File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/94/c94-1015_metho.xml
Size: 24,523 bytes
Last Modified: 2025-10-06 14:13:35
<?xml version="1.0" standalone="yes"?> <Paper uid="C94-1015"> <Title>Constituent lloundary Parsing for Exanll)lo-lkised Maclhine Tr,'inslation</Title> <Section position="4" start_page="705" end_page="705" type="metho"> <SectionTitle> 2 Pattern </SectionTitle> <Paragraph position="0"> A pattern represents meaningful units for linguistic structure and transfer in TDMT, and is defined as a sequence that consists of variables and synrbols representing constituent boundaries. A variable corresponds to some linguistic constituent, and a constituent boundary does not allow any two variables to be adjacent. A constituent boundary is expressed by either a functional word or a part-of-speech bigram marker l The explanations in this anti the subsequent two sections, use English sentence parsing.</Paragraph> <Section position="1" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 2.1 Part-of-speech </SectionTitle> <Paragraph position="0"> Table 1 shows tile English parts-of-speech, currently used in our English-to-Japanese TDMT system. This part-of-speech system does not necessarily agree with that of conventional grammar.</Paragraph> <Paragraph position="1"> Table 1 English parts-of-speech ~of-speech abbreviation example adjective adj large adverb adv exactly interjection i nterj oh common noun noun bus numeral num eleven proper noun propn Kyoto pronotm pron I wh-word wh what verb verb go be-verb be is auxiliary verb aux crm preposition prep ca conjunction co nj bta determiner det the suffix suffix a.m.</Paragraph> <Paragraph position="2"> In this part-of-st)eech system, a be-verb, auxiliary verb, preposition, conjtmction, deterntiner, and suffix, are classified into a functional word.</Paragraph> </Section> <Section position="2" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 2.2 Constituent I)()ulldary marke,&quot; exl)ressed </SectionTitle> <Paragraph position="0"> by a functional word One problem with pattern descriptions using surface 1 In this paper, variables, actual words, and part-of-speech abbreviations are expressed in calfital letters, italics, and gothic, respectively.</Paragraph> <Paragraph position="1"> words is the necessity for a large number Of patterns. To snppress the nnnecessary patterns, the surface words in patterns are in principle restricted to functional words, which occur frequently, and which modify or relate content words 2.</Paragraph> <Paragraph position="2"> Fnr instance, the expression, &quot;go to the station&quot; is divided into two constituents &quot;go&quot; and &quot;the station&quot;, and the l)reposition, &quot;to&quot; can be identified as a constituent boundary. Therefore, in parsing &quot;go to the station&quot;, we use tile l)attem, &quot;X to Y &quot;, which has two variables X and Y, and a constituent boundary, &quot;to.&quot;</Paragraph> </Section> <Section position="3" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 2.3 Constituent I)oundary marker expressed </SectionTitle> <Paragraph position="0"> by a pa,'t-nf-sl)eech hig,'anl The expression &quot;1 go&quot; can be divided into two constituents 'T' and &quot;go.&quot; But it has no surface word that divides tile expression into two constituents. In this case, a part-of-speech bigr,'un is used as a constituent boundary.</Paragraph> <Paragraph position="1"> Suppose th,qt a constituent X is immediately followed by a constituent Y. We express a boundary-marker between X and Y by A-B, where A is a part-of-speech abbreviation of X's last word, and B is a 1)art-of-speech abbreviation of Y's first word. For instance, 'T' and &quot;go&quot; are a pronoun and a verb, respectively, so the marker &quot;pron-verb&quot; is inserted as abot, ndary marker into &quot;1 go&quot;. Namely, &quot;I pron-verb go&quot;, i.e. with the boundary marker inserted into the original input, matches tile pattern &quot;X pron-verb Y.&quot;</Paragraph> </Section> <Section position="4" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 2.4 Linguistic level </SectionTitle> <Paragraph position="0"> Patterns are classified into (lffferent linguistic levels to limit the explosion of structural ambiguity during parsing. Table 2 shows typical linguistic levels in F.nglish patterns.</Paragraph> <Paragraph position="1"> Table 2 Typical levels in English patterns level exan_!p_le beginning phrase excuse me but X conlpotlnd sentence X when V simple sentence I would like to X verl) phrase X at Y noun phrase XofY, XatY c()mpound word X o'clock 2 Exceptions are canned expressions such as '7 would like to&quot; and &quot;in front of', or frdquent content words such as &quot;what.&quot; In Table 2, beginning phrase is the highest level, and compound word is the lowest. A variable on a given level is instantiated by a string described on that same level or on a lower level. For instance, in the noun phrase &quot;X of Y &quot;, the variables, X and Y cannot be instantiated by a simple sentence.</Paragraph> </Section> </Section> <Section position="5" start_page="705" end_page="705" type="metho"> <SectionTitle> 3 Derivation of Possible Structures </SectionTitle> <Paragraph position="0"> The algorithnl for constituertt l)oundary parsing is as follows; (A) Assignment of morphological inRn'nmtion to each woM of an input string (B) Insertion of constituent boundary nmrkcrs (C) Derivation of possible structures by top-down pattern matching (D) Structural disambiguation by distance calculation Note: we will explain (A), (B) and (C) in this section, and (D) in the next section, usirlg die following English sentence; (1) &quot;The bus leaves Kyoto at eleven a.m,&quot;</Paragraph> <Section position="1" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 3.1 Assignment of nlorphohlgical </SectionTitle> <Paragraph position="0"> int'ormathtn First, each word of the input string is assigned morphological information, such as its part-ol'-sl)eech and conjugation fc.rm. Through tiffs assignnient, we can get the lollowing part-of-speech sequence for (1). (2) dot, noun, verb, propn, prop, num, suffix hi addition, each word is also assigned a thesaurus code for distance calcnhltions ,'lnd ,'ill index for retrieving l)atterns. For instance, &quot;bits&quot; has a thesaurus code corresponding to tile semantic attribute 'car.' Moreover, from the word &quot;(it&quot;, we can obtain the index to the pattern &quot;X (at Y&quot;, whicll is found for both verb phrase and nOl.ln phrase.</Paragraph> <Paragraph position="1"> .3.2 Marker hiserlic, n A constituent boundary marker is inserted in an input string for pattern-matching. The marker is extracted \[rein the part-of-speech sequence of an input sentence. Since such bigrams as dot-noun belong to the same constituent, marker insertion by a part-of-sl)eech bigram is restricted according to the items below.</Paragraph> <Paragraph position="2"> (a) Neither A nor B is a part-of-speech relating two constituents, such as a preposition (b) A is not a l)art-of-speech nlodifying a latter constituent, such :.is a dotorinh/or.</Paragraph> <Paragraph position="3"> (c) B is not a l)art-of-sI)eech modifying a previous constituent, such as a suffix.</Paragraph> <Paragraph position="4"> We mainttlin a list of p:lrt-of-speech bigrams that are eligible as marke,'s because they satisfy the above conditions. Of the bigrams in (2), &quot;det-noun&quot;, &quot;propnprep&quot;, &quot;prop-nora&quot;, and &quot;nun>suffix&quot;, vioklte the above conditions, and are of course excluded. Thus, only &quot;noun-verb&quot; and &quot;verb-propn&quot; are inserted into sentence (1), as shown in (3).</Paragraph> <Paragraph position="5"> (3) &quot;The bus noun-verb leaves verb-propn Kyoto at eleven a.tn.&quot;</Paragraph> </Section> <Section position="2" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 3.3 al)l)liealhm of Ilaltel'ns </SectionTitle> <Paragraph position="0"> Our pattern-nlatchhlg nlethod parses an inpilt sentence in a top-down fashion. The highest level patterns of the input sentence :.ire applied first; then lmtterns at lower levels are applied. The application procedure is as follows.</Paragraph> <Paragraph position="1"> (I) Get indices to patterns from each woM of the sentence. With these indices, patterns are retrieved and chocked to determine if each of them can match tile sentence. Then exectlte (II).</Paragraph> <Paragraph position="2"> (ll)Try to apply the highest-level patterns first. If there is a pattern tlmt can be applied, execute (1II) with respect to the variable bindings. Otherwise, exectite (IV).</Paragraph> <Paragraph position="3"> (Ill)Try to apply surface words (content words registered in a dictionary). If lhe al)lflicalion succeeds, the application fo, that portion is finished successfully. ()thcrwise, execute (I1).</Paragraph> <Paragraph position="4"> (IV) If the pattern to be applied is at the lowest level, the api)lication fails. Otherwise, lower tile level of the patterns and execute (II).</Paragraph> <Paragraph position="5"> If pattern al~plication finishes successfully for all portions o\[&quot; an input sentence, one or more source strttctures are obtained: since there is a possibility that more ttmn one pattern can be apl)lied to an expression in step (II), structural ambiguity may occur. We seek all possible structures by breadth-first application, and select the most plausible structure by the total distance value (See Section 4.4).</Paragraph> <Paragraph position="6"> In step (I), indices to possible patterns :-ire obtained from several words and bigrams in the marker-inserted sentence (3), as shown in Table 3.</Paragraph> <Paragraph position="8"> After step (I) is finished, steps (II)-(IV) are repeated recursively. First, the highest level pattern of the input sentence is applied. This is &quot;X noun-verb Y &quot;, which is defined at the simple sentence level. Next, an attempt is made to apply patterns to the variable bindings &quot;the bus&quot; and &quot;leaves verb-propn Kyoto at eleven a.m.&quot;, which are bound to variables X and Y, respectively. To &quot;the bus&quot;, at compound word level p'tttern &quot;the X &quot; is applied first, and the surface word &quot;bus&quot; is applied to proso &quot;tile bus.&quot; Likewise, patterns and suri'aee words are appliecl Io tile remaining part, and tile al~plic:-nion is finished successfully.</Paragraph> <Paragraph position="9"> The pattern &quot;X at Y &quot; is found for both verb phrase and noun phrase. &quot;leaves verb-propn Kyoto at eleven a.m.&quot; thus has two possible structures, by the application of &quot;X at Y.&quot; &quot;X verb-propn Y &quot; at the verb phrase level and &quot;X a.m.&quot; at compotmd word level, are also applied. Fig. 1 is tile tree representation derived from the structure for sentence (1) where &quot;X at Y &quot; is a veal) phrase, while Fig. 2 is a tree representation derived from the slrnctllre in which &quot;X at Y &quot; is a noun phrase. A boldfilce denotes the head part in each pattent. This infer,nation is t, lilizcd for extracting an input for distance calculations (See section 4.3).</Paragraph> <Paragraph position="11"/> <Paragraph position="13"> Fig. 2 Struclure in which &quot;X at Y &quot; is a noun phrase tile thes:mrus, and varies from 0 to 1. Tim value 0 indicates that two semantic attributes belong to exactly the same category, and 1 indicates that they :-/re tmrclated.</Paragraph> <Paragraph position="14"> An expression consists of words. The distance between expressions is the sum of the (listance between words multiplied by each weight.</Paragraph> <Paragraph position="15"> The distance is calculated quickly bectutse of the simple mechanism employed. (Sumita, 1992) and (Furuse, 1992, 1994) give a clctailcd account of tile distance calculation mechanism we are aclopting.</Paragraph> </Section> </Section> <Section position="6" start_page="705" end_page="705" type="metho"> <SectionTitle> 4 Distance Calculatitm </SectionTitle> <Paragraph position="0"> In this ,ruction, a nlethod for structural disaml)iguation utilizing dist,'mce calculation, is described.</Paragraph> <Section position="1" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 4.1 Distance </SectionTitle> <Paragraph position="0"> The distance between two words is retluced to the distance between their respective sem;mtic attributes in a thesaurus. Words have associated thesaurus codes, which correspond to partietflar semantic attributes. The distance between the semantic attributes is determined according to the relationship of their positions in the hierarchy of</Paragraph> </Section> <Section position="2" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 4.2 Best-match by distance calcul:ltinn </SectionTitle> <Paragraph position="0"> The advantages of an example-based framework are mainly due to the distance calctdation, which achieves the bcst-malch operation between tile input and provided examples.</Paragraph> <Paragraph position="1"> In TDMT, translation is performed by applying stored empirical Iransl'er knowledge. In TDMT transfer knowledge, each source pattern has example words of variables and possible target patterns. The most * qppropriate target pattern is selected according to the calculated distance between, the input words and the example words. The English pattern &quot;X at Y &quot; at the verb phrase level, corresponds to several possible Japanese expressions, as shown in the folhlwing English-to-Japanese transfer knowledge: XatY => Y' de X' ((present, conference)..), Y' ni X' ((stay, hotel)..), Y' we X' ((look, it)..) The first possible target pattern is &quot; Y' de X' &quot;, with example set ((present, cotg'erenee)..). We will see that this target pattern is likely to be selected to the extent that the input variable bindings are semanticqlly similar to the example elements &quot;present&quot; and &quot;coati're|Ice.&quot; Within this pattern, X' is the target word correslx)nding to X, tile result of transfer. &quot;preset, l&quot; and &quot;con/~reaee&quot; are sample bindings for &quot; X at Y &quot;, where X = &quot;present&quot;, and Y = &quot;conference&quot;. The al)ove transfer knowledge is compiled from such translation examples as the source-target pair of &quot; presem a paper at the conference&quot; and &quot;kaigi de ronbun wo happ),ou-st~ru&quot;, where &quot;kaigi&quot; means &quot;conference&quot; and &quot;happyou-sltru&quot; means &quot;present&quot;.</Paragraph> <Paragraph position="2"> Tilt semantic distance from the input is calculated for all examples. Then lhe example with the least distance from the input is chosen, and the target expresskm of that example is extracted. If the input is closest to (stay, hotel), &quot;Y' ni X' &quot; is chosen as the target express ion.</Paragraph> <Paragraph position="3"> The enrichment of examples increases tile aCc,lracy Of determining the target expression and structure because conditions become more dclailed.</Paragraph> </Section> <Section position="3" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 4.3 lnl)ut of' distance calculation </SectionTitle> <Paragraph position="0"> An input for distance ealcuh.ltion consists of head words in variable parts. In &quot;X at Y &quot; for the structure in Fig. l, X and Y are substitumd \[or the compound expressions, &quot;leaves verb-propn Kyoto&quot; a1~d &quot;eleven a.m.&quot;, respectively. In such eases, it is necessary to extract head words as the input for the disEmce calculation about &quot;X at Y &quot;.</Paragraph> <Paragraph position="1"> In order to get head words, tile head part is (lcsignawd in each pattern (boldface in Figs. 1 and 2). For inslance, the t)attern &quot;X vorb-propn Y II e(li)t;lillg the information that X is a head part. So the head of &quot;leaves verb-propn Kyoto&quot; is &quot;leaves&quot;, and tile head or &quot;x a.m.&quot; is &quot;a.m.&quot;. Thus, in &quot;X at Y &quot; for Ihe strncture in Fig. 1, the ini)ut of the distance calculation is (leaves, a.m.).</Paragraph> <Paragraph position="2"> Table 4 shows tile result of distance cqlculation in &quot;X at Y &quot; in Fig. 1. The most plausible target structure &quot;Y' ni X' &quot; and its distance value 0.17 are obtained by the dislance calculation.</Paragraph> <Paragraph position="3"> Head words are passed upward from lower palterns to higher 1)atterns. Since the head of the verb phrase pattern, &quot;X at Y &quot; is assigned te X, the head of &quot;leaves verb-propn Kyoto at eleven a.m.&quot; is &quot;leaves&quot;, which is tile head of &quot;leaves wrb-propn Kyoto&quot;. The head of &quot;the bus&quot; is &quot;bus&quot; fi'om the head information that the Table 4 Result of distance calculation in &quot;X a/Y &quot; in lqg. 1 input:(leave, a.m.) AL~JPSELeXxl)ression closest example and |IS value :~ Y' de X' (arrive, a.m.) O. 17 Y' ni X' (serve, reception) 0.67 Y' we X' (look, it) 1.00 head of &quot;the X &quot; is X. Thus, rite input of tile distance calculation of &quot;X noun-verb Y &quot; is (bits, leave).</Paragraph> </Section> <Section position="4" start_page="705" end_page="705" type="sub_section"> <SectionTitle> 4.4 SI,'uetural dis:mlbignation </SectionTitle> <Paragraph position="0"> Distance calculqtion selects not only the most l)lausible target expression but also the most plausible source structure. When .strtlcttlral aml)iguity occttrs, the most apllrOl)riate structure is selected by comt)uting tl~o totals for all possible combinations of partizfl distance values. The structure with the least total distance is judged most consistent wilh empirical knowledge, and is chosen as Ihe most 1)lausil)le structure (Furuse 1992, 1994; Sumita 1993).</Paragraph> <Paragraph position="1"> Table 5 shows the result of each partial distance talc|Ha|ion for tile structure in Fig. 1. l:mm Table 5, we V.Ct Ihe total distance value 1.17 for the structure in l:it;. 1.</Paragraph> <Paragraph position="2"> Table 5 Result of each partial distance calculation for tile slructure in I,'ig. 1 souiee chosen l~lr..~c:\[ distance val,lg</Paragraph> <Paragraph position="4"> The difference in total distance value I)etween two l)OSsible structures for sentence (1) is due only to the distance value of &quot;X at Y &quot;, for the structure in Figs. 1 and 2. For the strucltne in Fig. 2, the distance valtl0 of &quot;X at Y &quot; at tile neun phrase level is given as 0.83, as shown in Table 6, and is given a total distance ef 1.83.</Paragraph> <Paragraph position="5"> Thus, the structure in Fig. 1 is selected as the 3 The:.;e vii\]ties were col//pu,ed based on Ihe present transfer knowledge of the T1)MT system.</Paragraph> <Paragraph position="6"> appropriate restflt because it has the least total distance knowledge for the pattern &quot;X pron-noun Y &quot;; value.</Paragraph> <Paragraph position="7"> Table 6 Result of distance calcul,ltion in &quot;X at Y &quot; in Fig. 2 input:(Kyoto, a.m.) target expression C/losest exampl0 and its value Y' no X' (room, hotel) 0.83 Y' deno X' (language, conference) 1.00 In macbine translation, it ix important to disambiguate tbe possible structures, l)ecause a difference in structure may bring about a translation difference. For instance, the structures in Figs.1 and 2 give different Japanese translations (4) and (5), respectively. (4) is selected because it is generated from the best structure with the least total distance value.</Paragraph> <Paragraph position="8"> (4) basu wa gozen 11 ji ni Kyoto we de masu 4 (5) basu wa gozen \] 1 ji ~_ Kyoto we de masu</Paragraph> </Section> </Section> <Section position="7" start_page="705" end_page="705" type="metho"> <SectionTitle> 5 Constituent Boundary Parsing in </SectionTitle> <Paragraph position="0"> Japanese Since a postposition is quite often used as a caseparticle in Japanese, tim botmdary markers expressed by a part-of-speech bigram may not be used less frequently than in English. However, in spoken Japanese, postpositions are frequently omitted. The Jqpanese sentence &quot;Kochira wa jimukyoku&quot; where kochira means this and jimukyoku means &quot;office&quot;, is translated into the English sentence &quot;77fis is the office&quot; by applying transfer knowledge such as the following5:</Paragraph> <Paragraph position="2"> But postpositions are often omitted in natural six)ken Japanese, e.g. in the sentence &quot;Kochira jimukyoku.&quot; The sentence can thus be divided into two noun phrases, &quot;kochira&quot; and &quot;jimukyoku.&quot; &quot;kochira&quot; is a pronotm, and &quot;jimukyoku&quot; is a noun. So, using the bigram method of marking boundaries, we get &quot;Kochira pron-noun jimukyoku&quot;, where the bigram &quot;pron-noun&quot; was inserted. The English sentence &quot;77fis is the oJfice&quot; can then be produced by applying the following transfer 4&quot;basu&quot;, &quot;de&quot;, and &quot;masu&quot; mean &quot;bus&quot;, &quot;leave&quot;, and a polite sentence-final form, respectively.</Paragraph> <Paragraph position="3"> 5 For simplicity, examples and other possible target expressions are omined.</Paragraph> <Paragraph position="4"> X pron-noun Y => X' be Y' In Japanese adnominal expressions, too, constituei~t bonndary markers ,'Ire inserted between the modifier and the modified.</Paragraph> </Section> <Section position="8" start_page="705" end_page="705" type="metho"> <SectionTitle> 6 Results </SectionTitle> <Paragraph position="0"> We have evaluated tim efficiency of our parsing method by utilizing a Japanese-lo-English (Jg) and English-to-Japanese (E J) TDMT prototype system (Furuse 1994; Sobashima 1994), which ix ,'unning on a Symbolics XL120(I, a LISP machine with IOMIPS performance. The system's domain is inquiries concerning international conference registrations. The efficency is evaluated with 154 Japanese sentences and 138 corresl)onding English sentences, which are extracted from 10 dialogues in the domain. The systeln has al)out 500 source p,'llterns for JE translation and about 35(1 source patterns for EJ transhttion.</Paragraph> <Paragraph position="1"> The test sentences mentioned above have already l)een tr:tined to investigate the efficiency of the method, and can be p-lrse(l correctly by the system. Table 7 outlines the 154 Japanese sentences and 138 corresponding English sentences.</Paragraph> <Paragraph position="2"> Table 7 Outline of test senlences _ Japanese E_j1Aj_I i sh words per inpnt sentence 9.8 8.7 average numl)er of ix)ssible structures 1.5 4.8 An l-nglish sentence tends to have more struclural ambiguities than a Japanese sentence, bec,'tnse of PF'altachment, the phenomenon that an English preposition f)rodtlCCS \[)()\[h a noun verb p\]lrasc \[Ilia a \[iolln phasc. In contrast, tile Jai)aneso l)ostposition does not generally produce different-level constituents.</Paragraph> <Paragraph position="3"> Table 8 shows how ,nuch time it takes to reach the best structure and translation output in our JE and EJ TDMT system. The processing time for distance calculation includes strnctnral disaml)iguation in addition to ktrget pattern selection.</Paragraph> <Paragraph position="4"> Tiffs demonstrates that the ot~r parsing method can get the best structure and translation output quickly wit\]fin the examl)lo-/xlsed framework.</Paragraph> </Section> <Section position="9" start_page="705" end_page="705" type="metho"> <SectionTitle> 6 JF. E,I </SectionTitle> <Paragraph position="0"> derivation of possible structures 0.25 (scc) 0.l 7 dislance calculation 1.32 0.14 whole tr,'lnsl;ition 2.17 1.07</Paragraph> </Section> class="xml-element"></Paper>