A Corpus-Based Learning Technique for Building A 
Self-Extensible Parser 
Rey-Long Liu and Von-Wun Soo 
Department of Computer Science 
National Tsing-Hua University 
HsinChu, Taiwan, 1LO.C. 
e-mail: sooQ<s.nthu.edu.tw 
Abstract 
IIuman intervention and/or training corpora tagged 
with various kinds of information were often assumed 
in many natural language acquisition models. This 
assumption is a major source of inconsistencies, er- 
rors, and inefficiency in learning. In this paper, we 
explore the extent to which a parser may extend it- 
self without relying on extra input from the outside 
world. A learning technique called SEP is proposed 
and attached to the parser. The input to SEP is raw 
sentences, while the output is the knowledge that is 
missing in the parser. Since parsers and raw sentences 
are commonly available and no human intervention is 
needed in learning, SEP could make fully automatic 
large-scale acquisition more feasible. 
Keywords: fully automatic natural language acqui- 
sition, self-extensible parser, corpus-based learning 
1 Introduction 
It is commonly believed in many psycholinguistics 
studies \[Pinker, 1984; Wexler & Culicover, 1980\] that 
extra input (in addition to raw sentences) is necessary 
for human language learners. Most existing compu- 
tational natural language acquisition models also as- 
sumed various kinds of the extra input (e.g. semantic 
associations \[Siskind, 1990; Webster & Marcus, 1989; 
Zernik, 1987\] and syntactic structures \[Berwick, 1985; 
Liu & Soo, 1992a\] of input sentences) and human 
intervention (e.g. information interactively given by 
the trainer \[tang & Ilirschman, 1988; Velardi et al., 
1991\]) during learning. The preparation of tile ex- 
tra input and human intervention may often cause 
inconsistencies, errors, and inefficiency in learning. It 
is often a bottle neck in scaling up natural language 
processing systems. 
Therefore, simple syntactic heuristics had been 
used to collect the extra input \[Brent, 1993; Sekine, 
1992\]. However, the information that may be collect- 
ed by the simple heuristics is limited. More sophisti- 
cated processors such as taggers and parsers had also 
been tested in collecting the extra input \[Webster & 
Marcus, 1989; Zernik, 1989\]. ttowever, since learning 
was based on the input that may be successfully and 
unambiguously analyzed by the processors, upgrad- 
ing tile performance of the processors became a new 
bottle neck of upgrading the performance of learning. 
Furthermore, since the heuristics and proccssors are 
separated from the learning component, the learning 
component cannot know what knowledge is actually 
missing in the heuristics and processors. The learn- 
ing component might learn the knowledge that the 
processors already have. 
In this paper, we investigate the robust ways of ac- 
quiring parsing knowledge without requiring extra in- 
put and truman intervention. The input to the system 
is raw sentences, while the output is the knowledge 
that is missing before learning. The parser takes an 
active role in collecting extra information for learn- 
ing. Thus a parser could be self-extensible in the 
sense that it couhl automatically acquire what it ac- 
tually lacks without relying on extra information from 
the outside world. The parsing capability is improved 
through learning, and in turn, the learning capability 
is improved due to the improved parsing capability. 
To achieve that, learning should be triggered when 
the parser fails to analyze input sentences. In that 
case, however, there might be a large number of hy- 
potheses to fix the failures. For example, suppose a 
parser finds an unknown word when parsing a sen- 
fence. In that sentence, tile unknown word may have 
many possible syntactic and semantic behaviors (e.g. 
parts-of-speech and argument structures) which may 
lead to many different sentence structures. Therefore, 
the major challenge of the study is the generation and 
verification of the hypotheses for missing knowledge. 
We thus propose a learning technique called SEP 
that is attached to the parser. SEP is triggered when 
tile parser fails in parsing raw sentences. In the next 
two sections we describe SEP's two modules: the hy- 
pothesis generation module and the hypothesis veri- 
fication module. For each training sentence, the two 
modules are triggered sequentially. Experimental re- 
sults are then presented in section 4. 
441 
2 Generation of hypotheses 
SEP generates hypotheses based on the partial results 
of failed parsing, universal linguistic constraints, and 
the parser's existing knowledge. 
2.1 Collecting partial results of pars- 
ing 
As the parscr's knowledge is incomplete for parsing an 
input sentence, the bottom-up chart parsing strategy 
is suitable in collecting partial results of parsing. The 
failures of parsing are due to the lack of some knowl- 
edge pieces for grouping the partial results into the 
target constituent (e.g. a major sentence Smaj). 
As an example, consider the sentence "taking exer- 
cises is good for your health". A parser with a com- 
plete knowledge base for parsing the sentence may 
construct a constituent Smaj after deriving all rele- 
vant constituents such as noun phrases (NPs) and ver- 
b phrases (VPs). The grouping of constituents should 
be based on both syntactic and semantic constraints. 
As the parser does not have adequate knowledge for 
parsing the sentence, failures will occur. For example, 
parsing will fail if "take" is an unknown word for the 
parser. There might be a large number of hypothe- 
ses to fix the failure. For example, "take" may be a 
noun or a verb. The learner may even hypothesize 
that Smaj may be constructed by matching the se- 
quence "taking NP VP". That is, without the help of 
additional information, a huge number of ridiculous 
hypotheses might be generated. 
2.2 Generating hypotheses based on 
universal constraints and the pars- 
er's knowledge 
There are universal linguistic constraints that may 
restrict the forms of missing knowledge. The X-Bar 
theory, for example, postulates that any maximal pro- 
jection (constituent) sbould be composed of at most 
three components: a specifier, an argument struc- 
ture of the lexical head (i.e. X-bar), and a modifier 
\[Chomsky, 1981\]. The set of possible subcategoriza- 
tion frames of lexical heads had also been set-up in 
many studies \[Gazdar, 1985\]. Based on these studies, 
SEP allows a constituent to have at most three com- 
ponents. This constraint (called Three-Components 
Constrttint ) is incorporated into SEP's hypothesis 
generation process which is composed of two phas- 
es: the top-down phase and the bottom-up phase. 
The bottom-up phase is triggered after the top-down 
phase is completed. 
2.2.1 The top-down phase 
In the top-down phase, SEP uses the parser's existing 
knowledge to perform top-down prediction of missing 
knowledge. This phase is for the case in which the 
parser has knowledge for constructing a constituent 
(e.g. Smaj may be constructed by an NP followed by 
a VP) excepts the knowledge for constructing all the 
individual components (e.g. the NP and the VP) of 
the constituent. 
The input sentence: 
Taking exercises is good for your health. 
Constituents already constructed: 
NP(2-2), VP(3-4), NP(6-7), PP(5-7), VP(3-7) 
The top-down phase: 
Pass 1: Sinai(i-7) :- NP(1-2), VP(3-7). 
pass 2: Ne0-2) :- VPO-2). 
Pass 3: VP(1-2) :- ?? 
The bottom-up phase: 
Step 1: Completed constituents: NP(2-2) 
Step 2: Hypothesis: VP(1-2) :- verb(i-i), NP(2-2) 
Step a: Checking the Three-Components Constraint 
Step 4: If valid, return the hypothesis 
Fig.1. An example trace of SgP's hypothesis generatiml 
As an example, consider "taking exercises is good 
for your health". Suppose "take" is an unknown 
word. Thus parsing fails, and SEP is triggered. The 
reasoning process is illustrated in Fig.1. In Fig.l, the 
numbers denote the starting positions and ending po- 
sitions of constituents, and the constituents that can- 
not be constructed in parsing are marked in the bold- 
face form. In the top-down phase, SEP first searches 
for the parser's knowledge pieces for constructing the 
top level goal constituent Sinai. Suppose one of the 
knowledge pieces says that Smaj may be constructed 
by an NP followed by a VP (Pass 1 in Fig.l). Sin- 
ce the VP may be instantiated by "is good for your 
health", SEP expects there should be an NP from po- 
sition 1 to position 2. Thus SEP retrieves all knowl- 
edge pieces for constructing NPs. Suppose that one 
of them says that a predicate NP may be constructed 
from a VP (Pass 2 in Fig.l). Thus, SEP attempts to 
retrieve VP rules. Ilowew.'r, since "take" is unknown, 
no knowledge may be retrieved for constructing the 
VP (Pass 3 in Fig.l). The top-down phase thus stops 
and the bottom-ul) phase i,~ triggered. 
2.2.2 The bottom-up phase 
In the bottom-up phase, SEP uses the partial results 
of parsing to perform bottom-up prediction of miss- 
ing knowledge. Thin phase is for the case in which 
the parser has knowledge for constructing all the in- 
dividual components (e.g. an NP and a VP) of a 
constituent (e.g. Sinai) excepts the knowledge for 
grouping these components (e.g. Smaj may be con- 
structed by the NP followed by the VP). 
For the above example, SEP has hypothesized that 
there is a VP from position 1 to position 2. In the 
bottom-up phase, SEP first observes the partial re- 
442 
sults of parsing from position 1 to position 2. S- 
ince only the NP "exercises" is constructed in this 
range (Step 1 in Fig.l), the NP is the only possible 
argument in the VP. Thus the hypothesis "VP(1-2) 
:- verb(l-l), NP(2-2)" is generated (Step 2 in Fig.l). 
Then the Three-Components Constraint is checked 
(Step 3 in Fig.l). Since the hypothesis satisfies the 
constraint, it may be returned as a hypothetic knowl- 
edge piece (Step 4 in Fig.l). If the hypothesis is con- 
firmed (see section 3), SEP acquires both a category 
and an argument structure of "take". 
Top-down search 
(/br conclu,vion part) } 
• C3 :- Clp, Cn. 
T Bottom-up search 
(for condinon parO 
Fig.2. The top-down phase and the bottom-up phase 
iv summary, the top-down phase and the bottom- 
up phase of SEP are complementary to each other. 
in the top-down phase, SEP hypothesizes the conclu- 
sion parts of the missing rules, while in the bottom-up 
phase, SEP hypothesizes the condition parts of the 
missing rules. A schematic view of the two phases is 
illustrated in Fig.2. The top-down phase starts from 
the constituents not constructed (marked with bold- 
face circles) to the constituents already constructed. 
The bottom-up phase starts from the constituents al- 
ready constructed to the constituents not construct- 
ed. The two phases meet at the possible failure points 
of parsing. A failure point indicates that there is a 
missing rule in the parser. 
It should be noted that, in generating hylmthc:ms , 
SEP might need to consider several reasoning trees 
such as the one in Fig.2. Tt, is is because there ntight 
be several rules for constructing a constituent (e.g. 
Smaj). Each rule indicates a path of reasoning, anti 
thus leads to a new reasoning tree. 
If the parser has only one missing rule for parsing 
a sentence, the top-down phase and the bottom-up 
phase will be able to meet at the corresponding fail- 
ure point. This is because in that case SEP will have 
enough rules (including grammar or lexical rules) to 
perform top-down traversal and enough constituents 
(already constructed in parsing) to perform bottom- 
up traversal. On the other hand, if the parser has 
more titan one missing rules for parsing a sentence, 
the top-down phase will stops at the points which the 
bottom-up phase cannot reach. IV'or example, if both 
Cp and Cn in Fig.2 cannot be constructed in parsing, 
the two phrases cannot meet. In that case, tt,e ambi- 
guity space may be too large to resolve. That is, we 
may have: 
The Cornplete-VChen-One-Misslng Theorem: 
Supoose the parser lacks only one rule R (either a 
lexical nile or a grammar rule) to completely parse a 
sentence. Then R will be included in the set of hy- 
potheses generated by SEP. Ttmt is, if the input sen- 
tertce is "not difficult" for the parser to learn knowl~ 
edge from it, the missing rule will be generated by 
SEP. 
Therefore, in each step of finding an existing rule to 
perform top-down traversal, SEP selects a rule only 
when all but one of the components in the condition 
part of the rule are constructed in parsing. If no such 
rules may be found, the top-down phase stops and the 
bottom-up phase is triggered. If the input sentence is 
not too difficult for the learner, the bottom-up phase 
and the top-down phase can meet at a failure point, 
and thus hypotheses of missing knowledge may be 
generated. If the parser has only one missing rule for 
parsing the input sentence, the missing rule will be 
in the hypothesis set generated. This from-simple-to- 
difficult learning sequence is assumed in many learn- 
ing models (and human learners as well). Since raw 
sentences are commonly available, SEP may easily get 
suitable training sentences. 
3 Verification of hypotheses 
SEP's hypothesis generation module might generate 
several hypotheses whose validities should be verified 
before assimilated into the parser's knowledge base. 
The algorithm of the hypothesis verification modnle 
is outlined in Fig.3. 
Algorithm: SEP's hypothesis verification 
lnlmt: Sets of hypothem~s generated 
Ontput: A hypothesis of the target missing rule 
Begin: 
For m~ch hypothesis .'let lI genen~ted for It sentence 
If there is only one hypothesis in II, return the hy- 
pothesis; (Step 1) 
Otherwise, increment the frequency of each hypoth- 
esis h in it by one; (Step 2) 
Return the hypothesis with the highest frequency; 
(Step a) 
End. 
Fig.3. 'File algorithm of SEWs hypothesis verification 
SEWs hytmthesis verification module makes decision 
based on the hypothesis sets generated for training 
sentences (one hypothesis set per training sentence). 
If there is only one hypothesis in any one of the hy- 
pothesis sets, SEP returns the hypothesis as the tar- 
443 
get missing knowledge (Step 1). Since only one hy- 
pothesis in each hypothesis set may be confirmed (i.e. 
only one rule is missing), other hypotheses in the hy- 
pothesis sets may be excluded. 
If more than one hypotheses are generated for a 
training sentence, the frequency of each of the hy- 
potheses is updated (Step 2). After considering all 
the hypothesis sets, SEP returns the hypothesis with 
the highest frequency of occurrence (Step 3). It is 
obvious that, a hypothesis with a higher frequency of 
being generated is more likely to be the target missing 
knowledge. 
As a hypothetic knowledge piece is confirmed, it 
should be annotated with critical syntactic and se- 
mantic features \[Liu & Soo, 1992a, 1992b\]. A knowl- 
edge piece without suitable feature annotation will 
be too general and thus useless. For example, sup- 
pose the learner acquires a knowledge piece for con- 
structing a predicate NP from a VP (e.g. "taking 
exercises"). It must annotate the NP with the fea- 
ture "NUM=singular'; otherwise the ungramrnatical 
sentence "taking exercises are good for your health" 
will be accepted as well. The annotation is based on 
universal linguistic principles such as the universal 
feature instantiation principles in generalized phrase 
structure grammar (GPSG \[Gazdar, 1985\]). For ex- 
ample, the feature "NUM=singular" is annotated by 
observing the fact that the verb "is" needs a singular 
external argument. 
4 Experiment 
In the experiment, SEWs hypothesis generation and 
verification modules were evaluated. We used a pars- 
er whose knowledge base included 2513 lexicon en- 
tries, 22 grammar rules, and 20 morphological rules. 
4.1 Evaluation of the hypothesis gen- 
eration module 
To compare the performance of SEP with a learner 
that is provided with extra input, a set of sentences 
that had been tested in previous experiments \[Liu & 
Soo, 1993a, 1992a, 1992b\] was entered to SEP. The 
difference was that, only raw sentences were entered 
to SEP. There were 165 sentences (about 1700 words) 
in the corpus. These sentences were majorly extract- 
ed from two articles of an English textbook. 
Among the 165 sentences, 80 sentences were suc- 
cessfully parsed by the parser, and hence SEP was 
not triggered for them. SEP was triggered for acquir- 
ing the missing knowledge for parsing the other 85 
sentences. There were totally 202 hypotheses gener- 
ated. Thus, on average SEP produced 2.38 (202/85) 
hypotheses per input sentence that cannot be parsed 
by the parser. Furthermore, among the 85 sentences 
that triggered learning, SEP successfully generated 
hypotheses from 55 sentences. That is, the 55 sen- 
tences were not too difficult for the parser. From this 
point of view, SEP produced 3.67 (202/55) hypothe- 
ses per missing rule. Therefore, SEP needs to collect 
more evidences in order to determine a target missing 
rule among 3.67 hypothetic knowledge pieces. 
(1) AP(3-11) :- NP(3-5), S(6-11). 
(2) NP(3-11):- NP(3-5), S(6-11). 
(3) VP(2-11) :-is(2-2), NP(3-5), S(6-11). 
(4) NP(1-6):- S(1-5), NP(6-6). 
(5) s(1-~1) :- s(1-5), s(~-n). 
(6) Sml~j(1-11) :- S(1-5), S(6-11). 
Fig.4. An example of the hypotheses generated by SEP 
As an example, consider the sentence "Lead is a soft 
metal that serves rnany purposes in home" in the cor- 
pus. The parser had a missing rule for constructing 
NPs with relative clauses. Six hypotheses were gen- 
erated by SEP They are illustrated in Fig.4 (for the 
illustrative purpose, syntactic and semantic features 
are omitted). IIypothesis (1) and (2) were generat- 
ed based on the existing argument structures of "is"; 
SEP thought that if an AP (Adjective Phrase) or 
an NP may be constructed from position 3 to po- 
sition 11, parsing may become successful, lIypothesis 
(3) was generated for learning a new argument struc- 
ture of"is". Ilypothesis (4) was generated when SEP 
hypothesized "serve" as the main verb of the sen- 
tence, tlypothesis (5) and (6) were generated since 
SEP thought the parser might need to know a new 
sentence structure. Among tim six hypotheses, the 
target missing rule is hypothesis (2), which is quite 
likely to have the highest frequency of being generat- 
ed in learning on a large corpus of sentences. 
4.2 Evaluation of the hypothesis veri- 
fication module 
In the experiment, we evaluated SEWs performance 
in hypothesis verification. Training sentences were 
extracted from the DJ corpus (Wall Street Journal 
articles). The size of the corpus was about 32 mega 
bytes. Since SEP only assmned raw sentences as 
input, other kinds of information (e.g. the part-of- 
speech information) in the DJ corpus were not con- 
sidered in the experiment. 
For verifying the hypotheses generated for parsing 
the above sentencc "Lead is a soft metal that serves 
many purposes in home" (Fig.4), 1000 sentences con- 
taining "that" were extracted from 19200 sentences 
(71294 words) of the DJ corpus. The 1000 sentences 
were fed into tile parser. As described above, since 
the knowledge for the NPs with relative clauses was 
missing, parsing failed and SEP was triggered for each 
sentence. In many cases, SEP could not generate any 
hypotheses, since there were many unknown words for 
444 
the parser. As hypotheses could be generated, SEP 
updated the frequencies of tile hypotheses (Step 2 in 
Fig.3). As learning proceeded to the sentence: 
... SEC, an agency that covets its independence, ... 
SEP generated only one hypothesis "NP :- NP, S". 
This was because, "an agency ..." was unambigu- 
ously segmented by commas and expected to be a 
noun phrase. Since only one hypothesis was generat- 
ed, tl, e hypothesis was returned as the target missing 
knowledge (Step 1 in Fig.a). ~l'hat is, SEP concluded 
that hypothesis (2) in Fig.4 was needed (and miss- 
ing in tile parser) for parsing both sentences. There- 
fore, although the sentences that are too difficult for 
the parser are skipped, SEP may still tind suitable 
sentences to learn since raw sentences are commonly 
available. 
The current version of SEP may be extended in the 
following two ways: 
• Acquisition of movement constructions: Movement 
constructions cannot he learned using SEP. For exam- 
ple, the movement construction in the sentence "IIere 
comes the dog" cannot be learned, since it cannot be 
detected if only raw sentences are entered to the learn- 
er \[Liu& Soo, 1992b\]. 
• The use of more universal linguistic constraints: 
More suitable universal linguistic constraints may be 
used to both reduce the number and promote the 
quality of the hypotheses generated. For example, 
hypothesis (1) in Fig.4 may be filtered out by consult- 
ing the fact that the head component of a constituent 
should be included in the condition parts of the rules 
for constructing the constituent. Since neither NP 
nor S may be the head of AP, the hypothesis may be 
discarded. As another example, according to X-Bar 
theory, each constituent is composed of a specifier, 
an argument structure, and a moditier. The possi- 
ble syntactic categories (e.g. NP, PP, VP, and AP) 
of the components of each kind of constituents have 
been identified in previous linguistics studies (e.g. a 
determiner may be the specifier of art NP). It is obvi- 
ous that, if SEP generates hypotheses hy considering 
the universal constraints, the quality of the generated 
hypotheses may be promoted. 
5 Related work 
Previous natural language acquisition models couht 
be characterized as interactive acquisition \[tang & 
tIirschman, 1988; Liu & Soo, I993a, 1993b; Velardi ct 
al., 1991\], corpus-based acquisition \[Brent, 1993; 3a- 
cobs & Zernik, 1988; Zernik, 1989\], dictionary-based 
acquisition \[Montemagni & Vanderwende, 1992; San- 
filippo & Pozanski, 1992\], statistics-based acquisition 
\[Smadja, 1991; Sekine et al, 1992\], and connectionist- 
based acquisition \[Paisal & Kwasny, 1990; McClelland 
& Kawamoto, 1986\]. 
Our motivation in the study is to provide the parser 
with the capability of extending itself without using 
extra input and intervention fi'om tile outside worht. 
From this point of view, interactive acquisition and 
connectionist-based acquisition will have difficulties 
in resolving tile problems of inconsistencies, errors, 
and inefficiency of learning, since tttey required the 
information encoded by the trainer. 
SEP could be characterized as corpus-based acqui- 
sition. '\]'he point here is that S~P only assumes 
raw sentences as the input. From the point of view, 
corpus-based acquisition and statistics-based acqui- 
sition that require pre-processed data will have dif 
ficulties in getting adequate extra input, since raw 
sentences are nmch more commonly available than 
pre-processed data. 
SEP collects observations from the parser which 
may grow through learning. From this point of 
view, collecting information using simple but non- 
extensible heuristics \[Brent, 1.99:1\] might miss many 
opportunities of learning (due to the inadeqnacy of 
the collected information), although raw sentences are 
assumed as the major input in that study as well. 
In addition to grammar and lexical information 
from the parser, other types of useful information 
slay include contextual, conceptual, and association- 
al semantic information \[SNkind, 1990;//aeons & Zer- 
I, ik, 1988; Wcbster & Marcus, 1989; Zernik, 1987\], al- 
though they are much more difliculL to collect in prac- 
tice (especially when the parser is incomplete and tile 
input sentences are noisy in many real-world applica- 
tions). '1'his paper explores the feasibility of allow- 
tug a parser (either preliminary or sophisticated) to 
extend itself with available raw sentences and infor- 
mation from itself. Noise-tolerant learning is imple- 
mented by allowing the learner to acquire knowledge 
hased on a large number (rather than one) of obser- 
vations. The practical learning method may make 
eftk:ien\[;, fully-antomatic, and large-scale acquisition 
ntore approachable. 
Parsers (or taggers) had been used in many pre- 
vious models as wel\]. They were pre-processors of 
learning \[Zernik & Jacobs, 199(I; Pustcjovsky et al., 
1199;t; Montemagni & Vanderwende, 19921 Sanfilippo 
& Pozanski, 1992; Zernik, 1989\] or post-processors of 
learning \[Smadja, \]991\]. In those models, the proces- 
sors were assumed to be complete and thus separated 
from the h;arning components. Learning was based 
on the "success" of syntactic processing. '1'o extend 
the capability of the processors, learning should be 
triggered when the processors fail to analyze input 
sentences. This is the reason why SEP atternpts to 
learn knowledge when the parser fails in parsing raw 
sentences. For those models assuming syntactic infor- 
mation its their necessary input, SEP may be attached 
to them to wake their processors more self-extensible. 
44.5 
We are currently integrating SEP with a framework 
of syntactic and semantic knowledge acquisition \[Liu 
& Soo, 1992a; Liu & Son, 1993b\]. 
6 Conclusion 
Building necessary knowledge bases is a major bottle 
neck of designing a practical parser. Previous studies 
on the problem had proposed many learning methods 
to reduce the difficulty. However, during learning 
most of them still need extra information or human 
intervention, which are the major sources of incon- 
sistencies, errors, and inefficiency in learning. This 
paper is thus dedicated to the fully automatic nat- 
ural language learning in which only raw sentences 
are entered to the system. We study how a parser 
may extend itself by observing its own experiences 
of parsing. The proposed learning technique SEP is 
triggered when the parser fails in parsing raw sen- 
tences. It is shown that hypotheses for missing knowl- 
edge may be generated based on the parser's existing 
knowledge, universal linguistic constraints, and par- 
tial results of the failed parsing. Those hypothetic 
knowledge pieces that are likely to facilitate success- 
ful parsing may then be extracted as the new parsing 
knowledge. Thus the parser may acquire the knowl- 
edge that it actually lacks in parsing. As more parsers 
become available, SEP may be attached to them in 
order to enhance their knowledge fully automatically. 
Acknowledgement This research is supported 
in part by NSC (National Science Council of R.@.C.) 
under the grant NSC83-0408-E-007-008. 
Reference 
Berwick R. C. (1985). The Acquisition of Syntac- 
tic Knowledge~ The MIT Press, Cambridgep Mas- 
sachusetts, London, England. 
Brent M. R. (1993). From Grammar to Lczicon: Un- 
supervised Learning of Lexical Syntaz~ Computational 
Linguistics, Vol. 9, No. 2, pp. 243-262. 
Chomsky N. (1981). Lectures on Government and 
Binding, Forts Publications- Dordrecht. 
Faisal K. A. and Kwasny S. C. (1990). Design of a 
Hybrid Deterministic Parser, Proc. of COLING. 
Gazdar G., Klein E., Pullum G. K., and Sag I. A. 
(1985). Generalized Phrase Structure Grammar, Har- 
vard University Press, Cambridge, MA. 
Jacobs P. and Zernik U. (1988). Acquiring Lexical 
Knowledge from Tezt: A Case Study, Proc. of AAAI. 
Lang F.-M. and Hirsehman L. (1988). Improved 
Portability and Parsing through Interactive Acquisi- 
tion of Semantic Information, Proc. of the second 
conference on Applied NLP, pp. 49-57. 
Liu R.-L. and Soo V.-W. (1993a). Parsing-Driven 
Generalization for Natural Language Acquisition, In- 
ternational Journal of Pattern Recognition and Arti- 
ficial Intelligence, Voi. 7, No. 3. 
Liu R.-L and Soo V.-W (1993b). An Empirical Study 
on Thematic Knowledge Acquisition Based on Syn- 
tactic Clues and Heuristics, Proc. of the ACL-93. 
Liu R.-L. and Soo V.-W. (1992a). Augmenting and 
Efficiently Utilizing Domain Theory in Ezplanation- 
Based Natural Language Acquisition, Proc. of the 9th 
International Machine Learning Conference. 
Liu R.-L and Soo V.-W. (1992b). Acquisition 
of Unbounded Dependency Using Explanation-Based 
Learning, Proc. of ROCLING V. 
McClelland J. L. and Kawamoto A. II. (1986). Mech- 
anisms of Sentence Processing: Assigning Roles to 
Constituents of Sentences, in Parallel Distributed 
Processing, Vol. 2, pp. 272-325. 
Montemagni S. and Vanderwende L. (1992). Struc- 
tural Patterns vs. String Patterns for Extracting 
Semantic Information from Dictionary, Proc. of 
COLING-92, pp. 546-552. 
Pinker S. (1984). Language Learnability and Lan- 
guage Development, The IIarvard University Press, 
Cambridge, Massachusetts, London, England. 
Pustejovsky J., Berger S, and Anick P. (1993). LexicaI 
Semantic Techniques for Corpus Analysis, Computa- 
tional Linguistics, Vol. 9, No. 2, pp. 331-358. 
Sanfilippo A. and Pozanski V. (1992). The Acquisi- 
tion of Lexical Knowledge from Combined Machine- 
Readable Dictionary Sources, Proc. of the Third Con- 
ference on Applied NLP, pp. 80-87. 
Sekine S., Carroll J. J., Ananiadou S., and Tsujii 3". 
(1992). Automatic Learning for Semantic Colloca- 
tion, Proc. of the Third Conference on Applied NLP. 
Siskind J. M. (1990). Acquiring Core Meanings of 
Words, Represented as Jackendoff-style Conceptual 
Structures, from Correlated Streams of Linguistic and 
Non-linguistic Input, Proc. of the ACL-90. 
Smadja F. A. (1991). From N-Grams to Collocations: 
An Evaluation of EXTRACT, Proc. of the 29th an- 
nual meeting of the ACL, pp. 279-284. 
Vclardi P., Pazlenza M. T., and Fasolo M. (1991). 
How to Encode Semantic Knowledge: A Method for 
Meaning Representation and Computer-Aided Acqui- 
sition, Computational Linguistics, Vol. 17, No. 2. 
Webster M. and Marcus M. (1989). Automatic Acqui- 
sition of the Le~ical Semantics of Verbs from Sentence 
IS.ames~ Proc. of the ACL-89, pp. 177-184. 
Wexler K. and Culicovcr P. W. (1980). Formal Prin- 
ciples of Language Acquisition, The MIT Press, Cam- 
bridge, Massachusetts, London, England. 
Zernik U. (1987). Learning ldioms - With and With- 
out Explanation, Proc. of IJCAI, pp. 133-136. 
Zernik U. (1989). Lexicon Acquisition: Learning from 
Corpus by Capitalizing on Lexical Categories, Proc. 
of IJCAI, pp. 1556-1562. 
Zernlk U. and .\]acobs P. (1990). Tagging for Learning: 
Collecting Thematic Relation from Corpus, Proc. of 
COLING, pp. 34-39. 
446 
