Grammar Writing System (GRADE) of Mu-Machtne Translation Project and its Characteristics Jun-tcht NAKAMURA. Jun-tcht TSUJII. Makoto NAGAO Department of Electrical Engineering Kyoto University Sakyo. Kyoto. Japan ABSTRACT A powerful grammar writing system has been developed. Thts grammar wrtttng system ts called GRADE (GRAmmar DEscriber). GRADE allows a grammar writer to write grammars Including analysts, transfer, and generation using the same expression. GRADE has powerful grammar writing facility. GRADE allows a grammar writer to control the process of a machine translation. GRADE also has a function to use grammatical rules written tn a word dictionary. GRADE has been used for more than a year as the software of the machine translation project from Japanese Into Engltsh. which ts supported by the Japanese Government and called Nu-proJect. 1. Objectives Vhen we develop a machine translation system, the intention of a grammar writer should be accurately stated tn the form of grammatical rules. Otherwise, a good grammar system cannot be achieved. A programming language to write a grammar, which ts composed of a grammar writing language, and a software system to execute tt. ts necessary for the development of a machine translation system (Bottet 82). If a grammar writing language for a machine translation system is to have a powerful writing facility, tt must fulfill the following needs. A grammar wrttlng language must be able to manipulate linguistic characteristics tn Japanese and other languages. The 11ngulstlc structure of Jcpanese tS largely different from that of Engltsh, for instance. Japanese does not restrict the word order strongly, and allows the omission of some syntactic components. Vhen a machine translation system translates sentences between Japanese and English, a grammar writer must be able to express such characteristics. A grammar writing language should have a framework to write grammars tn analysis, transfer, and generation phase using the same expression. It Is undeslrable for the grammar writer to learn several different expressions for different stages of a machine translation. There are many word specific linguistic phenomena tn a natural language. A grammar writer must be able to add word specific rules to a machine translation system one after another to deal wtth word specific linguistic phenomena, and improve hts machine translation system over a long period. Therefore. a grammar writing language must be able to handle grammatical rules written tn word dictionaries. There ts a natural sequence tn a translation process. For example, a parstng of noun phrases which do not contain sententtal forms is executed before a parsing of more complex noun phrases. An approximate parsing of compound sentences is executed before a parsing of complex sentences. Also. when an application sequence of grammatical rules are written explicitly, a grammar writing system can execute the rules efficiently. because the system Just needs to test the applicability of a restricted number of grammatical rules. So. a grammar writing language must be able to express several phases of a translation process in the expression explicitly. A grammar writing language must be able to treat the syntactic and semantic ambiguities tn natural languages. But tt must have some mechanisms to avoid a combinatorial explosion. Keeping these points in mind, we developed a new programming system, which ts composed of the grammar writing language and its executing system. Ve wtll call it GRADE (Grammar Describer). 2. Expression of the data for a processing The form of data to express the structure of a sentence during an analysis, a transfer, and a generation process has a strong effect on the framework of a grammar wrtttng language. GRADE uses an annotated tree structure for expressing a sentence. Grammatical rules tn GRADE are described tn the form of tree-to-tree transformation wtth annotation to each node. 338 The annotated tree tn GRADE ts a tree structure whose nodes have ltsts of property names and their values. Figure 1 shows an example of the annotated tree. ~ -CAT - S~ -NUMBER - SINGULAI LE-NUMBER - -SEM = HUMAN E-CAT : Engllsh Category Symbol E-NUMBER: English Number (SINGULAR or PLURAL) E-SEM :Engltsh Semantic Marker Ftgure 1 An example of the annotated tree tn GRADE The annotated tree can express a lot of Information such as syntactic category, number. semantic marker, and other thtngs. The annotated tree can also express a flag tn tts node. whlch ts stmtlar to a flag tn a conventional programming language, to control the process of a translation. For example, in a grammar of a generation, a grammatical rule ts applled to all nodes tn the annotated tree, whose processtngs are not finished. In such a case, a grammatical rule checks the DONE flag whether ttts processed or not. end sets T to the newly processed ones. 3. Rewriting Rule tn GRADE The bastc component of a grammar wrtttng language is a rewriting rule. The rewriting rule In GRADE transforms one annotated tree tnto anoti~er annotated tree. The rewriting rule can be used In the grammars of analysts, transfer and generation phase in a machtne translation system, because the tree-to-tree transformation by thts rewriting rule ts very powerful. A rewriting rule tn GRADE conststs of a declaration part and amatn part. The declaration part has the following four components. (1) Directory Entry part, whtch contains a grammar writer's name, a verston number of the rewrttin 9 rule, and the last date of the revision. Thts part ts not used at the execution ttme of the rewriting rule. A grammar wrtter ts able to see the information by ustng the help factltty of the GRADE system. (2) Property Definition part, where a grammar writer declares the property names and thetr values. (3) Vartable Intt. part, where a grammar wrtter declares the names of variables. (4) Matchtng Instruction part, where a grammar wrtter specifies the mode to apply the rewriting rule to an annotated tree. The matn pant specifies the transformation tn the rewriting rule. and has the following three parts. (1) Matchtng Condition part. where the condition of a structure and the property values of an annotated tree ts described. (2) Substructure Operation part, whtch specifies operations for the annotated tree that has matched wtth the condition wrttten tn the matching condition pant. (3) Creatton part, whtch spec|ftes the structure and the property values of the transformed annotated tree. 3.1. Matching Condition part The matchtng condition part specifies the condition of the structure and the property values of the annotated tree. The matchtng condition part allows a grammar writer to spectfy not only a rtgtd structure of the annotated tree, but also structures whtch may repeat several ttmes, structures which may be omttted, and structures tn which the order of thetn sub-structures ts not restricted. For example, the structure tn whtch adjectives (ADJ) repeat arbitrary ttmes and a noun (N) follows them tn Engllsh ts expressed as follows. ADJ ADJ N > matching_condition: • (ADJS N): AOJS: anyC~(ADJ)): The structure 11ke a combination of a verb (V) and an adverbial parttcle (ADVPART) tn thts sequence wtth or without a pronoun (PRON) tn between tn Engltsh tswrttten as follows. V (PRON) ADVPART > matching_condition: • (V PRON ADVPART): PRON: optional: Atyptcal Japanese sententtal structure tn whtch three adverbial phrases (ADVP). each composed of a noun phrase (NP) and a case particle (GA, WO. or NI) proceed an verb (V) tn no particular order ts expressed as follows. matching_condition; ~(A1 A2 A3 Y); A1. A3: disorder; ADVP1 ADVP2 ADVP3 V > Al: ~((ADVP1NP1GA)): A A A A2:zCCADVP2 NP2 WO)): NPl GA NP2 we NPa NZ A3: zCCAOVPa Ne3 .X)): The matchtng condition part allows a grammar wrtter to spectfy conditions about property names and property values for the nodes of the annotated tree. A grammar wrtter can compare not only a property value of a node wttha constant value, but also values between two nodes tn a tree. 339 For example, the number agreement between a subject noun and a verb Is written as follows. matching_condition: ~(NP UP): NP.NUNBER " VP.NUNBE~; 3.2. Substructure Operation part The substructure operation part spec'tftes operations for the annotated tree which has matched wtth the matchtng condition part. The substructure operation part allows a grammar writer to set a property value to a node. and to assign a tree or a property value to a variable, whtch is declared tn the variable tntt. part. It also allows htm to call a subgnammar, a subgnammar network, a dictlonary rule. a bullt-ln functlon, and a LISP function. The subgrammar, the subgramman network. the dicttonany rule, and the butlt-tn function w111 be discussed tn sectton 4 5., and 6. In addition, a grammar wntter can write a conditional operation by using the IF-THEN-ELSE form. An operaLion to set 'A' to the lextcal untt of the determiner node (DET.LEX). tf the number of the NP node |S SINGULAR. Is wrttten as follows. substnuctune_operatton: tf NP.NUMBER - 'SINGULAR': then DET.LEX <- "A': else DET.LEX <- "NIL'; end_tf: Transformation of matn part tn a newntttng rule: A A /b,, I B C O > E /t,,, B C D Transformation of a whole annotated tree: A A ABCD > A E /t,, I /b,, BCD E BCD /t,,, BCD Figure 2 An example of an application of the main part The matching tnstnuctton pant specifies the travense path of the annotated tree. There are four types of the traverse pathes, whtch are the combinations of <left-to-right or night-to-left> and <bottom-to-top on top-to-bottom>. When a grammar writer specifies left-to-right and bottom-to-top mode, the annotated tree w111 be traversed as follows. 5 3 / 3.3. Creation part The structure and the property values of the transformed annotated tree ts written tn the creation part. The transformed tree ts described by node names such as NP and VP, whtch are used in the matchtng condition part on the substructure operation part. A cneatton part to create the tree whose top node ts S and whtch has a NP sub-tree and a VP sub-tree ts wnttten as follows. creation: Z((S NP VP)): 3.4. Matching Instruction part The maln part of a rewrltlng rule In GRADE (the matching condition part, the substructure operation part. and the creatlon part) can be applied not only to a whole tree, but also to sub-trees. Figure 2 shows an example of the application of a maln part. 4. Control of the grammatical rule applications A grammar writing language must be able to express detailed phases of a translation process tn the expression expltctt]y. GRADE allows a grammar writer to divide a whole grammar into several parts. Each part of the grammar ts called a subgnammar. A subgrammar may correspond to a grammatical unit such as the parstng of a stmple noun phrase and the partng of a compound sentence. A whole grammar ts then described by a network of subgrammars. Thts network ts called a subgnammar network. A subgrammar network allows a grammar writer to control the process of a translation tn detatl. When • subgrammar network tn the analysts phase consists of a subgrammar for a noun-phrase (SG1) and a subgrammar for a verb-phrase (SG2) tn this sequence, the executor of GRADE first appltes SG1 to an input sentence, then appltes SG2 to the result of an application of SG1. 4.1. Subgrammar A subgrammar conststs of a set of rewriting rules. Rewriting rules tn a subgrammar have a prtontty ondertng tn their application. The n-th 340 rewriting rule tn a subgrammar tstrted before the (n+l)-th rule. A grammar wrtter can spectfy four types of application sequence of rewriting rules tn a subgrammar. Let us assume the situation that a set or rewriting rules tn the subgrammar ts composed of RR1. RR2 and RRn. that RR1 and RR|-I cannot be applied to an tnput tree. and that RRt can be applted to tt. When a grammar wrtter specifies the ftrst type. whtch ts called ORDER(I). the effect of the subgrammar execution ts the application of RRt to the tnput tree. When a grammar wrtter specifies the second type. which |s called ORDER(2). the executor of GRADE trtes to apply RRt+I RRn to the result of the application of RRt. So. ORDER(2) means that rewriting rui~s tn the subgrammsr are sequentially applted to an tnput tree. The thtrd and fourth type. whtch are called ORDER(3) and ORDER(4). are the Iteration type of ORDER(l) end ORDER(2) respectively. So, the executor of GRADE trtes to apply rewriting rules untt1 no rewriting rule Is applicable to the annotated tree. SEARCH-CANDIDATE-OF-HOUNS.sg: sg_mode: order(Z): rr_tn_sg: CANDIDATE-OF-NOUNS-t: UP-NP-TO-PNP: CANDIDATE-OF-NOUNS-Z; end_sg.SEARCH-CANDIDATE-OF-NOUNS: Ftgure 3 An example of a subgrammar Ftgure 3 shows an example of a subgrammsr. When thts subgrammar is applted to an annotated tree. the executor of GRADE ftrst trtes to apply the rewriting rule CANDIDATE-OF-NOUNS-1 to the tnput tree. If the appl|catton of thts rule succeeds, the tnput tree ts transformed to the result of the application of the rewriting rule CANDIDATE-OF-NOUNS-1. Otherwise. the tnput tree ts not modified. In etther case. the executor of GRADE next tr|es to apply the rewrtt|ng rule UP-NP-TO-PNP to the tnput tree. The executor continues such a process untt1 the application of the last rewriting rule CANDIDATE-OF-NOUNS-2 ts finished. 4.2. Subgramar Network A subgrammar network descr|bes the application sequence of subgrauars. The specification of a subgrammar network conststs of the following ftve parts. (1) Directory Entry part. whtch ts as the same as the one tn a rewriting rule. (2) Property Definition part. whtch Is the same as the one tn a rewriting rule. This part ts used as the default declaration tn rewriting rules. (3) Vsrtable ]ntt. part. which ts the same as the one tn a rewriting rule. The variables are used to control the transition of the subgrammar network. The variables are referred to and asstgned tn the substructure operation part of the rewriting rule. The variables are also referred tne 11nk specification part. whtch wtll be described later. (4) Entry part. whtch specifies a start node of the network. (5) Network part. whtch specifies a network of subgrammars, The network part spec|f]es the network structure of subgrammars, and conststs of node specifications and 11nk spectftcat|ons. The node specification has a label and a subgrammer or s subgnammar network name. whlch ts called when the node gets the control of the processing. The 11nk specification specifies the transit|on among nodes tn a subgramman network. The 11nk specification checks the value of a verteble whtch |s set tn • rewriting rule. and dectdes the label of a node whtch wtll be processed next. PRE.sgn; directory_entry: owner(J.NAKAHURA): verston(VO2L05): last_update(83/12/25): var_tntt; OPRE-FLAG tntt(T): entry: START: network: START: PRE-STEP-|osg; LOOP : PRE-STEP-2.sg; A: PRE-STEP-3.sg: B: PRE-END-CHECK.sg: |f OPRE-FLAG: then goto LOOP: else goto LAST: LAST: PRE-STEP-4.s9: extt: end_sgn.PRE; Ftgure 4 An example of a subgrammar network. Ftgure 4 shows an example of a subgrammar network. When the executor of GRADE appltes thts subgranunar network to an tnput tree. the executor checks the var-tntt part. then puts a new vartable OPRE-FLAG on a stack, and sets T to OPRE-FLAG as an tntttal value. After that. the executor checks the entry part and find the label of the start node START tn the network. Then the executor searches the node START and applles the subgrammar PRE-STEP-1 to the tnput tree. After the application, the executor appltes the subgrammer PRE-STEP-2 (node name: LOOP) and PRE-STEP-3 (node name: A) to the annotated tree tn thts sequence. Next. the executor applles the subgrammar PRE-END-CHECK (node name: B) to the tree. 341 Rewriting rules in PRE-END-CHECK examine the tree and set T or NIL to the variable ePRE-FLAG. The executor checks the link spectf|catJon part, which is started by IF. and examines the value of the variable @PRE-FLAG. The node in the network which will be activated next is the node LOOP if @PRE-FLAG is not NZL, otherwlse, the node LAST. Thus, while @FRE-FLAG ts not NIL, the executor repeats the applications of three subgrammars, PRE-STEP-2. PRE-STEP-3. and PRE-END-CHECK. to the annotated tree. When @PRE-FLAG becomes NIL. the subgrammar PRE-STEP-4 tn the node LAST ts applted to the tree. and the application of thls subgrammar network PRE Is terminated. 5. Handling the grannaatlcal rule tn the word dictionaries GRADE allows a grammar wrtter to write word specific grammatical rules as a subgramman In an entry of word dictionaries of a machine translation system. A subgramman written in a dictionary entry is called a dictionary rule. The dictionary rule is specific to a particular word In the dictionary. The dictionary rule is retrieved wttha entry word and a rule identifier as the key. and is applied to the annotated tree which is specified by a grammar writer, when CALL-DIC operation In the substructure operation part Is executed. Figure 5 shows an example of a rewriting rule which calls a dictionary rule. In thts case. a dictionary rule which ts written in an entry of a word as indicated by V.LEX (the value of the lextcal untt of verb). and whose name ts ANALYSIS. ts epplted to the sequence of NP1. V. NP2. and PP (noun phrase 1. verb phrase, noun phrase 2. and prepositional phrase). Then the result of the application of the dictionary rule Is assigned to the vartable aS. CASE-FRANE.rr: var_tntt: aS; matching_condition: Z(NPZ v Me2 PP): substructure_operation: @S <- ca11-dtc(V.LEX ANALYSIS Z(NP1V NP2 PP)): creation: ~(es): end_Pr.CASE-FRAME: Ftgure S An example of a rewriting rule which calls a dictionary rule 6. Treatment of Ambiguities A grammar wrtttng language must be able to treat the syntactic end semantic ambiguities in natural languages. GRADE allows a grammar writer to collect all the result of possible tree-to-tree transformations by a subgrammar. However, It must avoid a combinatorial explosion, when tt encounters the ambiguities. For instance, let us assume that a grammar writer writes a subgramman which contains two rewriting rules to analyze the case frame of • verb, that a rewriting rules ts the rule to construct VP (verb phrase) from V and UP (a verb and a noun phrase), and that the other ts the rule to construct VP (verb phrase) from V. NP and PP (a verb. a noun phrase, and a prepositional phrase). When he specifies NONDETERMINISTIC_PARALLELED mode to the subgremmar, the executor of GRADE 8ppltes both rewriting rules to an Input tree, constructs two transformed trees, and merges them tnto 8 new tree whose top node has 8 spectal property PARA. The top node of this structure is called a pare special node. whose sub-trees are the transformed trees by the rewriting rules. Figure 6 shows an example of thts mode and apara node. '7 V NP PP SG PARA VP PP VP A A",, V NP V NP PP Figure 6 An example of a pars speclal node A grammar writer can select the most appropriate one from the sub-trees under a pare special node. A grammar writer ts able to use built-in functlons. MAP-SG. MAP-SGN. SORT. CUT. and INJECTION in the substructure operation part to choose the most appnoprlate one. Figure 7 shows an example to use these bullt-Jn functions. substructure_operation: eX <= ca11-dtc(V.LEX CASE-FRAME Z(N NP PP)): eX <- ca11-butlt(map-sg ~(gX) tree EVALUATE-CASE-FRAME): @X <- call-built(sort Z(@X) tree SCORE): @X <- cell-built(cut [(eX) tree 1): 9X <- call-built(Injection ~(eX) tree 1): Figure 7 An example of bullt-ln functions In this substructure operation part. the executor of GRADE appltes the dictionary rule wrttten tn a word which ts the value of V.LEX (lexlcal untt of verb) to the tree. and sets the result to the vartable eX. When the nondetermtnisttc-paralleled mode ts used tn the dictionary rule. the value of eX ts the tree whose root node tsa pare spectel node. After that, the executor calls butlt-tn functton MAP-SG to apply 342 the subgrammar EVALUATE-CASE-FRAME to each sub-tree of the value of OK. and sets the result to eX again. The subgrammar EVALUATE-CASE-FRAME computes the evaluation score end sets the score to the value of the property SCORE tn the root node of the sub-trees. Next, the executor calls butlt-tn functton SORT. CUT. and INJECTION to get the sub-tree whose score Is the highest one among the sub-trees under the pare spectal node. This tree ts then set to 9X as the most appropriate result of the dictionary ru]e. The para spectal node ts treated as the same as the other nodes tn the current Implementation of GRADE. A grammar wrtter can use the para node as he want, and can select a sub-tree under a pare node at the later grammatical rule application. 7. System configuration end the environment The system configuration of GRADE ts Shown tn Figure 8. Grammatical rules written tn GRADE are first translated tnto tnternal forms, which are expressed by s-expressions tn LISP. This translation ts performed by GRADE translator. The Internal forms of grammatical rules are applted to an tnput tree. which ts an output of the morphological analysts program. Thts rule application Is performed by GRADE executor. The result of rule applications |s sent to the morphological generat4on program. Dictionary Grammar f J GRADE translator 1/ \ Dictionary Grammar (Internal form) rule ~ ~ r~ tnput_~ GRADE ~output sententtal tree|executor J sententtal tree Ftgure 8 The system configuration of GRADE GRADE system ts mrttten tn UTILISP (University of Tokyo Interactive LISP) and Implemented on FACON M382 wtth the additional functton of handllng Chatnese characters. The system ts also usable on Ltsp Machtne Symbollcs 3600. The program stze of GRADE system ts about 10.000 ltnes. the form of tree-to-tree transformation rtth annotation to each node. (2) Rewriting rule has • powerful wrtttng facility. (3) Grammar can be divided Into several parts and can be 11nked together as a subgrammar network. (4) Subgrammar can be written tn the dictionary entrtes to express word spectftc linguiStiC phenomena. (5) Spectel node ts provtded tn a tree for embedding ambiguities. GRADE has been used for more than a year as the software of the nattonal machtne translation project from Japanese Into English. The effectiveness of GRADE has been demonstrated tn thts project. The linguistic parts of the project such as the morphological analysts/generation programs, the grammars for the analysts of Japanese. the transfer from Japanese Into Engltsh and the generation of Engllsh. are discussed tn other papers (Sakamoto 84) (TsuJt1 84) (Raged 84). Thts study: "Research on the machtne translation system (Japanese-English) of scientific and technological documents" Is betng performed through Spectal Coordination Funds for Promoting Science & Technology of the Science and Technology Agency of the Japanese Government. ACKNOWLEDGEMENTS Ve would 11ke to acknowlege the contribution of N. Kogt. F. Ntshtno. Y. Sakane. M. Kobayasht. S. Sate. and Y. Senda. who programmed much of the system. We mould also 11ke to thank the other member of Me-project for their useful comments. REFERENCES Bottet. Ch., et el. Implementation and Conversational Environment of ARIANE 78.4. Proc. COLING82. 1982. RageD, M., et el, Dealtng wtth Incompleteness of Linguistic Kno~ledego on Language Translation. Proc. COLING84o ;964. Sakamoto, Y et al, Lextcon Features for Japanese Syntactic Analysts In Mu-ProJect-JE, Proc. COLING84, 1984. TsuJtt, J., et el, Analysts Grammar or Japanese tn Hu-ProJect, Proc. COLING84, ;984. 8. Conclusion The grammar wrtttng system GRADE ts discussed 4n thts paper. GRADE has the follow4ng featureS. (I) Rewriting rule ts an expression tn 343 . Grammar Writing System (GRADE) of Mu-Machtne Translation Project and its Characteristics Jun-tcht NAKAMURA. Jun-tcht TSUJII. Makoto NAGAO Department of Electrical. which ts composed of a grammar writing language, and a software system to execute tt. ts necessary for the development of a machine translation system