THE JAPANESE GOVERNMENTPROJECT FOR MACHINE TRANSLATION
Makoto Nagao, Jun-ichi Tsujii, and Jun-ichi Nakamura Department of Electrical Engineering Kyoto University Sakyou-ku, Kyoto, Japan 606
1 OUTLINE OF THE PROJECT post-editing, access to grammar rules, and dictionary maintenance. The project is funded by a grant from the Agency of The project is not primarily concerned with the devel- Science and Technology through the Special Coordi- opment of a final practical system; that will be developed nation Funds for the Promotion of Science and Technol- by private industry using the results of this project. ogy, and was started in fiscal 1982. The formal title of Technical know-how is already being transferred gradu- the project is "Research on Fast Information Services ally to private enterprise through the participation in the between Japanese and English for Scientific and Engi- project of people from industry. Software and linguistic neering Literature". The purpose is to demonstrate the data are also being transferred in part. Finally, complete feasibility of machine translation of abstracts of scientific technical transfer will be done under the proper condi- and engineering papers between the two languages, and tions. as a result, to establish a fast information exchange The Japanese source texts being used are abstracts of system for these papers. The project term was initially scientific and technical papers published in the monthly scheduled as three years from the fiscal year of 1982 JICST journal d Current Bibliography of Science and with a budget of about seven hundred million yen, but, Technology. At present, the project is only processing due to the present financial pressures on the government, texts in the electronics, electrical engineering, and the term has been extended to four years, up to 1986. computer science fields. English source texts will be The project is conducted by the close cooperation abstracts from INSPEC in these fields.. The sentence between four organizations. At Kyoto University, we structures used in abstracts tend .to be complex compared have the responsibility of developing the software system to ordinary sentences, with long nominal compounds, for the core part of the machine translation process noun-phrase conjunctions, mathematical and physical (grammar writing system and execution system); gram- formulas, long embedded sentences, and so on. The mar systems for analysis, transfer and synthesis; detailed analysis and translation of this type of sentence structure specification of what information is written in the word is far more difficult than ordinary sentence patterns. dictionaries (all the parts of speech in the analysis, trans- However, we have not included a pre-editing stage fer, and generation dictionaries), and the working manu- because we wanted to find the ultimate limitations on als for constructing these dictionaries. The Electro- handling this type of complex sentence structure. technical Laboratories (ETL) are responsible for the Our system is based on the following concepts: machine translation text input and output, morphological 1. The use of all available linguistic information, both analysis and synthesis, and the construction of the verb surface and syntactic. The writing of as detailed as and adjective dictionaries based on the working manuals possible syntactic rules. The development of a gram- prepared at Kyoto. The Japan Information Center for mar writing system that can accept any future level of Science and Technology (JICST) is in charge of the noun sophisticated linguistic theory. dictionary and the compiling of special technical terms in 2. The introduction of semantic information wherever scientific and technical fields. The Research Information necessary to enable the syntactic analysis to be as Processing System (RIPS) under the Agency of Engineer- . # . accurate as possible. The importance of semantic mg Technology is responsible for completing the machine information not over-estimated; a well-balanced translation system, including the man-machine interfaces usage of both syntax and semantics. Heavily seman- to the system developed at Kyoto, which allow pre- and
Copyright1985 by the Association for Computational Linguistics. Permission to copy without fee all or part of this material is granted provided that the copies are not made for direct commercial advantage and the CL reference and this copyright notice are included on the first page. To copy otherwise, or to republish, requires a fee and/or specific permission. 0362-613X/85/02091-111503.00
Computational Linguistics, Volume I I, Numbers 2-3, April-September 1985 91 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
tics-oriented analysis is very attractive and effective With these points in mind, we developed a new soft- for sentences within narrow limits, but a system of ware system for machine translation comprising the that type cannot cope with the complicated structures language used to specify the grammar rules and the found in descriptions of the wider world where execution system. We call it GRADE (GRAmmar DEscri- semantic description becomes almost impossible. ber). 3. There are many exceptional linguistic phenomena that 2.2 THE STRUCTURE OF GRADE are more word-specific than explainable in general linguistic theory. The system should be able to accept The data format used to express the structure of a word-specific rules. In our system, these rules are sentence during the analysis, transfer, and generation written into the lexical entries, with the priority given phases has a large influence on the design of the gram- to these grammar rules in the analysis, transfer, and mar writing language. GRADE uses an annotated tree synthesis phases. This mechanism allows the system structure to represent the sentence structure during the to be upgraded step by step by the accumulation of translation process. Grammatical rules in GRADE are linguistic facts and word-specific rules in the diction- described in the form of tree-to-tree transformations with ary and effectively bypasses any deadlock in system each node annotated. The annotated tree in GRADE is a improvement. tree structure whose nodes are annotated by sets of 4. The system must be able to produce an output with property-value pairs. This tree-to-tree transformation an imperfect sentence structure and containing gives a great power of expression to rewriting rules that untranslated original words rather than fail in cases can be used in the grammars for the analysis, transfer, where the analysis was imperfect. From the post- and synthesis phases of the machine translation system. editor's point of view, an imperfect output is far pref- Annotation parts can be used to express information such erable to no output at all. as syntactic category, number, semantic markers, and Many other concepts and methods have been devel- other properties. They can also be used as flags to oped in our machine translation system, and these are control rule application. explained in the sections following. This paper concen- A rewriting rule in GRADE consists of a declaration trates on the main features of the Japanese to English part and a main part. The declaration part has the follow- translation system. Details of the English to Japanese ing four components: system, which is also included in our national machine • Directory entry part, containing the grammar writer's translation project, is being developed, and the result will name, the version number of the rewriting rule, and the be published shortly. last revision date. This part is not used at execution time. The grammar writer can access the information 2 THE GRAMMAR WRITING SYSTEM, GRADE using the HELP facility in GRADE. • Property definition part, where the grammar writer 2.1 OBJECTIVES OF THE SOFTWARE SYSTEM declares the property names and their possible values. In developing a machine translation system, the grammar • Variable definition part, where the grammar writer rules should accurately reflect the intention of the gram- declares the names of the variables. mar writer. This is fundamental to the achievement of a • Matching instruction part, where the grammar writer good grammar system. One of the basic necessities of specifies the mode of application of the rewriting rule to any machine translation system is a programming an annotated tree. language to write the grammar composed of the language The main part specifies the transformation in the for specifying the grammar rules and the accompanying rewriting rule, and has the following three parts: execution system. • Matching condition part, which describes the conditions A grammar-writing language for machine translation for the structure of trees and the property values of that is powerful must fulfill the following requirements: nodes. 1. The language must allow manipulation of linguistic • Substructure operation part, which specifies the oper- characteristics in both source and target languages. ations for the parts of the annotated tree that match the The linguistic structure of Japanese differs greatly conditions written in the matching condition part. from that of English. For instance, in Japanese, the • Creation part, which specifies the structure and the restrictions on word order are not so strong, and property values of the transformed annotated trees. some syntactic components can be omitted. A gram- The matching condition part allows the grammar writ- mar writer must be able to reflect these sorts of char- er to specify not only a specific structure for an anno- acteristics. tated tree but also structures that may repeat several 2. It is desirable that the grammar-writing language use times, structures that are optional, and structures where the same framework for writing the grammars in the the order of the substructures is unrestricted. analysis, transfer, and synthesis phases. The grammar The substructure operation part specifies operations writer should not be forced to learn several different on the parts of the annotated tree that match in the systems for the different translation stages. matching condition part. It allows the grammar writer to assign a property value to a node, or to assign a variable
92 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 Makoto Nagao, Jun-ichi Tsujii, Jun-iehi Nakamura Japanese Government Project for MT
to a tree or property value. The variable is declared in A dictionary rule is called by the CALL-DIC function the variable declaration part. It also allows him to call a in the substructure operation part. When CALL-DIC is subgrammar, a subgrammar network (which is explained executed by an entry word and rule identifier as keys, the below), a dictionary rule, a built-in function, or a LISP dictionary rule is retrieved and is applied to the part of function. In addition, the grammar writer can specify a the annotated tree specified by the grammar writer. conditional operation using the IF-THEN-ELSE state- Any grammar writing language must be able to resolve ment. the syntactic and semantic ambiguities found in natural The structure and the property values of the trans- languages. GRADE allows the grammar writer to merge formed annotated tree are written in the creation part. the results of all possible tree-to-tree transformations for The transformed tree is described by node labels that are a particular subgrammar. However, it must avoid any used in the matching condition part or the substructure combinatorial explosion when it encounters ambiguities. operation part. For instance, let us take the case where a grammar The matching instruction parr of a rewriting rule speci- writer writes a subgrammar to analyze the case frame of fies the application path through the annotated tree. a verb, containing two rewriting rules; one rule is to Paths through the trees are specified by combinations of construct a VP (verb phrase) from a V and NP (verb and the four basic modes: left-to-right, right-to-left, noun phrase), and the other is to construct a VP from a bottom-to-top, and top-to-bottom. V, NP, and PP (verb, noun phrase, and prepositional GRADE allows the grammar writer to divide the whole phrase). When he specifies the NONDETERMINISTIC- grammar into several subgrammars and to describe the PARALLELED mode for the subgrammar, the GRADE phases of the translation process separately. A subgram- executor applies both rewriting rules to the input tree, mar may correspond to a grammatical unit such as the constructs two transformed trees, and merges them into a parsing of a simple noun phrase or the parsing of a new tree whose root node has the special PARA property. simple sentence. The network of subgrammars forming The root node is called a para node and the subtrees the whole grammar allows the grammar writer to control under this node are the trees that have been transformed the translation process in detail. If the subgrammar by the rewriting rules. Figure 1 shows this mode applied network in the analysis phase consists of the subgrammar to create a para node. for a noun phrase (SG1) and the subgrammar for a verb The grammar writer can select the most suitable phrase (SG2) in this sequence, the GRADE executor first subtree under the para node by applying an subgrammar applies SG1 to the input sentence, and then applies SG2 that assigns a priority value to each subtree and using a to the result. built-in function that orders the subtrees according to their values. 2.3 SOME SPECIFIC FEATURES OF GRADE A para node is treated the same as other nodes in the GRADE allows a grammar writer to write word-specific current implementation of GRADE. The grammar writer grammar rules as a subgrammar at the word dictionary can use the para node as he wants, and can select a entry level. A subgrammar written in a dictionary entry subtree under a para node at a later application of the is called a dictionary or lexical rule. A dictionary rule is grammar rule. specific to a particular word in the dictionary.
PARA A ~oo~ V NP PP SG
VP PP VP A A V NP V NP PP
Figure 1. Example of para node formation.
Computational Linguistics, Volume I 1, Numbers 2-3, April-September 1985 93 Makoto Nagao, .lun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
2.4 SYSTEM CONFIGURATION AND ENVIRONMENT differently. Distinguishing these different usages requires The system configuration of GRADE is shown in Figure careful investigation of the context around the verb. As 2. Grammar rules written in GRADE are first translated described in section 2, GRADE allows the definition of by the GRADE translator into internal forms, expressed grammar rules that are applied only to specific lexical as S-expressions in LISP. The internal forms of the gram- items. We use this capability to discriminate between mar rules are applied to the input tree that is output by verb usages. However, many verbs have only two or the morphological analysis program. The rules are three different usages at most. We prepared a fixed applied by the GRADE executor, and the results are sent format for the lexical coding of these verbs. Descriptions to the morphological generation program. in this format are converted to internal representation in The GRADE system program is written in UTILISP GRADE automatically by a program. For verbs that have (University of Tokyo Interactive LISP) and implemented a wide range of usages, where the rules need to be writ- on a FACOM M382 computer, which can handle Chinese ten based on a variety of heuristic information, the gram- characters. The system will also run on the Symbolics mar rules can be written directly in GRADE, bypassing 3600. The system program contains about 10,000 lines. the fixed format. In the fixed format, a verb can have several case frames corresponding to different usages. A case frame in Japanese is represented as a set of triplets 3 THE DICTIONARIES like: Because our system is based on the "transfer" approach, (Surface-Case-Mark Deep-Case Constraints-on NP) there are three separate dictionaries (for analysis, trans- fer, and synthesis). In this project, Japanese words are
Dlctlonary-,,. /Grammar / \ Dictionary Grammar rule (Interns] form)
sententtal tree [executor I sententtal tree
Figure 2. System configuration of GRADE.
94 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
I ~.-~ U ~ 4..I e-i m 0 0 0 0 U~ 1 4.1 oo •i-I ,.o 0 o Ill I..i O ~ .l,J i ,g CJ I~.~ fa.1 r.3 C..I 0 0 0 ~ aJ aJ ~ C~ 0 ~ O0 0 ~J 0 W 0 U O ~ ~J 0 ~'~ ,~ l ~ l I 0 E~ tJ U 0J
m o,,~ 0 ~ ~ m ~ [ ~ NN®
O cl I1 o tJ U "ICl O2 I
O N N N N 0 i...I.,-I ~ O
O ~ O ~ mCJ
IJ ~J
~J ~ g ~ m ~ ."4 O W 4-1 0 N I tJ gh ~ 4J ,IJ U g 4J m 0 ~J O~ ~ 0 m QJ 4J
0 Z ®~-40
0
cJ 1 4J ,.~ m ,JO bOO 0 ~J I ~ .,,4 O~ ~ 4J 4.I N N N N N
Figure 3. System of semantic primitives for nouns.
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 95 Makoto Nagao, Jun-ichi Tsujii, Juff-ichi Nakamura Japanese Government Project for MT
Table 1. English case labels, a. Subcategorization of nouns: proper noun, common 1. AGenT 17. RANge noun, action noun, adverbial noun, postpositional 2. Gausal-POtency 18. COmpaRison noun, conjunctive noun, complementizer. 3. EXPeriencer 19. TOOl b. Semantic codes: The semantic codes shown in Figure 3 are used. 4. OBJect 20. PURpose 5. RECipient 21. Space-FRom c. Information on collocation: Adjectives, nouns, etc., which often occur together with a noun are specified. 6. ORigin 22. Space-AT This information plays a role similar to the case frame 7. SOUrce 23. Space-TO 8. GOAl 24. Space-THrough of a verb, and is effective in discriminating the differ- ent usages and meanings of nouns. 9. COntent 25. Time-FRom We also have fixed formats for words in the other 10. PARtner 26. Time-AT 11. OPPonent 27. Time-TO morpho-syntactie classes. All words in the dictionary 12. BENificiary 28. DURation have, besides the above information, the properties listed in Table 2. 13. ACCompaniment 29. CAUse 14. ROLe 30. CONdition Table 2. Properties of Japanese words in the dictionary. 15. DEGree 31. RESult 16. MANner 32. COnCession lexical item subject code word length lexical entry in the Since one case frame corresponds to one usage of a transfer dictionary verb, and each usage corresponds to a different word stem semantic codes "meaning" of the verb, the lexical properties of verbs are (for nouns, verbs) represented by the properties of each case frame. The following properties are coded for each case frame. pronunciation thesaurus codes a. Aspectual features: stative, semi-stative, durative, (for nouns) resultant, transitional. part of speech idiomatic expressions b. Volition: volitional verb, non-volitional verb, sub-categorization of case frames c. Possible transformations of surface case markers: part of speech (for verbs, adjectives) Some auxiliary verbs that follow the verb and express conjugation passive, or causative voice, etc., change the surface case marking; that is, the postpositional case particles synonym described in SCM are changed. Which auxiliary verbs derivations (noun, verb, can follow, and what transformation of surface case adjective, adverb) markers is caused by an auxiliary verb, depend on the related words verb itself, and so are marked as a lexical property of the verb. 3.2 THE JAPANESE TO ENGLISH d. Idiomatic expressions: Information on collocation; for TRANSFER DICTIONARY example, which nouns and adverbs are often collocat- 3.2.1 TRANSFER DICTIONARY FOR VERBS ed with the verb are described in this column. e. Lexical entry in transfer dictionary: As described Different verb usages are discriminated during the analy- before, a verb may have more than one case frame] sis phase. This means that usage ambiguities are partially each of which corresponds to one "meaning" of the resolved before the transfer phase. However, the verb. The transfer dictionary contains an entry for concept of "meaning" (usage) applied to a single word is each meaning. Thus a single surface verb in Japanese very vague and in fact depends greatly on the language may correspond to several different entries in the pairs we have for translation. For example, the verb transfer dictionary. On the other hand, certain usages NOMU in Japanese can be used in the following ways: ,of different surface verbs may be reduced to a single Tabako-wo NOMU ~ smoke a cigarette entry in the transfer dictionary, if they are synony- Kusuri-wo NOMU ~ take medicine mous. Mizu-wo NOMU ~ drink water f. Semantic class: This property is used for semantic classification of verbs such as "mental-action", These three cases should be translated differently. In a "physical-transfer", etc. similar way, the English verb to wear is used as: g. Miscellaneous properties: Several other minOr proper- Wear a suit ~ Suutsu-wo KIRU ties are coded in the current dictionary. Wear black shoes ~ Kuroi Kutsu-wo HAKU 3.1.2 ANALYSIS DICTIONARY FOR NOUNS Wear spectacles ~ Megane-wo KAKERU Wear a wristwatch ~ Udedokei-wo SURU The following properties are described using the fixed noun format:
96 Computational Linguistics, Volume I !, Numbers 2-3, April-September 1985 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
These four cases should be translated differently into The fixed format for nouns includes the following Japanese. Some might claim that these verbs are very items: ambiguous and have different meanings; but this contra- a. Conditions on the sequence of words in the preceding dicts the intuitive conclusion that suggests it is reasonable part: to consider that the target language simply has more A set of default rules that translate Japanese postpo- specific verbs in these cases. In other words, discrimi- sitional case particles to English prepositions is nation in meaning at the analysis stage is not sufficient to provided in the transfer grammar. However, these select the appropriate target verb in these cases. default rules are often violated, because certain The verb transfer dictionary is divided into two parts: English nouns require specific prepositions. This kind a word selection part and a mapping part. The word of information is coded in this column. selection part is used to choose appropriate target verbs b. Conditions on the sequence of words in the succeed- by referring to semantic markers of the case elements. ing part: The semantic markers currently being used appear to be Postpositions that follow the noun often give a clue to insufficient to decide appropriate target verbs in certain the morpho-syntactic class conversion. cases. We cannot, for instance, distinguish medicine and c. Collocation with verbs: cigarette with the current set of semantic markers, which Certain combinations of nouns and verbs in Japanese is relevant to choosing appropriate English verbs for are translated into English as single verbs, and certain NOMU. However, we can treat such problems by specify- combinations of nouns and verbs such as kusuri ing word selection rules in the noun transfer dictionary. (medicine) and NOMU (to smoke, to drink, to take) The mapping part gives the correspondence of the require specific translation of the verb (to take). This deep cases in Japanese and English. In most cases, the is the kind of information coded here. Japanese deep case maps to the same deep case in 3.3 ENGLISH GENERATION DICTIONARY English. There are, however, certain deep cases that are interpreted differently in the two languages. The format for verbs includes the following items: Sometimes a single Japanese verb can not be trans- a. Components: In the transfer phase, certain Japanese lated into a single English verb and has to be paraphrased verbs are translated into English expressions contain- using a combination of a verb and another element such ing not only verbs but also prepositional phrases as a noun or a prepositional phrase. For example, and/or adverbial particles (off, up, etc.). These complex expressions have separate entries in the SHISAKUSURU ~ develop (something) on a trial basis generation dictionary, and the structural descriptions Such linguistic expressions are also treated in the for the complex expressions are given here. mapping part. b. Verb patterns: The verb codes from the Longman Although many verbs in the transfer dictionary are Dictionary of Contemporary English are used to speci- coded in this fixed format and converted to lexical rules fy the syntactic patterns a verb can take. in GRADE by a program, we also write lexical rules c. Aspectual features: stative, transitive, process, directly for verbs that have a wide range of usages. completive, momentary. d. Voice: usually passive, can be used in passive voice, 3.2.2 TRANSFER DICTIONARY FOR NOUNS cannot be used in passive voice. Some Japanese words that behave morpho-syntactically e. Volition: volitional verb, non-volitional verb. as nouns have to be translated into English words in f. Agent of to-infinitive: SUBject (I promise him to go), other morpho-syntactic classes. Such class conversions OBJject (I want him to go). should be treated in the transfer dictionary, because they g. Case frames: A case frame of a verb is expressed by a are highly dependent on the lexical item. For example: set of quadruplets like: (i) TAIWA-KEISHIKI-de JIKKOUSURU Surface-Case Deep-Case Syntactic-Form Semantic-Code (interaction) (to execute) SC DC SF SEC -- to execute interactively SF is a list of numbers, each of which expresses one syntactic form the case element can take: (ii) PUROGURAMU-MOODO-de JIKKOUSURU 1. noun phrase (program mode) (to execute) 2. infinitive without to -~ to execute in program mode 3. to-infinitive The above two examples have exactly the same struc- 4. -ing tures in Japanese (where the noun phrases 5. that-clause TAIWA-KEISHIKI-de and PUROGURAMU-MOODO-de fill 6. wh-clause the same deep case, "manner") but translate to different 7. adjective English structures simply because an appropriate lexical 8. -ed item exists for (i) but not for (ii). The formats for other parts of speech are described in the Japanese literature.
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 97 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
4 JAPANESE SENTENCE ANALYSIS 4.2 TYPICAL STEPS IN THE ANALYSIS GRAMMAR 4.1 ANALYSIS STRATEGIES 4.2.1 SIMPLE SENTENCES As pointed out by Wilks, semantic information cannot be As described in 3, the analysis dictionary for verbs used as constraints on single linguistic structures; it can contains verb case frames that are expanded to GRADE be used only as preference cues to help choose the most rules with unrestricted word order to obtain a match with feasible interpretation from among all the syntactically the input sentence structure. Certain verbs such as ARU, possible interpretations. We believe that many types of NARU, SURU, MOTSU, etc., which have a ,-,~le range of preference cues, besides semantic ones, exist in real texts, usages, are discriminated by directly coding SG~,, ":" the and these cannot be captured by CFG rules. By making dictionary. use of various types of preference cues, our analysis 4.2.2 RELATIVE CLAUSES grammar for Japanese can work almost deterministically to give the most preferable interpretation at the first Relative clause constructions in Japanese express several output, without extensive semantic processing. different relationships between modifying clauses (rela- In order to integrate heuristic rules based on various tive clauses) and their antecedents. Some relative clause levels of cues into a unified analysis grammar, we have constructions cannot be translated into English as rela- introduced the following principles in the analysis of tive clauses. We classified Japanese relative clauses into Japanese sentences: four types, according to the relationship between the 1. Explicit control of rule application: Heuristic rules clause and its antecedent. Because these four forms of can be ordered according to their strength. relative clauses have the same surface forms, like 2. Multiple relation representation: Various levels of ...... (verb) (noun) information including morphological, syntactic, Relative Clause Antecedent semantic, and logical are expressed in a single anno- tated tree and can be manipulated at any time during careful processing is required to distinguish between the analysis. This is required not only because many them. We have developed a sophisticated analysis proce- heuristic rules are based on heterogeneous levels of dure that uses the various levels of heuristic information. cues but also because the analysis grammar should be 4.2.3 NOUN PHRASE CONJUNCTIONS able to perform semantic/logical interpretation of sentences at the same time, and the rules for these Noun phrase conjunctions often appear in abstracts of phases should be written using the same framework scientific and technical papers. It is important to analyze as the syntae,tic analysis rules. them correctly, especially in correctly determining the 3. Lexicon-driven processing: We can write heuristic scope of the conjunction, because they often lead to a rules specific to a single or a limited number of words, proliferation of the analysis results. We have many such as rules concerned with collocation among heuristic rules based on various types of information. words. These rules are strong in the sense that they Some are based on surface lexical items, some on word almost always succeed. They are stored in the lexicon morphemes, and some on semantic information. They and invoked at the appropriate time during the analy- are used differently in different conjunctive structures. sis without decreasing efficiency. We can distinguish strong heuristic rules (that is, rules 4. Explicit definition of analysis strategies: The whole that almost always give correct scopes when applied) analysis phase can be divided into steps. This makes from others. In fact, there is some ordering of heuristic the whole grammar efficient, natural, and easy to rules according to their strength. In GRADE we can read. Furthermore, strategic consideration plays an define arbitrary ordering of rule applications by using essential role in preventing undesirable interpretations subgrammar networks and also by ordering rewriting from being generated. rules inside a subgrammar. This capability of being able Figure 4 shows the overall organization of our current to control the rule application sequence is absolutely analysis grammar. The main components are: necessary in integrating heuristic rules based on heter- 1. Morphological Analysis ogeneous types of information into a unified set of rules. 2. Analysis of Simple Noun Phrases 4.2.4 SENTENCE RELATIONSHIPS AND 3. Analysis of Simple Sentences OUTER CASE ANALYSIS 4. Analysis of Embedded Sentences (relative clauses) In Japanese there are several different syntactic 5. Analysis of Sentence Relationships constructions corresponding to English subordinators and 6. Analysis of Outer Cases coordinators like although, in order to, and, and so on. 7. Contextual Processing (processing of omitted case The correspondence between forms of Japanese and elements, interpretations of ha, etc.) English sentence constructions is not straightforward. The analysis produces dependency tree structures Some postpositional particles in Japanese express several showing the semantic relationships between the words in different semantic relationships between sentences, and the input sentence. therefore should be translated into different subordina-
98 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 Makoto Nagao, Jun-ichi Tsujil, Jun-ichi Nakamura Japanese Government Project for MT
START)
morphologica~I analysis J ?e:ision on1 scopesof~ imple sentences f Decisionon 1 scopesof h A~lysis~ ~alysis of -~ ~noun phraseconjunction~ [ relative I ~mantic _ ~f I ~/~clause~ ~lationship in~~Analysisof ~oun phrased.j/? T ~un phrases~ ~nalysis°f simplesentencey / @nversion of "ha' to "n~ / ~ecision:n aspec~ / ~nalysi'so~ semantiC/ Velationshipsbetween F \ simplesentences ~ ecision on ten~ ~Contextual1 analys:~ deletedcases etc Phrase structure1 t h pendencystructur~ (END)
Figure 4. Basicflow of processing.
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 99 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT tors in English according to the semantic relationships. Lexical rules for 62 case particles have also been imple- The postpositional particle TAME can express either a mented and tested. "purpose-action" relationship or a "cause-effect" relationship. In order to resolve the ambiguity in the 5 TRANSFER AND GENERATION OF ENGLISH semantic relationships expressed by TAME, a set of lexi- In principle we use the deep case dependency structure to cal rules is defined in the dictionary for the entry TAME. represent a sentence semantically. Theoretically it is The rules are roughly as follows, where the sequential possible to assign a unique case dependency structure to form ($1, S2) is assumed: each input sentence. In practice, however, the analysis (i) If $1 expresses a completed action or a stative phase may fail, or it may assign the wrong structure. assertion, the relationship is "cause-effect". Therefore, as an intermediate representation, we use a (ii)If $1 expresses neither a completed action nor a structure that makes it possible to annotate multiple stative assertion, and $2 expresses a volitional action, possibilities as well as multiple level representation. the relationship is "purpose-action". Properties at a node are represented as vectors, so that this complex dependency structure is flexible in the sense Note that whether S1 expresses a completed action or not that different interpretation rules can be applied to the is determined in a preceding phase by using rules that structure. utilize the aspectual features of the verbs described in the Transfer and generation rules are organized along the dictionary, and auxiliary verbs following the verb. We principle that "if a better rule exists, then the system uses have heuristic rules for 57 postpositional particles for it; otherwise, the system attempts to use a standard rule: sentences conjunctions like TAME. if that fails, the system uses a default rule". The gram- Postpositional particles that follow noun phrases and mar involves a number of stages of application of heuris- express case relationships are also very ambiguous in the tic rules. Figure 5 shows the process flow for the transfer sense that they express several different deep cases. and generation phases. Although the interpretations of inner case elements are To obtain a more neutral (or target-language oriented) directly given in the verb dictionary as the form of structure, some heuristic rules are activated immediately mapping between surface case particles and their deep after the standard analysis of the Japanese sentence is case interpretations, the outer case elements should be finished. We call such activation the pre-transfer loop. semantically interpreted by referring to the semantic Semantic and pragmatic interpretations are done in the categories of noun phrases and the verb properties. pre-transfer loop. The larger the number of heuristic rules applied in this loop, the better the results.
pre-transfer post-transfer loop ~ loop internal TRANSFER internal represen tat i on representation for Japanese for English ANALYS IS// t GENERATION
treeI structure transformation
MORPHOLOGICAL SYNTHESIS Figure5. Process flow for the transfer and generation phases.
100 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
Table 3. Word selection in target language by using semantic markers.
non-living substance form-i form X(obj) X ~ ,,,stru,cture social phenomena take place ,take"'place action,deed,movement occur-i X occur X reaction standard,property state,condition arise-I X arise relation ¢ produce-2 produce X non-living substance X~Y~ Y structure form-i X form Y phenomena, action cause-i X cause Y ¢ produce-2 X produce Y property improve-i 'X improve Y _h~Y~ Xfi~Y~&~ Y measure increase-2 X increase Y ¢ raise-i X raise Y Semanticmarker for X/Y
Table 4. Default rule for assigning a case label of English to the Japanese postposition ni.
J-SURFACE-CASE J-DEEP-CASE E-DEEP-CASE Default Preposition
t- (ni) RECipient REC. BENeficiary 1to (REC-- to. BEN -- for) ORigin ORI from
PARticipant PAR wilh TIMe Time-AT in ROLe ROt as GOAl GOA Io
eoe ooo Q @ •
5.1 WORDSELECTION IN ENGLISH column labelled 0 is applied to produce a default trans- USING SEMANTIC MARKERS lation. Word selection in the target language is a big problem in The expressive power of format-oriented descriptions machine translation. There are varieties of choices for is, however, insufficient for a number of common verbs translation of a word in the source language. The main such as SURU 'to make, to do, to perform, ...' and NARU principles adopted by our system are: 'to become, to consist of, to provide, ...'. In such cases, 1. Area restriction using field codes, such as electrical we can write the transfer rules directly in GRADE. There engineering, nuclear science, medicine, and so on. must be a constant effort to list varieties of usages with 2. Semantic codes are attached to a word in the analysis their corresponding English sentence structures and phase and used for the selection of the proper target semantic conditions. language word or phrase. A postposition in Japanese represents a case slot for a 3. The sentence structure involving the word to be verb, but it has a variety of usages; thus determination of translated is sometimes effective in determining the the English preposition corresponding to each Japanese proper word or phrase in the target language. postposition is quite difficult. It also depends on the verb Table 3 shows an example of part of the verb transfer that governs the noun phrase having that postposition. dictionary. Selection of the English verb is done from Table 4 illustrates part of a default table for determin- the semantic categories of the nouns related to the verb. ing deep and surface case labels when no higher level A number (i) attached to the verb, like form-1 or rule applies. This sort of table is defined for all case produce-2, labels the i-th usage of the verb. When combinations. In this way, we confirm at least one trans- semantic information on the nouns is not available, the lation to be assigned to an input. The particular usage of a preposition for a particular English verb is written in
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 101 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT the lexical entry for the verb, and the information is used tal differences between the internal representation of for English sentence generation. Japanese and of English are absorbed in the pre-transfer Many odd structures are still left after the pre-transfer loop. But before the English generation phase, some loop and the lexical selection, and the internal English structural transformations are still required for cases such representation must be adjusted further into more natural as (a) embedded sentence structures, and (b) complex forms. We call this part the post-transfer loop. sentence structures. These structural adjustments are Global sentence structures are completely different in performed in the post-transfer loop. Japanese and English, and correspondingly the internal The steps comprising the transfer phase are shown in structures are also completely different. The fundamen- Figure 6.
[Apply heuristic rules (I t Simplify structure
~ransfer nouns, adjectives and determiners
l raos eooopouo oooos I eolde prepositions ,L output with adjective or not ITransfer adverbs 1 ITransfer verbs I
Transfer optional case, subordination and coordinatio~
Decide tense, aspect and moda4
Apply heuristic rules (II
Figure 6. Outline of the transfer phase.
102 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
5.2 ENGLISHSURFACE STRUCTURE GENERATION Any big structural transformations required in the translation come from the essential differences between After transferring from the Japanese deep dependency English, which is a DO-language, and Japanese, which is structure to the English one, the structure is converted to a BE-language. In English, case slots such as tool, a phrase structure tree with all the surface words cause/reason, and some others often appear in the attached to the tree. subject position, while in Japanese such expressions are The conversion is performed top-down from the root never used. Transformations of this kind are incorpo- node of the dependency tree to the leaf. Therefore, rated in the generation grammar as shown in Figure 7. when a governing verb demands a noun phrase They produce more natural English expressions. The expression or a to-infinitive expression for its dependent stylistic transformation part of the process is still very phrase, a structural change must be made to the phrase. primitive. We need to accumulate much more linguistic Noun-to-verb transformations and noun-to-adjective knowledge and lexical data before we can produce really transformations are often required due to the difference natural English expressions. in expressions between Japanese and English. This proc- ess moves down from the root node to all the leaf nodes. After this phrase structure generation process, some 6 EVALUATION OF TRANSLATION QUALITY sentential transformations are performed. For example: The following two aspects of the machine translation • When the agent is missing, a passive transformation is output have been adopted to evaluate translation quality. applied. They are to some extent independent indicators. • When the agent and the object are both missing, the 1. Intellibility: An evaluation of the extent to which the predicative verb is nominalized and made the subject by translated text can be understood by a native speaker supplementing verb phrases such as is made or is of the target language. In Japanese to English trans- performed. lation, we evaluate the extent to which an average • When the subject phrase has a big tree, the anticipatory British or American reader can understand the output subject it is introduced. without any reference made to the Japanese original. • In compound and complex sentences, same subject 2. Accuracy: The degree to which the translated text nouns are pronominalized. conveys the meaning of the original text is evaluated, • Duplication of head nouns in conjunctive noun phrases and a measure of the amount of difference between is eliminated. For example, "uniform component and the input and output sentences. The evaluation is non-uniform component" is reduced to "uniform and done by Japanese translators specializing in Japa- non-uniform components". nese-to-English translation.
earthquake building collapse
collapse destroy
SUB~ ~AUSE CPO BJ
building earthquake earthquake building [CPO:causal potency] = The building collapsed due to the earthquake. The earthquake destroyed the buildings.
Figure 7. An example of structural transformations in the generation phase.
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 103 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
6.1 INTELLIGIBILITY 3) While the content of the input sentence is generally Evaluation of intelligibility is based on a scale of 1 to 5; conveyed faithfully to the output sentence, there are the categories are described below. [See Appendix A for some problems with things like relationships between translation examples.] phrases and expressions, and with tense, voice, 1. The meaning of the sentence is clear, and there are no plurals, and the positions of adverbs. There is some questions. Grammar, word usage, and style are all duplication of nouns in the sentence. appropriate, and no rewriting is needed. 4) The content of the input sentence is not adequately 2. The meaning of the sentence is clear, but there are conveyed to the output sentence. Some expressions some problems in grammar, word usage, and/or style, are missing, and there are problems with the relation- making the overall quality less than 1. ships between clauses, between phrases and clauses, 3. The basic thrust of the sentence is clear, but the eval- or between sentence elements. uator is not sure of some detailed parts because of 5) The content of the input sentence is not conveyed to grammar and word usage problems. The problems the output sentence. Clauses and phrases are missing. cannot be resolved by any set procedure; the evalu- 6) The content of the input sentence is not conveyed at ator needs the assistance of a Japanese evaluator to all. The output is not a proper sentence; subjects and clarify the meaning of those parts in the Japanese predicates are missing. In noun phrases, the main original. noun (the noun positioned last in the Japanese) is 4) The sentence contains many grammatical and word missing, or a clause or phrase acting as a verb and usage problems, and the evaluator can only guess at modifying a noun is missing. the meaning after careful study, if at all. The quickest As the evaluation number increases on the above scale solution will be a retranslation of the Japanese from 0 to 6, the accuracy decreases. This part of the sentence because too many revisions would be need- evaluation was done by four Japanese translators, each ed. of whom has one or two years experience in Japanese to 5) The sentence cannot be understood at all. No English translation. The whole evaluation process is amount of effort will produce any meaning. monitored by a Japanese translation specialist with As the evaluation number increases on the above scale extensive experience in translation work. from 1 to 5, intelligibility decreases. The evaluator uses 6.3 RESULTS OF EVALUATION the above scale to evaluate the output sentence without any reference to the Japanese original in the first place. We describe here the results of the evaluation of the When the output sentence contains untranslated words in translation of 1,682 sentences taken from the monthly Japanese, the English translation of those words is JICST journal A Current Bibliography of Science and provided by a Japanese rewriter before the evaluation. Technology. Of these, 791 were the ones often referred to This evaluation work has been carried out to date by one for the development of the analysis grammar, and the British and one American evaluator, neither of whom has remaining 891 were added as the test material this time. the ability to read or evaluate Japanese. Both evaluators All the sentences were given to the machine translation have one year's experience in proofreading and checking system with no pre-editing. The 791 sentences forming translations of general scientific and technical literature, the first group were originally selected out of 1000 after but neither has specialized knowledge in the field of elec- eliminating 120 that contained ungrammatical Japanese trical engineering, which has been used for the input expressions, and a further 90 that contained long math- material up to now. ematical or chemical formulae. The deletion of the latter was because, in the early stages, the analysis grammar 6.2 ACCURACY that would deal with formulae had not been completed. Accuracy is evaluated on a scale of 0 to 6; that is, seven The second group of 891 were all those that were in the categories. [See Appendix B for translation examples.] abstract, without any such selection. 0) The content of the input sentence is faithfully Tables 5 and 6 present the evaluation results for intel- conveyed to the output sentence. The translated ligibility and accuracy for the two groups of abstracts. sentence is clear to a native speaker and no rewriting Table 7 gives a comparison of the two groups. As the is needed. system was not tuned to the sentences in the second 1) The content of the input sentence is faithfully group, there were many unknown grammatical structures conveyed to the output sentence, and can be clearly and missing words in the dictionary, which made the understood by a native speaker, but some rewriting is evaluation result worse than the first group. needed. The sentence can be corrected by a native As these tables show, when the accuracy of translation speaking rewriter without referring to the original goes down, so too does the intelligibility. We did not text. No Japanese language assistance is required. find any examples of intelligibility being low when accu- 2) The content of the input sentence is faithfully racy was high, but we did find a reasonable number of conveyed to the output sentence, but some changes cases where the translation accuracy was evaluated as are needed in word order. low, but intelligibility was rated high. Table 8 lists typical sample sentences for each evaluation type.
104 Computational Linguistics, Volume I 1, Numbers 2-3, April-September 1985 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
Table 5. Evaluation results for the first group of 791 abstracts.
accuracy llncelllgibility m ---7- 1 2 I 3 4 5 do receive! total T 2 0 116 4 0 2-;;-9 32.7 3 0 16 3 259 4 24 12.5
5 36 1 58
total 82
percentage of ¢o ~.',1 23.6 22.4 12.4 12.4 ! 5.8 10.4
Table 6. Evaluation results for the second group of 891 abstracts.
uccurscy intelligibility I 0 pe:¢en~9= 1 2 3 4 5 l 6 defective ¢O~Sl of to~l f 1 61 0 7 5 0 3 7 0 83 9.3
2 0 142 22 27 8 13 9 0 221 24.8 3 0 0 138 68 44 26 17 4 297 33.3
4 0 0 I0 24 35 16 37 i 4 126 14.1
t 5 0 0 0 1 6 7 149 1 164 18.4 to=oz 61 141 177 125 93 65 219 9 891 perconCege of =o.= ~ 6.8 15.8 14.0 1.0 19.9 I0.4Il 7.3 24.6j
Table 7. Comparison between first and Table 8. Typical sample sentences in second groups for intelligibility. the different evaluation categories.
tntolligi first group soco~ group I second group accuracy flrst7g lgr°uP -~izity 7 9 1 8 9 1 891
1 14.7% 9.3% 0 12.4~ 6.8~ I 2 32.7~ 24.8~ l 23.6% 15.8%
3 32.7% 33.3~ 2 22.4% 19.9~
4 12.6% 14.1~ 3 12.4% 14.0%
5 7.3% 18.4% 4 12.4~ 10.4%
5 8% 7 3%
lO 4% 24 6%
0 6% I 0%
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 | 05 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
Just as there are no clear and objective criteria for naturalness of sentential expressions, as well as exact evaluating the quality of Japanese to English translations meaning, become more important. Depending on these done by humans, standard criteria for judging the results situations, the yardstick for intelligibility will change as of machine translation have yet to be established. The well. evaluation methods proposed in this paper are still in the trial stage, and much more refining and improving is still ACKNOWLEDGMENT needed. We are deeply indepted to professor Toyoski Nishida The translation quality and the amount of post-editing (Kyoto University), Mr. Yoshiyuki Sakamoto (ETL), needed is closely related to the quality and nature of the Tsuyoshi Toriumi (JICST), and Masayuki Sato (JICST) for original text. It is quite natural to expect that simple taking part in this project, and to many other people from sentences can be translated accurately and intelligibly. private companies for their help in the development of We need to develop some way to evaluate the degree of the system. We are grateful to the Science and Technol- difficulty of the original text along with the translation ogy Agency and to the Agency of Engineering Technolo- evaluation. Only within this wider context can accuracy gy for their constant funding of and guidance for the and intelligibility be meaningfully discussed. project. Professors Yutaka Kusanagi, Shinobu Takamat- The JICST abstracts used in this project were written su, Makoto Hirai, and others gave us comments during primarily with the aim of condensing as much informa- the development of the system. tion as possible into a few sentences. This means that there are many long sentences, many of which are not REFERENCES very correct from a linguistic point of view. This is one Nagao, Makoto; Nishida, Toyoaki; and Tsujii, Jun-ichi 1984 Dealing obvious factor contributing to the poor evaluation results with the Incompleteness of Linguistic Knowledge in Language shown in Tables 5 to 7. Translation. In Proceedings of COL1NG 84, Stanford University, Evaluation of the quality of machine-translated California: 420-427. sentences is closely linked to the way in which the Nakamura, Jun-ichi; Tsujii, Jun-ichi; and Nagao, Makoto 1984 Gram- mar Writing System (GRADE) of Mu-Machine Translation Project machine translation output is to be used, hence to the and its Characteristics. In Proceedings of COLING 84, Stanford ease with which post-editing can be done. Only a mini- University, California: 338-343. mum of post-editing will be necessary to convey the Sakamo, Yoshiyuki 1984 Lexicon Features for Japanese Syntactic technical meaning of the original to the specialist in a Analysis in Mu-Project-JE. In Proceedings of COLING 84, Stanford University, California: 338-343. particular field for the purpose of information service. Tsujii, Jun-ichi; Nakamura, Jun-ichi; and Nagao, Makoto 1984 Analy- However, when the translated text is for wide circulation sis Grammar of Japanese in the Mu-Project. In Proceedings of or publication (for example, technical manuals), style and COLING 84, Stanford University, California: 338-343.
The following addendum is in rough form as it arrived just as the issue was going to press. Recent achievement of the translation evaluation (as of September 1985)
,-.t,ib.it~ 0 1 2 3 4 1 256 1 34 13 6 6 2 0 0 318 .~4.5~1 2 1 274 36 31 26 11 1 0 1 381 ~.9.31 3 0 0 157 87 79 26 29 0 6 384 Z9.51 4 0 0 19 39 60 16 24 1 3 162 L2.5~ 5 0 0 2 3 18 2 10 0 0 35 Z.?Z ? 1 1 18 2O L.5Z • .,-.~ 257 275 248 174 190 61 66 1 28 1300 "'"'".t,.0,~,~ 19.SZ21.2119.11I3.4~t!I4.614.7Z i.lg ).Ig .~.~
106 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
APPENDIX A TRANSLATIONS AND THEIR INTELLIGIBILITY EVALUATION RESULTS
Intelligibility = l
NO. 535:E82060091 3 1 ~~i~v~h=u 7 h I~±~ ~, ~I~-~.~, ~©~S~~(:~ b~o
Distributions of size, charged particle density and electric fields of drift currents during the high frequency non-seLf-sustaining discharge were studied numericaLLy.
NO. 87:E82060092 4 1 ~ b~l~,~ = ~ ~~v)~,=tz~i7 " ~ ~ ~I~ b'd;@~ b, ~f~l~,~-d~ < o
Conditions for realizing stabilized simuLtaneousLy-operated arcs are analyzed as non-baLanced bridge circuits, and the solution is made by using computers.
NO. 97:E82060095 5 1 ~ ~: c~=c ~" ~'-~¢~i~ ~~-F-~~I: • -~~ < o
Mass, charges, momentum and energy preservation formulas are solved by the iteratiue finite differential method under appropriate boundary conditions.
Intelligibility = 2
NO. 81:E82060020 3 1
CompLex amplitudes and equivalent surface current distribution of the surface waves formed by the diffraction of the plane waves emitted to dielectric plates in vacua are found by numerical solutions of integral equations. NO. 86:E82060020 6
As the plane wave arrival direction is close to tangents, the nonuniform component amplitude of solutions in the neighborhood of the irregular section is increased, and the number of contributions of the scattering by edges increases.
NO. 87:E82060021 3 1
MuLtipLier photo tube current spectrum root-mean-square values have the periodicity when freezing hyootheses ~re established in the intensity fluctuation.
Intelligibility = 3
NO. 102:E82060026 3 1
SpecuLar points of the conductor surfaces when wave sources and observation points are given have to be known when geometrical optics theories are applied to the caLcuLation of reflected waves.
N3. 107:E82060025 4 1 f~ ~ 7 cg~l~-~-~ ~- v H 0 0 cg~J~ < (=~ -'< 9 h s~ b~:l~,~-~ h ~ ~ b ~ 6 ~
It WaS done that the motion of charged particles accompanies the diffractive radiation bursts which frequency spectra concentrated in the neighborhood of frequencies of quasi-static resonance mode H O0 be obvious.
go. 108:E82060025 5 I , . B ~ ~ s~ °- t~t~,~ ~ ~ h L~<~ ~ ~±~-7-1~~o~--~1~t ~o
As ~or the radiation energy it is maximum when particles fly resonator coupled slots, and the nonuniformity of the particle intrinsic electromagnetic field is reflected.
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 107 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nak~mura Japanese Government Project for MT
Intelligibility = 4
NO. 62:E82060011 3 I
~ as for causes of generation of charges in every place of the Living environment and the Industrial environment, and forms, conductivity, dielectric constants, humidity of surfaces, and so on are important as weLL as materials and movement.
NO. 47:E82060012 S 1
There were many cases in which the parts exist which ~ in real machine be axt-symmetric but correction factors were introduced for the modeLin9 of th~s case, and it was incorporated in nonlinear iterative computations.
NO. SO: E82060013 ; I
The equiamptitude distribution of the spatial composite electric fields and magnetic fields to intervals of the conductin 9 cylinders and radii of the conducting cylinder in front rear scattering cross section and the neighborhood of the cylinder was shown, and characteristics of scattering waves were done be obvious ~hen the emission was made from sides and array directions.
Intelligibility = 5
NO. 352:E82060195 S 1
The various insertion machine tiLL automatic forming machines is introduced by the manual operation.
NO. 366:E82060200 3 1 7 ° ~ ~ ~ ~ ~ ~[]~t~L~/~t~, ~-~'-.~- v~i~,~l~l:~J~ ~o
Zntegrated circuits proceed in the printed wiring itself, and the wiring of the mother board reverse is mainly the wire wrapping.
NO. 367:E82060200 4 I
Ratios of the wiring by automatic iachinea increased, and they became more than two twice of the manual operation.
APPENDIX B TRANSLATIONS AND THEIR ACCURACY EVALUATION RESULTS
Accuracy = 0
NO. 70:E82060017 3 1
The reflection of space loduLated waves fro., the plane nonhomogeneous medium with smaLL-scaLe heterogeneity is studied.
NO. 73:E82060017 7 1
RefLection characteristics of periodicaLLy modulated signals fro., media of random nonhomogeneoua Layers are studied.
NO. 7k: E82060018 2 1
Mean values of ordered products of two functiona[s.
108 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
Accuracy = 1
NO. 77:E82060018 6 1 ~:©;~,:lZ~j~ - ~ ~, ~ -~ :/h"d~, ~1~O a u s s~~l~l~'~ ~o
High-order averages of responses are expressed by the ordinal cumulant, and the Gauss characteristic of physical systems can be utilized to the utmost.
80. 78:E82060019 2 1
Uniform asymptotic characteristics of electromagnetic fields in half shadows of transversal waves.
NO. 79:E82060019 ~ 1
Uniform asymptotic equations of the fields established in arbitrary distances from the half shadow region of transversal waves by the analysis of asymptotic characteristics of correct answers of the reflection field potential were obtained.
Accuracy = 2
80 172:E82060044 2 1
The analysis of accelerated charged fields to bradyohic and tachyonic parts.
NO. 176:E820600~5_3_1
Tt is important from research and several applications of dynamics of electron beams to know the angle distribution of electrons in circular accelerators.
80. 182:E82060046 5 1
The current density under the cutoff is given by functions of voltages and gap distances, and the current density under the cutoff contradicts experiments.
Accuracy = 3
NO. 271:E82060070 4_1
Realization conditions include iatrices, and basic differences between single-variable networks are stressed muttivariable.
NO. 30:E82060071 4 1
Several examples of the logical composition are given after describing outlines such as methods for connection of the slave side and requests for slaves.
NO. 36:E82060072 5 1
Next, production and markettn9 situations of the future expansibility and system elements are described after discussing diagnostic functions.
Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985 109 Makoto Nagao, Jun-ichi Tsujii, Jun-ichi Nakamura Japanese Government Project for MT
Accuracy = 4
NO. 31:E82060071 5 1
The ~J~ address togtcat circuits and data control circuits executing the stave logic next are shown.
NO. 44:E82060075 k I
The error correction is carried out by processing the measuring terminal output digitally, and measurement examples such as reflection coefficients of cable connectors and losses at the time of the mismatching are described, and the accuracy is also refered to.
NO. 57:E82060079 6 ~.
Related equations were based on the integral type of taws of electromagnetic field as methods by using finite differential methods.
Accuracy = 5
NO. 203:E82060051 8 1
The asymptotic stability in the regions by which there is an unique self-excited vibration network of the DC solution and for giving the unique self-excited vibration network of the DC solution can be discussed.
NO.~l~ 213:E82060054 ~<~b-~-~,8~ 7 1 ~ ~n~--h~~ ~]~0~ ~~t~
This condition becomes the expansion of passivity and tosstessness conditions of time region and frequency region of Linear and time-invariable welt known concentrated constant n-port networks to nonlinear cases.
NO. 222:E82060056 6 1
Electric circuits are ~i~-~ ed by using system equations mainly in this method.
Accuracy = 6
NO. 630:E82060176 S 1
It becomes the length of At tubes o~ more than 100 km in wholes, and junctions and installation methods are shown, and introduced as application examples utilizing features of the At.
NO. 309:E82060182 3 1
Unlike conventional interactive systems, pin positions are digitized be manual, and they do not have to be inputted.
NO. 310:E82060182 4 1 ~,~, ~y~, ~,P~I:~8~~N~(=~8o
Pin dimensions by pin positions, number of pins and current capacity are optimized automatically.
NO. 326:E82060188 3 1
The soldering forms above-mentioned compounds in interfaces.
110 Computational Linguistics, Volume 11, Numbers 2-3, April-September 1985