Translation by Understanding: A Machine Translation Sy,i;tem LUTE 
Hirosato NOMURA, Shozo NAITO, Yasuhiro KATAGIRI, and Aldra SHIMAZU 
NTT Basic Research Laboratories, Masashino-shi, Tokyo, 180, Japan 
Abstract 
This pal)or presents a linguistic model for language 
understanding and describes its application to an experimental 
machine translation system called LUTE. The language 
understanding model is an interactive model between the memory 
structure and a text. The memory structure is hierarchical and 
represented in a frame-network. Linguistic and non-linguistic 
knowledge is stored and the result of understanding the text is 
assimilated into the memory structure. Tim understanding process is 
interactive in that the text invokes knowledge and the understanding 
procedure intcrprots the text by using that knowledge. A linguistic 
model, called the Extended Case Structure model, is defined by 
adopting three kinds of information: structure, relation and concept. 
These three are used rccursively and iteratively as the basis for 
memory organization. These principles are applied to the design and 
implementation of the LUTE which translates Japanese into English 
and vice versa. 
1. Introduetiml 
Since the early 1970s, a variety of approaches to language 
understanding have been proposed. In particular, the importance of 
knowledge organization has been emphasized, and linguistically 
structured knowledge such as Script \[11 and knowledge 
representation frameworks such as Frame \[2\] and Semantic Network 
\[31 have been proposed. At the santo time, the linguistic approach has 
been adopted to reveal the discourse structure, the cognitive approach 
has attempted to explain phenomena such as focus, topic and 
intention, and the formal semantic approach has been used to 
establish semantics based on tim logical model theory. 
We propose an interactive model between the memory structure 
and the text (or utterance) as a language understanding model. In the 
model, knowledge stored in the memory structure plays the principal 
role such that the text invokes knowledge and the understanding 
system interprets the text using that knowledge. The knowledge 
consists of linguistic knowledge and non-linguistic knowledge. They 
are closely related each other and incorporated into the memory 
structure simultaneously. As a result of understanding, the system 
assimilates tile meaning structure of the text into its memory 
structure. The bases for representing the knowledge are structure, 
relation and concept which are the fundamental cmnponents for 
constructing and representing the memory structure including the 
meaning structure era sentence. For the purpose oI'clear definition of 
linguistic information, a linguistic model, called the Extended Case 
Structure model (ECS), which is capable of treating the structures of 
complex sentences, is provided. 
These principles have been applied to the design of a new version 
of the experimental machine translation system called LUTE 
(Language U nderstander, Translator and Editor) \[4\]. This paper 
deals mainly with the current Japanese-English version of LUTE 
(LUTE-JE versioml) \[5\]. LUTE has following processing 
characteristics: 1) Not only syntactic but also semantic relations 
(dependencies) between modifiers and modificants are analyzed 
simultaneously. 2) All kinds of information such as syntactic 
patterns, meaning structures, lexical items, and knowledge are 
represented in a uniform framework, called a Frame-Network. 3) 
Analysis produces a ~most plausible meaning structm'e' based on the 
prediction of syntactic structures and the integration of semantic 
structures. 4) Transfer is realized as a general fl'amework for 
manipulationg the frame network. 
2. Language Understanding Model 
2.1 Memory Organization 
Knowledge can be organized into various memory structures 
depending on the type of knowledge. TheMe structures are usually 
hierarchical and consist of three layers; 1) long-term memory, 2) 
discourse memory, and 3) episodic memory (or short-term memory). 
Long-term memory stores knowledge such as dictionaries, grammars, 
experiences, cmnmon-sense knowledge, expert knowledge, and 
procedural knowledge such as how to infm" a fact from a collection of 
facts. Knowledge also contains recta-level knowledge such as 
knowledge about the characteristics of knowledge and the nsage of it. 
Discourse memory stores knowledge concerning the situation as an 
environment of utterances, and the history of understandings such as 
"Who is the author?", "What is the topic?", and "What is the purpose 
of the discourse segment?", Episodic memory steres the meaning 
structure of the ongoing segment of the text and its construction is the 
main issue in the discussion of sentence analysis. 
The memory model described above can be applied to account for a 
number of linguistic phenomena. For example, the difference 
between two Japanese anaphoric expressions "sono ( 7¢ ~),, and "ano 
( ~ ¢))" (both expressions correspond to the determiner "the" in most 
English contexts) is explained by using the memory structure model 
as follows: a referent of the noun modified by "sono" is found in tile 
discourse memory, and a referent of the noun modified by "ano" is 
found in the long-term memory. 
2.2 Basis for Memory Organization 
The memory is constructed by assembling three kinds of basic 
elements; 1) structure, 2) relation and 3) concept. Structure is a 
packet of memory organization. A variety of structures can be used to 
represent linguistic knowledge, the situation of utterance, and the 
meaning structure of a sentence. Concept is associated with 
structures which include all kinds of constituent structures; words, 
phrases, sentences, etc. Hence, this definition of concept, in a sense, is 
language-dependent. There are two kinds of concepts, semantic 
categories and word meanings. Thus a word and its meaning are 
strictly distinguished. Relation integrates structures to form a 
compomld structure. Examples of compound structures are compound 
norms, ease structures, and complex sentences. There are several 
kinds of relations such as casc relations, conjunctive relations and 
taxonomic relations between semantic categories. 
2.3 Understanding Process 
In the understanding process, operations such as matching, 
searching, deletion, replacement, integration, and generation are 
executed in the memory structures. For example, in a morphological 
analysis process, using their literal expressions as search keys, the 
search for words to be identified is made using lexical entries in the 
dictionary, and in the case analysis process, a search is made for case 
instances that match prototype cases in ease frames. 
As understanding proceeds, the essence of episodic memory is 
assimilated into discourse memory and the essence of discourse 
memory is assimilated into long-term memory. Discourse memory 
(long-term memory) is not simply an accumulation of the contents of 
episodic memory (discourse memory), but is a structured memory 
coherently organized from the episodic memory (discourse memory). 
As a prelimlnary model of discourse memory, we define a Local Scene 
Frame (LSF), which is a collection of cases and predicates in preceding 
sentences already analysed. LSF is partly viewed as describing a 
6271 
q Contextual structure ~nted case structure ntodel ~ F 
L ..... E11! PS i-s ...... / '.. _ _ _ Pr e2 up_p2 sitlon _ _ _ / 
q Complex Selltenco structure (Extended case structure model ) F 
r ............................... . 
t ........ _ _ _ c2~i2~2~t;2 ~_ ............ d 
-~ Unit sent ...... tructure F 
L c°_'~p-°u'~'Ln°<-'"- J \[ Y ,_,,b_o a_a~ a ~._o,_~t_~,_~_ \] ..... 
Simple sentence structure 
i L ~._. __t,_No_<kqa_~__ J t 
..... ...... ,,.uot .... ¼ 
i 
Fig. 1 Extended Case Structure Model 
situation in which the utterance is carried out. Information in the 
LSF is used for filling in missing cases, and resolving anaphora. A 
discussion of the detailed procedure for the assimilation would be 
beyond the scope of this paper. 
3. Extended Case Structure Model 
3.1 General Framework 
The Extended Case Structure Model (ECS) is a linguistic model 
for representing the meaning structures of the text. Thus the ECS 
presents a representation scheme for the episodic memory. Figure 1 
shows its fundamental construction. The traditional case structure 
(Fillmorean type) is a structure for a unit sentence which consists 
mainly of relations between nouns and a verb. This is not sufficient to 
represent structures of real sentences which sometimes have complex 
noun phrases, compound sentences, etc. Also, the ECS has to have 
facilities for representing other structures involving relations 
between a noun and a noun, a verb and a verb, etc. The ECS has been 
designed to integrate those structures into one linguistic model. Its 
nature is hierarchical as to the compoundness of constituents, 
iterative as to conjunction, and recursive as to embedding. Using 
these formalisms, the syntactic and semantic structures of sentences 
can be represented uniformly and correctly. 
3.2 Semantic Structure in ECS 
There are two types of semantic structurcs, composite and primi- 
tive structures. A cmnposite structure is made by integrating seman- 
tic structures using semantic relations. A primitive structure, by 
definition, cannot be divided into further substructures. In general, a 
single word corresponds to a primitive structure, and a phrase 
corresponds to a composite structure. Since syntactic information can 
also contribute to define meaning structures, each semantic structure 
simultaneously incorporates not only meaning information but also 
syntactic information. 
We do not assume a language-independent universal semantic 
representation. Thus, it is necessary to define a proper ECS for each 
language: Japanese ECS (J-ECS) \[6\] for Japanese language and 
English ECS (E-ECS) \[7\] for English language. In the translation 
process from Japanese into English, the analysis procedure generates 
a J-ECS for a Japanese sentence, and the transfer procedure 
generates an E-ECS corresponding to the J-ECS. 
3.3 Semantic Relation in ECS 
Senmntic relation connects semantic structures and builds a 
larger semantic structure, ranging from a word structure to a 
sentence structure. Figure 2 shows types of semantic relations, and 
each of them can be explained briefly as fellows: 
1) Noun relation: Relationship between nouns; Examples are 
whole-part, upper-lower, possession, material, etc. 
2) Case relation: Relationship between a case element and a 
predicate; Examples are object, agent, instrument, place, etc. 
622 
3) Embedded relation: Relationship between an embedded sentence 
and a noun phrase, which can be categorized into three types; a) 
case re'lation between a modified noun phrase and the predicate iu 
a modifier embedded sentence, b) noun relation between a 
modified noun phrase and a noun phrase in a modifier embedded 
sentence, and c) an appositive or subsidiary relation between a 
modified noun phrase and a modifier embedded sentence. 
4) Conjunctive relation: Relationship between sentences; Examples 
are cause-result, time-advance, assumption, etc. 
3.4 Concept in ECS 
Concepts are associated with structures mentioned above. Among 
them, concepts associated with word structures represent word 
meanings which appear when the words are used in a sentence. A 
word meaning is represented by principal concepts, supplenmntary 
concepts, and their semantic dependencies. Principal and 
supplementary concepts are dcfined by using semantic categories, and 
prepared for nouns, adverbs, verbs, adjective-verbs and modalities as 
shown in Figure 3. Semantic dependencies are defined by using 
semantic relation fi'ames and semantic structure frames. Semantic 
categories, semantic relation fi'amcs and semantic structure frames 
have the following characteristics: 1) There are two types of concepts: 
prototype and instance. Prototypes play a part of selectional 
constraint to define semantic dependency structures. Instances show 
an assimilated structure which satisfies the selectienal constraints. 
2) They shows semantic commonness and analogy between two 
structures. This allows the system to share information and to 
provide facilities for paraphrase. 3) Semantic categories make tip a 
hierarchical structure. This provides the system with inheritance 
ability and information sharing. 
4. Dictionaries, Knowledge and Their Representation 
4.1 Dictionary 
There are two typos of dictionaries in LUTE. Mono-liugual 
dictionaries are used in analysis and generation, while bMingual 
dictionaries are used in transfer. Mono-lingual dictionaries have the 
following information about words and concepts: 1) tIow the word is 
expressed, 2) how the word is used in the syntax of a sentence, and 3) 
what concept the word corresponds to. Bi-lingual dictionaries has 
information on the correspondence of concepts in two different 
languages, and will be explained in section 6. (Note that concepts are 
defined here by associating structures which are generally language 
dependent.) Figure 4 shows the contents of a word dictionary. 
A word meaning can be regarded as an entry to the conceptual 
knowledge description. The LUTE dictionaries contain the following 
semantic information: 
1) Semantic category (for word meanings): Principal concepts 
associated with the word meaning. Those for nouns and adverbs 
are used as selectional constraints in semantic relation analysis. 
Those for predicate s are used to analyse modality. 
2) Case fi'ame (for predicate word meanings): Constraints and case 
relations which are applied to construct unit sentence semantic 
structures. There are three types of ease frames: intrinsic for 
each predicate word nmaning, common for several predicate word 
meanings, optional for outer ease relations. 
3) Noun relation frame (for noun word meanings): Constraints and 
semantic relationships which ace applied to construct semantic 
structures made up of two nouns. Case frames are also used as a 
ldnd of object relation frames for predicate-type nouns. 
4) Event relation frame (for predicate word meanings): Constraints 
and semantic relationships to be applied to construct complex 
sentence semantic structures. An example is the relation between 
the verb in a main clause and the verb in a subordinate clause. 
5) Heuristics (for semantic categories and relation fi'ames): This is 
used for resolving ambiguity of semantic categories, semantic 
relations, and semantic structures by linguistic information such 
N~oun-ielalot~ .... tri~ci;on ca~se~rel--ati~on ;gent:a-ctio~n~objec~t--actio~i-nst--ru--n,e,;t~;ctio~n- ' -Catego-ries for-~ouns-a~ld adverbs:: =,-~a-ture~m-a-teriT;i lnstri,,,~;~,i i .... 
time-action location-action destination-action source-aetim~ co-object-action manner- I society organization \[ cultllre i buman l action l state \[ number l degree l enmtion l 
action freque cy-action object-state action-location action-time action-result action- time l location| a bstract l concrete animate J plant I others 
degree I state'object property-object possessor-object number-object material-object Categories for verbs:: = voice active statlve movemental I transitional\] 
Iota, on-object object.property object-element object-number object-location species- emotional I thinking perceptual existential judgenlental non-willing 
object l re ative-location Iocat on-specificatio ~ time-specification human-relation noun- v oi c e : : = passive l affected-passive possible l sponraneous causa tive \] pet fective 
I cative:: =momentallcontinual suffix prefix-noun \[ parallel l others 
Case relation:: = OBJECT-TYPE METHOD-TYPE DfRECTION-TYPEJ TIME-SPACE-TYPE 1 stative:: = ~ teiru\[ -teiru 
Categories for adlectives and adjective-verbs:: = SUPPLEMENT-TYPE I MODIFICATION-TYPE 
0 gJ E CT-TY PE :: = object I co-object \[ statem ent-object J compared-object I seconds ry-obje ct I 
theme J agent \[ experiencer 
METHOD-TYPE:: = method I instrument I material\[ element I ca use 
DIRECTION-TYPE:: = source I destination I purpose I result \] giver I recipient 
TIME-SPACE-TYPE:: = location J time 
SUPPLEMENT-TYPE::=Ocasion content role contrast region 
MODIFICATION-TYPE:: = manner \[ frequency degree tbing rate I number emphasis I true tf 
Embedded relation:: = case relation\] relation that modified noun phrase modify case 
instance in tJm enlbedded sentencel apposition.Event-result 
Con unctive relation:: = condition right-affirmative cause lpurposelright assumption\[ 
contrary-assumption contrary-affirmativeljuxtaposition introducr on pallare t me- 
relation before simultaneous after continuation limitecl-continuationlduring I 
examplilTcation selection interrogative-contentslcitationlexplanationlspecificlnlinimal. 
limit proportion degree I limit 
stative I charactedstk \] relational I emotional 
Categories for modalities:: = 
as pect:: = beghlning just-before-beginning Ijust-after.beginning I continuous I 
repetitive I perfective \]just-before.perfective Ijust-after-perfective I perfecrive- 
state I others 
tense:: =past present future 
modal:: = negation possibility necessarity obligation I necessity I inevitability I 
favorability \[ sufficiency I guess I affirmative \] confidential-guess I uncertahl- 
a ffirnlative estimation guess uncertain-guess l hearsay l intention willingness I 
plan hope try causative secood-pelson command interrogative request I 
pernlission invitation third-person l causa tive \] 
request I passive I spon tanity \[ bene factive \[polite I respect I o ,hers 
manner:: =limited degree \[ extr eme-e xample l stress l exan\]plification \] 
parallel ladditionlselection uncertainty distinction others 
Fig. 2 Semantic Relations Fig. 3 Semantic Categories 
as preference over several semantic relations, semantic relation 
fillers, and remaining semantic relation fi'ames not yet filled. 
4.2 Knowledge 
Both comnmn.-sense knowledge and expertdmowledge are 
constructed using basic elements such as concepts, ,'elations and 
structures as well as linguistic structures. Thus the non-lingulstlc 
knowledge manipulated in LUTI,; is not represented in a simple data- 
base fl'amework but rather incorporated into the memory structure. 
Although re.troy language processing systems use the term 
"knowledge" rather vaguely, LU\[\['Itl gives a concrete form to 
knowledge in the sense of franmmetworks corresponding to word 
meanings. The current version of I,UTE defines the following types of 
knowledge in terms of semantic relations: 
1) Concept Relation: Relations such as hyponymy, synonymy, 
antonymy, whole-part, and possession. One example is "whole- 
part" rehttion between "densha ( ;~ ii'-') (train)" and "made (~) 
(window)". (A window can be a part era train.) 
2) Event State Relation: Relations between two events or between 
an event and a state. One example is "subsidiary situation" rela- 
tion that "nioi (~ ~ ) (smell)" results from "yaku ('l')'~ < ) (grill)". 
3) Mete knowledge: This is used for reasoning, such as in traversing 
the concept networks, and checMng semantic consistency 
according to concept networks. 
4.3 Frame-Network 
All information manipulated in LUTE is represented in a uniform 
frameworlt, called a Frame-Network. F, ach type of {lictiouary infer- 
marion such as semantic category, case frame, noun relation frame, 
Entry Information (Kanji-Code, Root-Form) I 
tactic I~forma~ion (Part-of-~peeeh. Co~jugat~omtype) \] 
onin  _3 • • • 
• . . 
--I Obi~ct a~,at,o,,~ .... 1~ ... I Object Relation F .... S \] 
--1 E~entaelationFramel\] "°° \[Eventaelatio~F .... t\] 
Fig'. 4 Contents of dictionary (nlono-lingual) 
and event relation fi'ame is represented by frames with correspondlng 
frame names. These fi'amcs consist of subframes representing 
semantic relation information. Slots of a frame which represents 
semantic relation infermation contain information such as semantic 
category and cast particles stipulating the semantic relation. An 
idiomatic expression between a noun and a verb is represented by a 
co-relation fi'ame. This is the convention for sharing case slots in case 
Dames to yield an effective processing for case analysis and selection 
of word meaning. These frames are also provided for each noun. 
IIcuristies are defined as methods (daemons) in fl'ames. The concept 
relation of knowledge can be represented by inheritance and semantic 
relation slots of noun relation fl'ames. Event state relation is 
represented by event-object relation fl'ames, and expressed in a word 
meaning of the eorresponding noun. Using this relation fFame, 
semantic relations in a phrase, "Sakana we yaku nioi (,(rE {" ;t}'~ { ~ ~,) 
(Smell of fish grilling)" can be analysed. Me,a-knowledge is repre- 
sented as a procedure for unifying frames to select a word meaning, 
inheritance mechanism, and methods in frames as well as heuristics. 
5. Extended Case Analysis 
l,',xtended Case Analysis (F, CA) builds the meaning structure of a 
sentence which is expressed by tim fi'amework based on ECS. The 
ECA integrates both syntactic and semantic analysis using Structure 
Patterns. Analysis proceeds in ,~t manner ill which, top down structure 
prediction and bottom-up structure integration are intertwined. 
Viewing the analysis from the standpoint of the activation of 
lcnowledge, an expression activates a word, a word activates a word 
meaning, a word meaning activates concepts, and coneel)ts activate 
concept relations. We will describe the prccedure for analyzing 
Japanese sentences in the following sections. 
5.l Flow and Control in ECA 
It is assumed here that the morphological analysis process has 
already segmented a sentence into a sequence of words. The ECA 
procedure can be explained roughly as follows. First, the ECA 
predicts a sentence structure in a top-down manner using Structure 
Patterns. Second, it analyzes semantic structures for the predicted 
sentence structure using Semantic Structure Frames, which describe 
constraints for integrating the substructures. Finally, those 
substructures are integrated into a bigger structure. These 
procedures are performed recursively for each level of constituent 
construction until an integrated meaning structure is obtained for the 
entire sentence. When information concerning semantic structure 
frames or knowledge is missing, the ECA does not attempt to nmke a 
unique integrated meaning structure. Rather it utilizes a variety of 
heuristics, thus making it possible to order multiple possible meaning 
structures in terms of likelihood or plausibility based on a score given 
to each meaning strueture. A rough outline of this analysis is 
presented in Figure 5. 
623 
syntactic prediction meaning integration 
\[ unit structure \]~ ................... \[ unit structure~ 
\[ phrase_. __~structure ~- ........... ~~phrase structure 
\[ ~ord structure l---~ word ,tru~t~,~l 
Fig. g Rough sketch of analysis flow 
tIistorieal information, including both the success and failure of 
the processing, is stored so that the ECA can avoid analyzing the 
same sequence of substructures in the backtracking process. 
5.2 Structure Pattern 
A structure pattern is a package of knowledge for predleting 
syntactic constructions between pairs of modifiers and modifieants 
among the constituent structures of a sentence. Based on this 
prediction, an analysis procedure is invoked to analyze their 
semantic structures. If this analysis succeeds, the 
modifier/modificant pair is integrated into a new unified structure. 
Structure patterns are assigned to each structure type in the ECS. An 
example of structure patterns for a unit sentence is shown in Figure 6. 
A structure pattern eenslsts of three parts: 1) the condition for 
applying the pattern, 2) the procedure for semantic structure 
analysis, and 3) newly integrated structure type. The first part 
describes whether this structure pattern can be applied to the 
structure sequence. The second part performs a semantic relation 
analysis of the structure sequence which satisfy the above condition. 
The third part describes the structure type to be produced by the 
above procedure. A structure pattern might be viewed as a context 
fi'ee gramnmr (CFG) rule augmented with a semantic relation 
analysis. In this case, the condition part corresponds to the right hand 
side of the CFG rule, the integrated structure type part corresponds to 
the left hand side of it, and the procedure part can be seen as a 
procedure to derive the left hand side from the right hand side. 
5.3 Semantic Strueture Analysis 
For each constituent construction predicted, the semantic relation 
between modifier and modifieant in the construction is analyzed using 
semantic relation frames. Depending on the differences in structure 
types of the modifierhnodifieant pair, different types of semantic 
relations can be analyzed. In addition, the word meanings of the word 
structure and the categories for the integrated structure can also be 
analyzed. 
Semantic relation analysis can be explained by the analogy of a 
key and key-hole. A modifieant has a number of possible key-holes, 
and a modifier can be regarded as the key which can match it. The 
procedure is to search for the best matching key hole for the key. The 
shapes of keys and key-holes are determined by syntactic (case 
particles) and semantic (semantic category) information. 
The score given to the integrated structure represents the degree 
of syntactic and semantic mismatch recognized in the integration 
process. It is represented by a two-dimensional vectm', whose first 
argument is for syntantic mismatch, and second is for semantic 
mismatch. At each stage of analysis, if syntactic constraint is not 
pattern-name: usent-pattern-1 variables: (case-instance case-particle sequence usent) 
structure: structure-class= usent substructures: 
substructure: substructure-label1 = sub1 structure-class = case 
patterns = (.case-instance (restrict >case-particle case-particlep)) 
substructure: substructure-label2 = sub2 structure-class = usent 
patterns = (.sequence (restrict >usent usentp)) semantic-analysis-function: (case-analysis subl sub2) 
Fig. 6 Example of Structure Pattern (Unit Sentence) The argument with the prefix symbol * can match any nanlber of elements, and the 
argument with the prefix symbol > can match a single element. 
624 
satisfied, two points are added to the syntactic mismatch score, and if 
it is satisfied by modal particles, one 1-mint is added to it. As for 
semantic eon.~traints, if they are not satisfied, two points are added to 
the semantic mismatch score, and if they are satisfied through 
inheritance of semantic categories, one point is added to it. 
5.4 Case Analysis 
Case analysis is the process of matching a ease instance and 
prototype eases in the case fl'ame and of selecting the best matched 
prototype case. Then, the value of the case relation between the case 
instance and the predicate is determined to be the case relation of the 
selected prototype case. 
A modifier element may have co-case slots. It is true that some 
modifiers are strongly associated with partlcular word meanings of 
predicate words. I"or example, a verb "hiku ( iJI < )" has multiple 
meanings, and its exact meaning in a sentence is determined when it 
occurs simultaneously with object cases such as "kaze we hiku (~J{ N~ 
~l < ) (catch a cold)", "jisho we hiku ( ~}~ {q~" ~" ~J\[ < ) (consult a 
dictionary)" ancl "denwa we hiku ( 7E ;,~, ~ ~ I < ) (establish a telephone 
service)". When a modifier element definitely determines the word 
meaning of a modifieant element, it is not efficient to test all possible 
word meanings of the modificant. Therefore, if the same case slot is 
shared by both a modifier and a modificant element, the meaning 
which shares this same case slot is selected as the word meaning of 
both elements without analysing another possibilities. 
5.5 Modality analysis \[8\] 
The classification of modality information and the procedure for 
analysing thmn have presented in the reference thus we will describe 
here only the outline. Modality analysis consists of the following 
three modules combined with case analysis and conjunctive analysis: 
(l) Pro-ease-analysis: A modality which causes a change in the case 
structure is analyzed at this stage. The case frame to be assigned to 
the predicate is modified by utilizing the result of this analysis before 
starting the ease analysis. As for semantically ambiguous auxiliary 
verbs which are also related to the modification of the case structure, 
their role is only predicted at this stage, and after case analysis, the 
likelihood of the prediction is evaluated. 
(2) Post-ease-analysis: A medaiity whose analysis requires case 
structure information is analysed at this stage as follows : 
a) If the category of the modality expression is unique, this category 
is assigned to the nmaning structure. 
b) If a daemon (a procedure to resolve ambiguities using heuristics) 
is attached to the modality expression, it performs the following 
three tasks: i) disambiguating the category of the nmdality word, 
ii) determining the operational scope of the modality, iii) adding 
the implicative meaning caused by the modality word. 
(3) Post-conjunetive-analysls: Following the conjunctive analysis 
between the subordinate clause and the main clause, this module is 
activated to determine whether the medality in the main clause also 
operates on the subordinate clause. For negation in the main clause, 
the transfer of negation is considered. 'resting whether or not the 
modifier event is subsidiary to the oceurenee of the main event is 
accomplished using the semantic relation frames assigned to the 
predicate of the main clause. 
5.6 Determination of Word Meaning 
Word meaning is an entry fl'mn a word to the conceptual network 
consisting of dictionary information and knowledge. Since a word has 
multiple word meanings, it is possible that the word might have 
multiple entries. The information available for the determination of 
word meaning is the accumulated situation (discourse information) 
and the accumulated word meanings (accumulated concepts). If no 
such information is available, a default value is borrowed as the most 
likely word meaning. In the early stage of semantic relation analysis, 
tentative word meanings are assumed. These word meanings may not 
be accurate because they have heen determined solely by the local 
analysis. It is possihle that some of the rejected meanings at this 
stage might be more adequete as the exact word meanings for a given 
word in the context of the entire sentence. Therefore, the system must 
retain all possible word meanings as candidates so that it can change 
the meanings after obtaining enough information to determine the 
exact meaning. 
5.7 Determination of Category 
At the st;tge of building a meaning structure for a sentence, 
categories for each constituent structure are also deterlnined. 
Categories for a structure are usually the same as the categories of 
the head constituent. But if a structure is exoeentrie, categories for 
the structure can be obtained by some operation on its constituent 
substructures. For example, tile category for "omocha no heitai ( }S g 
/5 ~ ~') 3~ t~) (a toy soldier)" is non-animate, although the category of 
"heitai (>fg IN) (a soldier)" is hmnan (therefore, animate). 
In order to determine the categories of asmnantically anabigvtous 
structure or a exoeentrie structure, an attached procedure is invoked. 
For example, the Japanese noun "tame ( ?d &)" is ambiguous because 
it has two categories, purpose and cause. To resolve this ambiguity, a 
daemon is invoked after the noun phrase containing "tame" is 
analyzed. 'Phi,.; daemon performs tile following heuristics: 
1) If "tame" is followed lay both a ease particle "ni ( l.= )" and a modal 
particle "ha ( 12)" (that is, in ease of"tameuiha ( & a5 l= I~)" form), 
the category is determined to be "purpose". 
2) If "tame'is succeeded lay an embedded sentence and the predicate 
shows a perfective aspect (that is, the end part of tile embedded 
sentence contains the auxiliary verb "ta ( t:)" or "teiru ( -C ~' 7o )"), 
or the semantic category of the predicate is "state", the category is 
determined to be "cause". 
3) Otherwise, "purpose". 
6. Transfer 
6.1 Transfer Functions 
Discrepancies among ECS's for different languages arise for 
several reasons. One is essential in nature. We believe that syntactic 
information should be preserved as far as possible in FCS. But 
semantically equivalent information is often realized differently in 
the syntax of different languages. Conceputual systems are also 
different in different language communities. These differences must 
be reflcctcd in BCS's. 
Transfer process should fill these gaps between the ECS's of two 
different languages. At the transfer stage from Japanese to English, 
structures, relations and concepts in J-ECS arc transferred into those 
in g-ECS. Since concepts and relations are integrated into structures, 
the transfer of concepts and relations is performed at the same time as 
the transfer of structures. 
6.2 Transfer of elements of ECS 
In the course of the transfer processes, ECS's in the source 
language are converted by reeursively traversing original structures 
from top nodes, and creating corresponding target structures. So, the 
transfer process consists of transfering components of the ECS's, i.e., 
concepts that make up the ECS and relations which hold among them. 
13ut there are cases which don't suit this scheme well, and hence 
require special treatment. They are idiosyncratic to \[exical items and 
specific procedures have to be triggered by certain concepts included 
in the original structm'es. Idiosyncratic transformations include: 
1) delctlon: certain structures in the source structures are deleted 
and no counterpart structures are embodied in target structures; 
for example, eomt~ound structures are transferred into primitive 
struetures, as in the transfer from "Sakana we tsuru ( ,((t ~" $"; .,.o )" 
in Japanese to "fish" in l'haglish, 
2) addition: certain structures Ihat have no counterpart in the source 
structures are added to target structures; for example, primitive 
structures In'o transferred into compound structures, as in tim 
transfer from "Samidare ( .It J\] H:i)" in Japanese into "early 
summer rain" in English, and 
3) modifieation: .~;ource structures are non-trivially changed in the 
process of transfer, as in the transfer 5'om Japanese adjective 
sentence "Jisuu ga eel ( :-~: ~ \]/~g v, )" into Plnglish "There are ..." 
sentence structure, or types of input and output are different, as in 
the transfer from Japanese phrase "... suru toM (... -4" Za II~'i,)" 
("time" case instance) into the English subordinate clause 
construetion "Whel~ ...". 
The transfer ef concepts consists of 1) transfer of semantic 
categories, and 2) transfer of word meanings. A transfer dictionary 
for a pair of languages is prepared to give information on the 
eorrespondence between concepts in hoth languages. An entry of the 
dictionary consists era triad or fi'alnes, that is, a source concept fi'ame, 
a target concept flame, and a mediating frame. Since concept 
correspondence is in general not one-to-one, there may be several 
target concepts given one source concept and vice versa. Mediating 
fi'ames can provide infm'mation on conditions to make it possible to 
choose auaong alternatives. Concepts that would trigger idiosyncratic 
procedm:es lmve the information in the dictionary in the form of 
transfer rules. 
Transfer of relations consists of transfer of four types of relations 
described in 3.3. Correspondence information is also placed in the 
transfer dictionary. But inforlnation on case relation transfer are 
stored as verbal concepts, since they might be specific to individual 
verbs or classes of verbs. 
6.3 Transfer process 
The transfer process is essentially a manipulation of fl'ame 
networks. A rule-based system was devised to facilitate easy 
specification of the complex pattern of network manipulations. Au 
example of the transfer rule is shown in Figure 7. Similar to structure 
patterns, a transfer rule consists of three parts: a matching part, 
execution part, and a return part. The matching part specifies the 
conditions under which the rule should be invoked. It also contains 
variables, which are bound during matching process and the 
information will be passed to and used in the later stage when the 
matching is successful. The execution part specifies the transfer of 
substructures and concepts, value assignment to the variables, 
fnrther conditional branching, and other operations. Lisp code can he 
invoked in this part. The return part specifies the target structure 
that has to be construeted and returned on the basis of tile application 
of the entire rule. 
(defrule J:USENT (if (self = (J:USENT(*verb(varj-verb)) *meaning(varj-mns))(*m da ity(varj m d))(*cases restj cases) )) 
(i-verb~ (J:WOR,D (*stem-pos (optional (vat j-stem-type))) )) then lexec uocal r-toO ~rest e-modif)) 
{LISP (cond (l-stem-type (setq r-fun #'(lambda (t-frm) (isa* t-ffm 'T:noun-verb})) (send* j-mns 'put: '$restriction r-fun)))) (j-mns-> e-mns) (j-mod -> e-mod) (((for-all}j-cases) -> e-cases) 
(LISP (and j-stem-tyRo (send* j-runs 'remove: $restriction r-fun)) (setq e-modi (mapcan #'(Eambda (q) (and (isa* q 'E:Modifier-Clause) (neons q))) e-cases)) 
\[setq e-cases (subtract e-cases e-modif)) ) (if(LISPe-modif) then (exec (return (! E:CSENT e-csent)) 
where (e-csent = (E:CSENT(*main (! E:Predicatee-pred)) (*mod\]fier-clausee-modif))) (e-prod1 = (E:Predlcate( meaning e-runs) (modahtye-mod) ( cases e-cases)))) 
else (exec (return (\[ E:Predicate e-pred2)) , , • where (e-pred2 = (E:Predicate ( meaning e-mns) ( modality e-mod) ( cases e-cases)))))))) 
Fig. 7 Example of transfer rule (Unit sentence) 
625 
The frame system presented here has a elass-lnstanee hierarchy, 
which adopts an "object-oriented" style of implementation for the 
frame network manipulation in the transfer process. Transfer rules 
specifying how the network should he handled are written for each 
type of structm'es. These are converted into executable forms, and 
attached to class frames of the structure as methods. When the top 
node of the input ECS is given a "transfer" message, corresponding 
methods in the class frame, to the instances of which the top node 
belongs, will be invoked and handle the network as is specified in the 
original rules. 
7. LUTE Experiments 
The LUTE is an experimental machine translation system 
between Japanese and English developed by applying ~he 
investigations mentioned above. The dictionary of each language has 
about 3000 entries. It has been implemented on a Symbellcs Lisp- 
machine by using ZetaI,isp. The size of the system is 850KB of 
programs and 4MB of dictionaries and knowledge, 
LUTE was not deveIoped for practical use but to provide a part of 
the computer environment, RESOLUTE (Reclprocal Envlronment for 
the Study o_f Language .Understander, ~J_'ranslator & Editor), on which 
theoretical works concerning computational linguistics can be 
examined. As a result, IZ/"~SOLUTE contains many facilities for man- 
machine interaction via a multi window screen and consists mainly of 
a frame editor and facilities for conducting program executing. In 
this environment, it is possible to pertbrm translation experiments 
such as analyzing texts, transfering the meaning structures, 
generating phrases and sentences, developlng dictionaries, editing 
knowledge base and examining programs both separately and 
simultaneously. For example, I,UTE can regenerate a sentence of the 
GBJ g &g 
lii;~ ° ti~l!~'lti~l!., i<) i b~. 
lilffili~i~ T qJgl@ ~ 5ILTZ I) ~ b ~. 
~ l~I/l #ll~ ~i T @III@ i 515tl l) ~ b ~. 
source language, while showing the deleted parts in the source 
sentence, from a meaning structure produced by the analysis of a 
source sentence. Also, any intermediate representation can be 
modified to examine the transfer and generation as a whole or a part. 
Since all of the data are represented in a frame network, this 
environment system provides a general fi'amework for frame- 
manipulation facilities. A snapshot of the translation experiment is 
shown in Figure 8. 
The child acquired the+ cunc~pt of" ~aSSo 
The child can solve the problem. 
The child that acquired the cmmept uf mass can solve the problem. 
The child that acquired the concept of mass c\[In solve the_ problem, 
T 2"" '°X Y' "I 
J2USENT-99 ----- 
*INSTANCE-OF SURLUE J~USEMT XOLRSS* 
*VERB ~qRLUE JRNORD-583 -IMSTRNCE-139 \] 
*CASE-CONVERSION ~VRLUE REMTAI 
JRORSE-361 1CASES SVRLUE JRSRSE-332 
*MODRLITY -~ J2MODRLITY-20 *SCORE ~ .(2 2) 
 dL2 7 E.%oot 
M~ In.~t \]n~t 
E~VERB-38 
*IRSTAMCE-OF I E20RSE-32 Remove 
.~ *POS :IMSTAMCE-OF Nodify *MODRLITY Copy 
" atlon *MEANING *NAME Move 
,MRRKER-I#STAMCE ~E~e~xT\]i~Eqx *CASES 
*IMSTRMSE E~~6n 
*UNFILLEO-INDISP *MRRKER-CmEGORY Edit Oaemon Hack ~ 
: Ty~ QFILE ~ervln9 LUTE-3600-4 
Fig. 8 Snapshotofan experimenton the LUTE 
626 

References 

\[1\] Schank, R. C., and Abelson, tL P. Scripts, Plcms, GorEs, and 
Undertanding, LEA, t977. 

\[2\] Minsky, M. L., A Framework for Representing Knowledge, in The 
Psychology of Computer Vision, Winston ted J, McGraw liill, 1975. 

\[3\] Quill)an, M., R, Semantic Memory, in Semantic \[nyorraatio~ 
Processing, Minsky ted.), MIT Press, 1968. 

\[4\] Nomura, II., Towards the tIigh Ability Machine Translation, 
Joint Japancsed'~uropcan Work Shop on Machine Translation, 
Nov. 1983. 

\[51 Nomura, II., I!',xperimental Machine Translation Systems LUTI:,, 
Second Joint Japanese European Work Shop on Machine 
Translation, Dee. 1985. 

\[6\] Shimazn, A., Naito, S., and Nomura, lI., Japanese Language 
Semantic Analyzer based on an Extended Case Frame Model, 
Proceedings of the International Joint Conference on Artificial 
Intelligence, Aug. 1983. 

\[7\] ilda, lI., Ogura, K., and Nomura, tI., A Case Analysis Method 
Cooperating with ATNG and its Application to Machine 
Translation, Proceedh~gs of the \] 0th Interim)tonal Conference on 
Computational l,inguisties, July 1984. 

\[81 Naito, S., Sblmazu, A., and Nomura, i1., Classification of 
Modality Function and its Application to Japanense I,anguage 
Analysis, Preeeedings of Association for Colnputational 
Linguistics Conference, July 1985. 
