Generation of Japanese Sentences From Conceptual Representation

I Introduction

This is an attempt to generate Japanese sentences from conceptual representation obtained from Spanish newspaper stories [1]. This generator was first built as part of the Machine Translation Project in the Computer Science Department of Yale University in the summer of 1982. It has been sub­ sequently i~proved at the Electrotechnical Labora­ tory in Japan.

Conceptual representation is based on the Conceptual Dependency theory[2] and also on Memory Organization Packets (MOPs) [3]. The general idea behind MOPs is to store knowledge which is common to many different situations in only one processing structure, and then to make this processing struc­ ture available for use in a variety of appropriate situations. MOPs are useful for a Japanese genera­ tor because their contents are independent from the language, although CD theory and MOPs are con­ structed by using English . Such independency may be verified by applying these theories to Japanese sentences, structurally different from Romance lan­ guages, such as English, Spanish and French.

This generator has several features. First, it infers the temporal order of events included in the conceptual representation using causal chains and MOPs. Such inference is necessary to generate Jap­ anese sentences because most clauses and sentences are ordered temporally in Japanese. Second, appro­ priate conjunctives are inserted between the events using their causal relations. Third, subjects are omitted in relevant situations and two case markers for subjects are used in the generator to make the sentences suitable for Japanese. Other features are described in the following sections.

CONCEPT M-MOCK-TRIAL

ACTOR HUMll=

CONCEPT TERRORIST

ORG OBJ3= CONCEPT TERRORIST-ORG

MEMBERS HUMll

GENDER MALE

TYPE GUERRILLA

WEARING OBJO= CONCEPT CLOTHING

TYPE

SUIT

COLOR OLIVE-COLORED

PERSON

AT-LEAST 60

HUM6= CONCEPT

NUMBER

RESIDENCE COUNTRY

SCENE2

ACCl= CONCEPT ACCUSE

OBJECT HUM6

BAD-ACT UNDl

ACTOR HUMll

IR-FROM UNDl

SCENE4 EXEO=

CONCEPT EXECUTE

ACTOR HUMll

PLACE LOCO= CONCEPT CITY

NAME SAN PEDRO PERULAPAN

OBJECT HUM6

IR-FROM UNDl

UNDl= CONCEPT UNDESIRABLE-ASSISTANCE

JUDGER HUMll

OBJECT OBJS= CONCEPT GOVERNMENT

ACTOR HUM6

TRYO= CONCEPT TRY

ACTOR HUMll

OBJECT HUM6

OBJECT

SCENEl

SCENE3

Fig.l Conceptual Representation for Story 1.

sukunakutomo 60 nin no noumintachi ga seifu ni AT-LEAST 60 PEASANTS GOVERNMENT

kyoryokushita node, terososhiki ni zokusu COLLABORATED BECAUSE TERRORIST-ORG

oribuiro no fuku o kita geriratachi wa sono OLIVE-COLORED SUIT WEARING GUERRILLAS THE

noumintachi o kokuhatsushite, saiban-ni-kaketa. PEASANTS ACCUSED TRIED

sorekara, sono geriratachi wa san pedoro perurapan THEN THE GUERRILLAS SAN PEDRO PERULAPAN

toiu machi de sono noumintachi o shokeishita. CITY THE PEASANTS EXECUTED

Fig.2 Generated Japanese Sentences and Correspond­ ence with English Words.

The generator has a Japanese has Japanese entries for verbs, which correspond to the concepts dictionary. It nouns and others and the proper

nouns such as shown in Fig.l.

IV GENERATION OF JAPANESE SENTENCES

One of the main functions of this generator is the inference of temporal order of events using MOPs and causal chains. The temporal order is im­ portant because most clauses and sentences are ordered temporally in Japanese . The parser gives the conceptual representation and its main node to the generator. The generator first extracts all the events connected to the main node.

Then the events are sorted temporally and ap­ propriate conjunctives are inseted between them. For example , in Story 1,

(UNDl node, ACCl, TRYO. sorekara EXEO)

is the whole structure to be generated for the Story 1, where all the events are ordered temporal­ ly and two conjunctives are inserted there. The generator starts to translate the left side events and then processes the next to the right until 'the list becomes NIL.

A. Temporal Ordering of Events Using MOPs

1. Ordering With Scenes in MOPs

Every event is related to another event by a causal chain or a scene relation in MOPs. The caus­ al chain , indicated by such relations as LEAD-TO, REASON, GOAL, PRECONDITIONS etc. specifies its tem­ poral order between the two events. The scene rela­ tion shown in Fig .l gives a temporal order with other scenes in the MOP; SCENEl, SCENE2,etc. are in a temporal order. For example, in the MOP of M­ POLICE-CAPTURE, SCENEl is a description of a crime and the scene causes SCENE2, POLICE-SEARCH. As a result of the scene, AUTHORITY will arrest BAD-GUY in SCENE3 . These events are ordered temporally us­ ing these scenes if ACTORs and OBJECTs satisfy the conditions at each scene.

2. Ordering With Causal Relations And MOPs

The parser extracts causal relationships among events in stories when some of the events can not be expressed with scenes in MOPs.

'

However there are some cases where the temporal orders can not be determined by the causal chains given by the parser. For example, in Story3, the main node contains three events:

KIL2 A POLICEMAN KILLED A CRIMINAL.

TRAO : A POLICEMAN TRANSPORTED A CRIMINAL.

ESCl : A CRIMINAL TRIED TO ESCAPE.

The input story has a structure such as ,

KIL2 during TRAO. KIL2 when ESCl.

Then the parser gives the following temporal orders in the conceptual representation,

TRAO ~ KIL2 and ESCl -7 KIL2. The temporal relation between TRAO and ESCl is not clear here. The generator has to infer the relation using MOPs and the conceptual representation to generate Japanese sentences .

Connectivity of locations is one of the clues of inference when either of the event has a con­ cept of PTRANS(position transition). If the loca­ tion where ESCl occured is confirmed to be the des­ tination or on the way of.- TRAO, then the generator can infer that ESCl occured during or after TRAO occued. But there is no such information in the conceptual representation of Story 3.

• Next, the generator looks at CONTROL and DIS­ ABLE- CONTROL relation. The concept of TRAO, TRANS­ PORT-CRIMINAL, has the template as follows,

TEMPLATE ACTOR

AUTHORITY

OBJECT

BAD-GUY

PRECONDITIONS (CONTROL ACTOR OBJECT) .

This template means that an authority keeps a badguy under control during the transportation. On the other hand, ESCAPE is the ESCl's concept and it has an antecedent, DISABLE-CONT. Therefore, ESCAPE sat­ isfies the properties of DISABLE-CONT. After check­ ing the correspondence between ACTOR of TRAO and OBJECT of ESCl and also OBJECT of TRAO and ACTOR of ESCl, the generator infers that the temporal order of the events should be TRAO ~ ESCl. The whole structure of Story 3 thus becomes

(TRAO toki, ESCl node, KIL2).

B. EVENT GENERATION

An event is generated by the following func­ tions,

(GEN-SUBJECT), (GEN-VERB-MODIFIERS), (GEN-VERB). (GEN-SUBJECT) picks up an ACTOR slot from the event. The function chooses an OBJECT slot and the event is expressed in a passive form when the ACTOR slot is not filled.

A case marker "wa" usually follows a subject. Another case marker "ga" is also used often for a subject. The difference between them is so subtle and highly semantic that it is not easy to treat. Roughly speaking, "ga" is used in subordinate clauses which express time, reason, goal, etc. And "wa" is usually used in main clauses. This approx­ mation is adopted in this generator.

(GEN-VERB-MODIFIERS) generates OBJECT and other modifiers for the verb. It has standard pairs of properties and case markers listed below,

((TIME ni) (PLACE de) (FROM kara) (TO e)

(INST de) (OBJECT o)).

These modifiers for verbs are generated with this order. A case marker must be specified at the verb's entry in the Japanese dictionary when the property has another type of case markers.

(GEN-VERB) generates a verb which is appropri­ ately inflected according to the tense and the mode of the event. The tense ·inflection of the verbs are specified according to their endings.

C. NOUN PHRASE GENERATION

A noun phrase is generated by the following functions when the noun is processed for the first time, (GEN-NOUN-MODIFIERS), (GEN-NOUN). The second time a noun is needed, it is generated by (GEN-PRONOUN).

For example, when CRIMINAL FIDEL GONZALEZ is generated for the first time, (GEN-PRONOUN) is used the next time. The function generates THE CRIMINAL instead of pronoun "he". Because in Japanese, espe­ cially in newspapers, pronouns corresponding to "he", "she" etc. are rarely used. There is a tend­ ency to use more specific words shown above than such pronouns.

An order of the modifier generation for a noun is specified as, TYPE NUMBER AGE NAME STATUS etc. Other modifiers not specified here are generated according to the order in the node representation

and they precede these modifiers.

D. ELLIPSES

Subjects in main clauses are usually omitted when their subordinate clauses have the same sub­ jects in Japanese. In Story 1 shown in Fig. 1, the subject and the object in event ACCl are same as those in TRAO. In the Japanese translation, not only is the subject omitted in TRAO, but also the object is as shown in Fig.2. An object is not omit­ ted if the two events have no common subject.

V CONCLUSIONS

This generator works well in translating con­ ceptual representation into Japanese sentences dur­ ing the processing of 15 terrorism stories. Concep­ tual representation is expressed to be language in­ dependent if the culture or the customs behind the language are not very different from those of the representation. In those stories, the MOPs are not changed for the generation.

The problem of focus is not discussed in this paper. Word order and nested structure should be taken into account if this problem is treated by the generator. But the temporal ordering of events and other methods proposed here are necessary for generation regardless of the focus problem.

ACKNOWLEDGEMENTS

The author first thanks Professor R.C. Schank for his guidance and encouragement. And he also thanks S.L. Lytinen for many of the functions used in the generator. Additional thanks to D.Littleboy, L. Birnbaum and MT Project members for their impor­ tant suggestions and cooperations, and to Dr. H. Tanaka who supported this work at ETL.