academic 
Press, New York, NY, 117-150. 

Faisal, K.A. 1990. Cormectionist Deterministic Parsing. 
D.Sc. Dissertation, Department of Computer Science, Wash- 
ington University, St. Louis, Missouri. 

Faisal, K.A. and S.C. Kwasny. 1990. Deductive and Induc- 
tive Learning in a Connectionist Deterministic Parser. Proc 
lntl Joint Conf Neural Networks, 1~471-474. 

Kitzmiller, C.T., and J.S. Kowalik. 1987. Coupling Symbolic 
and Numeric Computing in Knowledge-Based Systems. AI 
Magazine 8, no. 2, 85-90. 

Kwasny, S.C. and K.A. Faisal. 1989. Competition and 
Learning in a Connectionist Deterministic Parser. Proc llth 
Conf Cog Sci Society, 690-697. 

Kwasny, S.C. and N.K. Sondheimer. 1981. Relaxation Tech- 
niques for Parsing Ill-Formed Input. Am J Comp Ling 7, no. 
2, 99-108. 

Kwasny, S.C. 1980. "Treatment of Ungrammatical and 
Extra-Grammatical Phenomena in Natural Language Under- 
standing Systems." Indiana University Linguistics Club, 
Bloomington, Indiana. 

Marcus, M. P. 1980. A Theory of Syntactic Recognition for 
Natural Language. MIT Press, Cambridge, MA. 

McClelland, J. L., & A. H. Kawamoto. 1986. "Mechanisms 
of Sentence Processing: Assigning Roles to Constituents of 
Sentences." In Parallel Distributed Processing, D.E. 
Rumelhart and J.L. McClelland, MIT Press, Cambridge, MA, 
272-325. 

Milne, R. 1986. Resolving Lexical Ambiguity in a Deter- 
ministic Parser. Comp Ling 12, No. 1, 1-12. 

Rumelhart, D. E., G. Hintoh, and R.J. Williams. 1986. 
"Learning Internal Representations by Error Propagation." 
In Parallel Distributed Processing, D.E. Rumelhart and J.L. 
McCMland, MIT Press, Cambridge, MA, 318-364. 

Weischedel, R.M. and N.K. Sondheimer. 1983. Meta-Rules 
as a Basis for Processing Ill-Formed Input. AJCL
9, No. 3-4, 161-177. 

Werbos, P. 1974. "Beyond Regression: New Tools for Pred- 
iction and Analysis in Behavioral Science." Ph.D. Thesis. 
Harvard University, Cambridge, Ma. 
