Generic NLP Technologies: Language, Knowledge and
Information Extraction
Junichi Tsujii
Department of Information Science, Faculty of Science
University of Tokyo, JAPAN
And
Centre for Computational Linguistics, UMIST, UK
1 Introduction
We have witnessed signi#0Ccant progress in
NLP applications such as information ex-
traction #28IE#29, summarization, machine trans-
lation, cross-lingual information retrieval
#28CLIR#29, etc. The progress will be accelerated
by advances in speech technology, which not
only enables us to interact with systems via
speech but also to store and retrieve texts in-
put via speech.
The progress of NLP applications in this
decade has been mainly accomplished by the
rapid development of corpus-based and sta-
tistical techniques, while rather simple tech-
niques have been used as far as the structural
aspects of  are concerned.
In this paper, we will discuss how we
can combine more sophisticated, linguistically
elaborate techniques with the current statis-
tical techniques and what kinds of improve-
mentwe can expect from suchanintegration
of di#0Berent knowledge types and methods.
2 Argument against linguistically
elaborate techniques
Throughout the 80s, research based on lin-
guistics had #0Dourished even in application ori-
ented NLP research such as machine transla-
tion. Eurotra, a European MT project, had
attracted a large number of theoretical lin-
guists into MT and the linguists developed
clean and linguistically elaborate frameworks
such as CTA-2, Simple Transfer, Eurotra-6,
etc.
ATR, a Japanese research institute for
telephone dialogue translation supported by
a consortium of private companies and the
Ministry of Post and Communication, also
adopted a linguistics-based framework, al-
though they changed their direction in the
later stage of the project. They also adopted
sophisticated plan-based dialogue models as
well at the initial stage of the project.
However, the trend changed rather dras-
tically in the early 90s and most research
groups with practical applications in mind
gave up such strategies and switched to more
corpus-oriented and statistical methods. In-
stead of sentential parsing based on linguis-
tically well founded grammar, for example,
they started to use simpler but more ro-
bust techniques based on #0Cnite-state models.
Neither did knowledge-based techniques like
plan-recognition, etc. survive, which presume
explicit representation of domain knowledge.
One of the major reasons for the failure
of these techniques is that, while these tech-
niques alone cannot solve the whole range of
problems that NLP application encounters,
both linguistsand AI researchers made strong
claims that their techniques would be able to
solve most, if not all, of the problems. Al-
though formalisms based on linguistic theo-
ries can certainly contribute to the develop-
ment of clean and modular frameworks for
NLP, it is rather obvious that linguistics the-
ories alone cannot solve most of NLP's prob-
lems. Most of MT's problems, for example,
are related with semantics or interpretation
of  which linguistictheories of syntax
can hardly o#0Ber solutions for #28Tsujii 1995#29.
However, this does not imply, either, that
frameworks based on linguistic theories are of
no use for MT or NLP application in general.
This only implies that we need techniques
complementary to those based on linguistic
theories and that frameworks based on lin-
guistic theories should be augmented or com-
bined with other techniques. Since techniques
from complementary #0Celds such as statistical
or corpus-based ones have made signi#0Ccant
progresses, it is our contention in this paper
that we should start to think seriously about
combining the fruits of the research results of
the 80s with those of the 90s.
The other claims against linguistics-based
and knowledge-based techniques which have
often been made by practical-minded people
are :
#281#29 E#0Eciency: The techniques such as sen-
tential parsing and knowledge-based in-
ference, etc. are slow and require a large
amount of memory
#282#29 Ambiguity of Parsing: Sentential
parsing tends to generate thousands of
parse results from which systems cannot
choose the correct one.
#283#29 Incompleteness of Knowledge and
Robustness: In practice one cannot
provide systems with complete knowl-
edge. Defects in knowledge often cause
failures in processing, which result in the
fragile behavior of systems.
While these claims may have been the case
during the 80s, the steady progress of such
technologies have largely removed these dif-
#0Cculties. Instead, the disadvantages of cur-
rent technologies based on #0Cnite state tech-
nologies, etc. have increasingly become
clearer; the disadvantages such ad-hocness
and opaqueness of systems which prevent
them from being transferred from an appli-
cation in one domain to another domain.
3 The current state of the JSPS
project
Ina#0Cve-year project funded by JSPS #28Japan
Society of Promotion of Science#29 which
started in September 1996, we have focussed
our research on generic techniques that will
be used for di#0Berent kinds of NLP application
and domains.
The project comprises three university
groups from the University of Tokyo, Tokyo
Institute of Technology #28Prof. Tokunaga#29
and Kyoto University #28Dr. Kurohashi#29, and
coordinated by myself #28at the University of
Tokyo#29. The University of Tokyo has been
engaged in development of software infras-
tructure for e#0Ecient NLP, parsing technology
and ontology building from texts, while the
groups of Tokyo Institute of Technology and
Kyoto University have been responsible for
NLP application to IR and Knowledge-based
NLP techniques, respectively.
Since wehave delivered promisingresults in
research on generic NLP methods, we are now
engaged in developing several application sys-
tems that integrate various research results to
show their feasibility in actual application en-
vironments. One such application is a system
that helps biochemists working in the #0Celd of
genome research.
The system integrates various research re-
sults of our project such as new techniques
for query expansion and intelligent indexing
in IR, etc. The two results to be integrated
into the system that we focus on in this paper
are IE using a full-parser #28sentential parser
based on grammar#29 and ontology building
from texts.
IE is very much in demand in genome re-
search, since quite a large portion of research
is now being targeted to construct systems
that model complete sequences of interac-
tion of various materials in biological organ-
isms. These systems require extraction of rel-
evant information from texts and its integra-
tion in #0Cxed formats. This entails that the
researchers there should have a model of in-
teraction among materials, into which actual
pieces of information extracted from texts are
#0Ctted. Such a model should have a set of
classes of interaction #28event classes#29 and a set
of classes of entities that participate in events.
That is, the ontology of the domainshouldex-
ist. However, sincethe buildingof an ultimate
ontology is, in a sense, the goal of science,
the explicit ontology exists only in a very re-
stricted and partial form. In other words, IE
and Ontology building are inevitably inter-
twined here.
In short, we found that IE and Ontology
building from texts in genome research pro-
vide an ideal test bed for our generic NLP
techniques, namelysoftware infrastructurefor
e#0Ecient NLP, parsing technology, and ontol-
ogy building from texts with initial partial
knowledge of the domain.
4 Software Infrastructure and
Parsing Technology
While tree structures are a versatile scheme
for linguistic representation, invention of fea-
ture structures that allow complex features
and reentrancy #28structure sharing#29 makes
linguistic representation concise and allows
declarative speci#0Ccations of mutual relation-
ships among representation of di#0Berent lin-
guistic levels #28e.g.: morphology, syntax, se-
mantics, discourse, etc.#29. More importantly,
using bundles of features instead of simple
non-terminal symbols to characterize linguis-
tic objects allow us to use much richer statis-
tical means such as ME #28maximum entropy
model#29, etc. instead of simple probabilistic
CFG. However, the potential has hardly been
pursued yet mostly due to the ine#0Eciency and
fragility of parsing based on feature-based for-
malisms.
In order to remove the e#0Eciency obstacle,
we have in the #0Crst two years devoted our-
selves to the developmentof:
#28A#29 Software infrastructure that makes pro-
cessing of feature-based formalisms e#0E-
cient enough both for practical applica-
tion and for combining it with statistical
means.
#28B#29 Grammar #28Japanese and English#29 with
wide coverage for processing real world
texts #28not examples in textbooks of lin-
guistics#29. At the same time, processing
techniques that make a system robust
enough for application.
#28C#29 E#0Ecient parsing algorithm for
linguistics-based frameworks, in particu-
lar HPSG.
We describe the current states of these three
in the following.
#28A#29 Software Infrastructure #28Miyao
2000#29:
We designed and develop a programming sys-
tem, LiLFeS, which is an extension of Pro-
log for expressing typed feature structures in-
stead of #0Crst order terms. The system's core
engine is an abstract machine that can pro-
cess features and execute de#0Cnite clause pro-
gram. While similar attempts treat feature
structure processing separately from that of
de#0Cnite clause programs, the LiLFeS abstract
machine increases processing speed by seam-
lessly processing feature structures and de#0C-
nite clause programs.
Diverse systems, such as large scale English
and Japanese grammar, a statistical disam-
biguation module for the Japanese parser, a
robust parser for English, etc., have already
been developed in the LiLFeS system.
We compared the performance of the sys-
tem with other systems, in particular with
LKB developed by CSLI, Stanford Univer-
sity,by using the same grammar #28LinGo also
provided by Stanford University#29. A parsing
system in the LiLFeS system, which adopts
a naive CKY algorithm without any sophis-
tication, shows similar performance as that
of LKB which uses a more re#0Cned algorithm
to #0Clter out unnecessary uni#0Ccation. The de-
tailed examination reveals that feature uni-
#0Ccation of the LiLFeS system is about four
times faster than LKB.
Furthermore, since LiLFeS has quite a few
built-in functions that facilitate fast sub-
sumption checking, e#0Ecient memory manage-
ment, etc., the performance comparison re-
veals that more advanced parsing algorithms
like the one we developed in #28C#29 can bene#0Ct
from the LiLFeS system. Wehave almost #0Cn-
ished the second version of the LiLFeS system
that uses a more #0Cne-grained instruction set,
directly translatable to naive machine code of
aPentium CPU. The new version shows more
than twice improvement in execution speed,
which means the naive CKY algorithm with-
out any sophistication in the LiLFeS system
will outperform LKB.
#28B#29 Grammar with wide coverage
#28Tateisi 1998; Mitsuishi 1998#29:
While LinGo that we used for comparison is
an interesting grammar from the view point
of linguistics, the coverage of the grammar is
rather restricted. We have cooperated with
the University of Pennsylvania to develop a
grammar with wide coverage. In this co-
operation, we translated an existing wide-
coverage grammar of XTAG to the framework
of HPSG, since our parsing algorithms in #28C#29
all assume that the grammar are HPSG. As
we discuss in the following section, we will
use this translated grammar as the core gram-
mar for information extraction from texts in
genome science.
As for wide-coverage Japanese Gram-
mar, we have developed our own grammar
#28SLUNG#29 . SLUNG exploits the property
of HPSG that allows under-speci#0Ced con-
straints. That is, in order to obtain wide-
coverage from the very beginning of grammar
development, we only give loose constraints
to individual words that may over-generate
wrong interpretations but nonetheless guar-
antee correct ones to be always generated.
Instead of rather rigid and strict con-
straints, we prepare 76 templates for lexical
entries that specify behaviors of words be-
longing to these 76 classes. The approach
is against the spirit of HPSG or lexicalized
grammar that emphasizes constraints speci#0Cc
to individual lexical items. However, our
goal is #0Crst to develop wide-coverage gram-
mar that can be improved by adding lexical-
item speci#0Cc constraints in the later stage
of grammar development. The strategy has
proved to be e#0Bective and the current gram-
mar can produce successful parse results for
98.3 #25 of sentences in the EDR corpus with
high e#0Eciency #280.38 sec per sentence for the
EDR corpus#29. Since the grammar overgen-
erates, wehave to choose single parse results
among a combinatorially large numberofpos-
sible parses. However, an experiment shows
that a statistic method using ME #28we use the
program for ME developed by NYU#29 can se-
lect around 88.6 #25 of correct analysis in terms
of dependency relationships among ! ! bun-
setsu's - the phrases in Japanese#29.
#28C#29 E#0Ecient parsing algorithm
#28Torisawa 2000#29:
While feature structure representation pro-
vides an e#0Bective means of representing lin-
guistic objects and constraints on them,
checking satis#0Cability of constraints by lin-
guistic objects, i.e. uni#0Ccation, is computa-
tionally expensive in terms of time and space.
Oneway of improvingthe e#0Eciency isto avoid
uni#0Ccation operations as much as possible,
while the other way is to provide e#0Ecient soft-
ware infrastructure such as in #28A#29. Once we
choose a speci#0Cc task like parsing, genera-
tion, etc., we can devise e#0Ecient algorithms
for avoiding uni#0Ccation.
LKB accomplishes such reduction by in-
specting dependencies among features, while
the algorithm wechose is to reduce necessary
uni#0Ccation by compiling given HPSG gram-
mar into CFG. The CFG skeleton of given
HPSG, which is semi-automatically extracted
from the original HPSG, is applied to pro-
duce possible candidates of parse trees in the
#0Crst phase. The skeletal parsing based on ex-
tracted CFG #0Clters out the local constituent
structures which do not contribute to any
parse covering the whole sentence. Since a
large proportion of local constituent struc-
tures do not actually contribute to the whole
parse, this #0Crst CFG phase helps the second
phase to avoid most of the globally mean-
ingless uni#0Ccation. The e#0Eciency gain by
this compilationtechnique dependson the na-
ture of the original grammar to be compiled.
While the e#0Eciency gain for SLUNG is just
two times, the gain for XHPSG #28HPSG gram-
mar obtained by translating the XTAG gram-
mar into HPSG#29 is around 47 times for the
ATIS corpus #28Tateisi 1998#29.
5 Information extraction by
sentential parsing
The basic arguments against use of sentential
parsing in practical application suchasIEare
the ine#0Eciency in terms of time and space,
the fragility of systems based on linguistically
rigid frameworks and highly ambiguous parse
results that we often have as results of pars-
ing.
On the other hand, there are arguments
for sentential parsing or the deep analysis
approach. One argument is that an ap-
proach based on linguistically sound frame-
works makes systems transparent and easy to
re-use. The other is the limit on the qual-
ity that is achievable by the pattern match-
ing approach. While a higher recall rate of
IE requires a large amount of patterns to
cover diverse surface realization of the same
information, wehave to widen linguistic con-
texts to improve the precision by preventing
extraction of false information. A pattern-
based system may end up with a set of pat-
terns whose complex mutual nullifythe initial
appeal of simplicityofthe pattern-based ap-
proach.
As we see in the previous section, the e#0E-
ciency problem becomes less problematic by
utilizing the current parsing technology. It
is still a problem when we apply the deep
analysis to texts in the #0Celd of genome sci-
ence, which tend to have much longer sen-
tences than in the ATIS corpus. However, as
in the pattern-based approach, we can reduce
the complexity of problems by combining dif-
ferent techniques.
In a preliminary experiment, we #0Crst use a
shallow parser #28ENGCG#29 to reduce part-of-
speech ambiguities before sentential parsing.
Unlike statistic POS taggers, the constraint
grammar adopted by ENGCG preserves all
possiblePOS interpretations just by dropping
interpretationsthat are impossibleingivenlo-
cal contexts. Therefore, the use of ENGCG
does not a#0Bect the soundness and complete-
ness of the whole system, while it reduces sig-
ni#0Ccantly the local ambiguities that do not
contribute to the whole parse.
The experiment shows that ENGCG pre-
vents 60 #25 of edges produced by a parser
Based on naive CKY algorithm, when it is ap-
plied to 180 sentences randomly chosen from
MEDLINE abstracts #28Yakushiji 2000#29. As a
result, the parsing by XHPSG becomes four
times faster from 20.0 seconds to 4.8 second
per sentence, which is further improved by us-
ing chunking based on the output of a Named
Entity recognition tool to 2.57 second per sen-
tence. Since the experiment was conducted
with a naive parser based on CYK and the
old version of LiLFeS, the performance can
be improved further.
The problems of fragility and ambiguity
still remain. XHPSG fails to produce parses
for about half of the sentences that cover
the whole. However, in application such as
IE, a system needs not have parses covering
the whole sentence. If the part in which the
relevant pieces of information appear can be
parsed, the system can extract them. This is
one of the major reasons why pattern-based
systems can work in a robust manner. The
same idea can be used in IE based on sen-
tential parser. That is, techniques that can
extract information from partial parse results
will make the system robust.
The problem of ambiguity can be treated in
a similar manner. In a pattern-based system,
the system extracts informationwhen parts of
the text match with a pattern, independently
of whether other interpretations that compete
with the interpretation intended by the pat-
tern exist or not. In this way, a pattern-based
system treats ambiguity implicitly. In case
of the approach based on sentential parsing,
we treat the ambiguity problemby preference.
That is, an interpretation that indicates rel-
evant pieces of information exist is preferred
to other interpretations.
Although the methods illustrated in the
above make IE based on sentential pars-
ing similar to the pattern-based approach,
the approach retains the advantages over the
pattern-based one. For example, it can pre-
vent false extraction if the pattern that dic-
tates extraction contradicts with wider lin-
guistic structures or with the more preferred
interpretations. It keeps separate the general
linguistic knowledge embodied in the form of
XHPSG grammar that can be used in any do-
main. The mapping between syntactic struc-
tures to predicate structures can also be sys-
tematic.
6 Information extraction of named
entities using a hidden Markov
model
The named entity tool mentioned above,
called NEHMM #28Collier 2000#29, has been de-
veloped as a generalizable supervised learning
method for identifying and classifying terms
given a training corpus of SGML marked-up
texts. HMMs themselves belong to a class of
learning algorithms that can be considered to
be stochastic #0Cnite state machines. They have
enjoyed success in a wide number of #0Celds in-
cluding speech recognition and part of speech
tagging. We therefore consider their exten-
sion to the named entity task, which is es-
sentially a kind of semantic tagging of words
based on their class, to be quite natural.
NEHMM itself strives to be highly gen-
eralizable to terms in di#0Berent domains and
the initial version uses bigrams based on lex-
ical and character features with one state
per name class. Data-sparseness is over-
come using the character features and linear-
interpolation.
Nobata et al. #28Nobata 1999#29 comment on
the particular di#0Eculties with identifying and
classifying terms in the biochemistry domain
including an open vocabulary and irregular
naming conventions as well as extensive cross-
over invocabulary between classes. The irreg-
ularnamingarises inpart because of the num-
ber of researchers from di#0Berent #0Celds who
are working on the same knowledge discov-
ery area as well as the large number of pro-
teins, DNA etc. that need to be named. De-
spitethe beste#0Borts of majorjournalsto stan-
dardize the terminology, there is also a sig-
ni#0Ccant problem with synonymy so that of-
ten an entity has more than one name that is
widely used such as the protein names AKT
and PKB. Class cross-over of terms is another
problem that arises because many DNA and
RNA are named after the protein with which
they transcribe.
Despite the apparent simplicity of the
knowledge in NEHMM, the model has proven
to be quite powerful in application. In the
genome domain with only 80 training MED-
LINE abstracts it could achieve over 74#25 F-
score #28a common metric for evaluation used in
IE that combines recall and precision#29. Simi-
lar performance has been found when training
using the dry-run and test set for MUC-6 #2860
articles#29 in the news domain.
The next stage in the development of our
model is to train using larger test sets and
to incorporate wider contextual knowledge,
perhaps by marking-up for dependencies of
named-entities in the training corpus. This
extra level of structural knowledge should
help to constrain class assignment and also
to aid in higher levels of IE suchasevent ex-
traction.
7 Knowledge Building and Text
Annotation
Annotated corpora constitute not only an in-
tegral part of a linguistic investigation but
also an essential part of the design methodol-
ogy foranNLP systems. Inparticular, the de-
sign of IE systems requires clear understand-
ing of information formats of the domain, i.e.
what kinds of entities and events are consid-
ered as essential ingredients of information.
However, such information formats are often
implicit in the minds of domain specialists
and the process of annotating texts helps to
reveal them.
It is also the case that the mapping be-
tween information formats and surface lin-
guistic realization is not trivial and that cap-
turing the mapping requires empirical exam-
ination of actual corpora. While generic pro-
grams with learning ability may learn such
a mapping, learning algorithms need training
data, i.e. annotated corpora.
In order to design a NE recognition pro-
gram, for example, wehavetohave a reason-
able amount of annotated texts which show
in what linguistic contexts named entities ap-
pear and what internal structures typical lin-
guisticexpressionsof namedentitiesofa given
#0Celd have. Such human inspection of anno-
tated texts suggests feasible tools for NE #28e.g.
HMM, ME, decisiontrees, dictionarylook-up,
etc.#29 and a set of feasible features, if one uses
programs with learning ability. Human in-
spection of annotated corpora is still an in-
evitable step of feature selection, even if one
uses programs with learning ability.
More importantly, to determine classes of
named entities and events which should re-
#0Dect the views of domain specialists requires
empirical investigation, since these often exist
implicitlyonly inthe mindof specialists. This
is particularly the case in the #0Celd of med-
ical and biological sciences, since they have
a much larger collection of terms #28i.e. class
names#29 than, for example, mathematical sci-
ence, physics, etc.
In order to see the magnitude of the work
and di#0Eculties involved, we chose a well-
circumscribed#0Celd and collected texts #28MED-
LINE abstracts#29 in the #0Celd to be annotated.
The #0Celd is the reaction of transcription fac-
tors in human blood cells. The kinds of infor-
mation that we try to extract are the infor-
mation on protein-protein interactions.
The #0Celd was chosen because a research
group of National Health Research Institute
of the Ministry of Health in Japan is building
a database called CSNDB #28Cell Signal Net-
work DB#29, which gathers this type of infor-
mation. They read papers every week to ex-
tract relevant information and store them in
the database. IE of this #0Celd can reduce the
work that is done manually at present.
We selected abstracts from MEDLINE by
the key words of "human", "transcription fac-
tors" and "blood cells", which yield 3300 ab-
stracts. The abstracts are from 100 to 200
words in length. 500 abstracts were chosen
randomly and annotated. Currently, seman-
tic annotation of 300 abstracts has been #0Cn-
ished and we expect 500 abstracts to be done
by April #28Ohta 2000#29.
The task of annotation can be regarded as
identifying and classifying the terms that ap-
pear in texts according to a pre-de#0Cned clas-
si#0Ccation scheme. The classi#0Ccation scheme,
in turn, re#0Dects the view of the #0Celds that bio-
chemists have. That is, semantic tags we use
are the class names in an ontology of the #0Celd.
Ontologies of biological terminology have
been created in projects such as the EU
funded GALEN project to provide a model
of biological concepts that can be used to
integrate heterogeneous information sources
while some ontologies such as MeSH are built
for the purpose of information retrieval Ac-
cording to their purposes, ontologies di#0Ber
from #0Cne-grained to coarse ones and from as-
sociative to logical ones. Since there is no
appropriate ontology that covers the domain
that we are interested in, we decided to build
one for this speci#0Cc domain.
The design of our ontology is in progress,
in which we distinguish classi#0Ccation based
on roles that proteins play in events from
that based on internal structures of proteins.
The former classi#0Ccation is closely linked with
classi#0Ccation of events. Since classi#0Ccation is
based on feature lattices, we plan to use the
LiLFeS system to de#0Cne these classi#0Ccation
schemes and their relationships among them.
8 Future Directions
While the researches of the 80s and 90s in
NLP focussed on di#0Berent aspects of lan-
guage, they have been so far considered sepa-
rate development and no serious attempt has
been made to integrate them.
In the JSPS project, wehave prepared nec-
essary background for suchintegration. Tech-
nological background such as e#0Ecient parsing,
a programming system based on types, etc.
will contribute to resolving e#0Eciency prob-
lems. The techniques such as NE recogni-
tion, staged architecture in conventional IE,
etc. will give hints on how to incorporate sev-
eral di#0Berent techniques in the whole system.
A reasonable size of semantically annotated
texts, together with relevant ontology, have
been prepared.
We are engaged now in integrating these
components in the whole system, in order to
showhow theoretical work, together with col-
lection of empirical data, can facilitate sys-
tematic development of NLP application sys-
tems.
References
Collier, N., Nobata, C., and Tsujii, J.: "Extract-
ing the Namesof Genesand Geneproducts with
a Hidden Markov Model", COLING'2000 #28Au-
gust#29, 2000
Mitsuishi, Y. et.al.: HPSG-style Underspeci#0Ced
Japanese Grammar with Wide Coverage, in
Proc. of Coling-ACL 98, Montreal, 1998
Miyao, Y., Makino, T., et.al.: The LiLFeS Ab-
stract Machine and its Evaluation with LinGo,
A Special Issue on E#0Ecient Processing of
HPSG, Journal of Natural Language Engineer-
ing, Cambridge University Press, 2000 #28to ap-
pear#29
Nobata, C., Collier, N., and Tsujii, J.: "Au-
tomatic Term Identi#0Ccation and Classi#0Ccation
in Biology Texts", in proceedings of the Nat-
ural Language Paci#0Cc Rim Symposium #28NL-
PRS'99#29, Beijing, China, 1999.
Ohta, T., et.al.: A Semantically Tagged Corpus
based on an Ontology for Molecular Biology,in
Proc. of JSPS Symposium 2000, Tokyo, 2000
Tateisi, Y. et.al.: Translating the XTAG English
Grammar to HPSG, in Proc. of TAG+4 work-
shop, UniversityofPennsylvania, 1998
Torisawa, K. et.al.: An HPSG Parser with CFG
Filtering, A Special Issue on E#0Ecient Process-
ing of HPSG, Journal of Natural  Pro-
cessing, Cambridge University Press, 2000 #28to
appear#29
Tsujii, J.: MT Research:Productivity and Con-
ventionality of Language, RANLP-95,Tzigov
Chark,Bulgaria,14-16 September,1995
Yakushiji, A.: Domain-Independent System for
EventFrame Extractionusing an HPSG Parser,
Bsc Dissertation, Department of Information
Science, UniversityofTokyo, 2000
