File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/98/p98-2223_metho.xml
Size: 12,905 bytes
Last Modified: 2025-10-06 14:15:02
<?xml version="1.0" standalone="yes"?> <Paper uid="P98-2223"> <Title>A Pattern-based Machine Translation System Extended by Example-based Processing</Title> <Section position="3" start_page="0" end_page="1369" type="metho"> <SectionTitle> 2 Pattern-based Translation </SectionTitle> <Paragraph position="0"> Here, we briefly describe how the pattern-based translation works. (See \[6, 7\] for details.) A translation pattern is a pair of source CFG-rule and its corresponding target CFG-rule. The followings are examples of translation patterns.</Paragraph> <Paragraph position="1"> (pl) take:VERB:l a look at NP:2 =~ VP:I</Paragraph> <Paragraph position="3"> The (pl) is a translation pattern of an English colloquial phrase &quot;take a look at,&quot; and (p2) and (p3) are general syntactic translation patterns. In the above patterns, a left-half part (like &quot;A B C =~ D&quot;) of a pattern is a source CFG-rule, the righthalf part (like &quot;A C/:= B C D&quot;) is a target CFG-rule, and an index number represents correspondence of terms in the source and target sides and is also used to indicate a head term (which is a term having the same index as the left-hand side 2 of a CFG-rule).</Paragraph> <Paragraph position="4"> Further, some features can be attached as matching conditions for each term.</Paragraph> <Paragraph position="5"> The pattern-based MT engine performs a CFGparsing for an input sentence with using source sides of translation patterns. This is done by using chart-type CFG-parser. The target structure is constructed by the synchronous derivation which generates a target structure by combining target sides of translation patterns which are used to make a parse.</Paragraph> <Paragraph position="6"> Figure 2 shows how an English sentence &quot;She takes a look at him&quot; is translated into Japanese. 2we call the destination of an arrow of a CFG rule description the left-hand side or LHS, on the other hand, we call the source side of an arrow the right-hand side or RHS. S ......................................................................... S VP ................................................................... VP .........................</Paragraph> <Paragraph position="7"> NP ............. .. -.. NP ... NP NP pron verb det noun prep pron pron cm pron cm verb :&quot; \] : . ha i wo miru She take a look at him : .</Paragraph> <Paragraph position="8"> ........ &quot; ......................... &quot; .......... ::.:: .... - ........ &quot; ..... -&quot;i .... ..... &quot; ................................. .... :::: ........ ,::::i .... .... ......... i ...... i kanojo ha kare wo (she) (subj) (he) (dobj) In this figure, a dotted line represents the correspondence of terms in the source side and the target side. The source part of (p3) matches &quot;She&quot; and &quot;him,&quot; the source part of (pl) matches a segment consisting &quot;take a look at&quot; and a NP(&quot;him&quot;) made from (p3), and finally the source part of (p2) matches a whole sentence. A target structure is constructed by combining target sides of (pl), (p2), and (p3). Several terms without lexical forms are instantiated with translation words, and finally a translated Japanese sentence &quot;kanojo(she) ha(sub j) kare(he) wo(dobj) miru(see)&quot; will be generated.</Paragraph> </Section> <Section position="4" start_page="1369" end_page="1369" type="metho"> <SectionTitle> 3 Pruning Techniques </SectionTitle> <Paragraph position="0"> As mentioned earlier, our basic principle is to use many lexical translation patterns for producing natural translation. Therefore, we use more CFG rules than usual systems. This causes the slow-down of the parsing process. We introduced the following pruning techniques for improving the performance.</Paragraph> <Section position="1" start_page="1369" end_page="1369" type="sub_section"> <SectionTitle> 3.1 Lexical Rule Preference Principle </SectionTitle> <Paragraph position="0"> We call a CFG rule which has lexical terms in the right-hand side (RHS) a lexical rule, otherwise a normal rule. The lexical rule preference principle (or LRPP} invalidates arcs made from normal rules in a span in which there are arcs made from both normal rules and lexical rules.</Paragraph> <Paragraph position="1"> Further, lexical rules are assigned cost so that lexical rules which has more lexical terms are preferred. null For instance, for the span \[take, map\] of the following input sentence, He takes a look at a map.</Paragraph> <Paragraph position="2"> if the following rules are matched, (rl) take:verb a look at NP (r2) take:verb a NP at NP (r3) take:verb NP at NP (r4) VERB NP PREP NP then, (r4) is invalidated, and (rl),(r2), and (r3) are preferred in this order.</Paragraph> </Section> <Section position="2" start_page="1369" end_page="1369" type="sub_section"> <SectionTitle> 3.2 Left-Bound Fixed Exclusive Rule </SectionTitle> <Paragraph position="0"> We generally use an exclusive rttle which invalidates competitive arcs made from general rules for a very special expression. This is, however, limited in terms of the matching ability since it is usually implemented as both ends of rules are lexical items.</Paragraph> <Paragraph position="1"> There are many expression such that left-end part is fixed but right-end is open, but these expressions cannot be expressed as exclusive rules. Therefore, we introduce here a left-bound fixed exclusive (or LBFE) rule which can deal with right-end open expressions.</Paragraph> <Paragraph position="2"> Given a span \[x y\] for which an LBFE rule matched, in a span \[i j\] such that i<x and x<j<y, and in all * Rules other than exclusive rules are not applied, and * Arcs made from non-exclusive rules are invalidated. null Fig.2 shows that an LBFE rule &quot;VP ~= VERB NP ''3 matches an input. In spans of (a),(b), and (c), arcs made from non-exclusive rules are invalidated, and the application of non-exclusive rules are inhibited.</Paragraph> </Section> <Section position="3" start_page="1369" end_page="1369" type="sub_section"> <SectionTitle> 3.3 Preproeessing </SectionTitle> <Paragraph position="0"> Preprocessing includes local bracketing of proper nouns, monetary expressions, quoted expressions, Internet addresses, and so on. Conversion of numeric expressions and units, and decomposition of unknown hyphenated words are also included in the preprocessing. A bracketed span works like an exclusive rule, that is, we can ignore arcs crossing a bracketed span. Thus, accurate preprocessing not only improved the translation accuracy, but it visibly improved the translation speed for longer sentences. null</Paragraph> </Section> <Section position="4" start_page="1369" end_page="1369" type="sub_section"> <SectionTitle> 3.4 Experiments </SectionTitle> <Paragraph position="0"> To evaluate the above pruning techniques, we have tested the speed and the translation quality for three documents. Table 1 shows the speed to translate documents with and without the above pruning techniques. 4 The fourth row shows the recorded about two years ago and the latest version is much faster.</Paragraph> <Paragraph position="1"> number of sentences tested with pruning which become worse than sentences without pruning and sentences with pruning which become better than without pruning.</Paragraph> <Paragraph position="2"> This shows the speed with pruning is about 2 times faster than one without pruning at the same time the translation quality with pruning is kept in the almost same level as one without pruning..</Paragraph> </Section> </Section> <Section position="5" start_page="1369" end_page="1371" type="metho"> <SectionTitle> 4 Extension by Example-based Pro- </SectionTitle> <Paragraph position="0"> cessing One drawback of our pattern-based formalism is to have to use many rules in the parsing process. One of reasons to use such many rules is that the matching of rules and the input is performed by the exact-matching. It is a straightforward idea to extend this exact-matching to fuzzy-matching so that we can reduce the number of translation patterns by merging some patterns identical in terms of the fuzzy-matching. We made the following extensions to the pattern-based MT to achieve this example-based processing.</Paragraph> <Section position="1" start_page="1369" end_page="1371" type="sub_section"> <SectionTitle> 4.1 Example-based Parsing </SectionTitle> <Paragraph position="0"> If a term in a RHS of source part of a pattern has a lexical-form and a corresponding term in the target part, then it is called a \]uzzy-match term, otherwise an exact-match term. A pattern writer can intentionally designate if a term is a fuzzy-match term or an exact-match term by using a doublequoted string (for fuzzy-match) or a single-quoted string (for exact-match).</Paragraph> <Paragraph position="1"> For instance, in the following example, a word make is usually a fuzzy-match term since it has a corresponding term in the target side (ketsudansuru), but it is a single-quoted string, so it is an exact-match term. Words a and decision are exact-match terms since they has no corresponding terms in the target side.</Paragraph> <Paragraph position="2"> 'make':VERB:l a decision =v VP:I VP:I ~= ketsudan-suru:l Thus, the example-based parsing extends the term matching mechanism of a normal parsing as follows: A term TB matches another matched-term TA (LexA,POsB) s if one of the following conditions holds.</Paragraph> <Paragraph position="3"> (1) When a term TB has both LexB and POSB, (1-1) LexB is the same as LeXA, and PosB is (1-2) TB is a fuzzy-match term, the semantic distance of LexB and LexA is smaller than a criterion, and PosB is the same as ROSA.</Paragraph> <Paragraph position="4"> (2) When a term TB has only LexB, (2-1) LexB is the same as LeXA.</Paragraph> <Paragraph position="5"> (2-2) LexB is a fuzzy-match term, the semantic distance of LexB and LexA is smaller than a criterion.</Paragraph> <Paragraph position="6"> (3) When TB has only PosB, then PosB is the same as RosA.</Paragraph> </Section> <Section position="2" start_page="1371" end_page="1371" type="sub_section"> <SectionTitle> 4.2 Prioritization of Rules </SectionTitle> <Paragraph position="0"> Many ambiguous results are given in the parsing, and the preference of these results are usually determined by the cost value calculated as the sum of costs of used rules. This example-based processing adds fuzzy-matching cost to this base cost. The fuzzy-matching cost is determined to keep the following order.</Paragraph> <Paragraph position="1"> (1-1) < (1-2),(2-1) < (2-2) < (3) The costs of (1-2) and (2-1) are determined by the fuzzy-match criterion value, since we cannot determine which one of (1-2) and (2-1) is preferable in general.</Paragraph> </Section> <Section position="3" start_page="1371" end_page="1371" type="sub_section"> <SectionTitle> 4.3 Modification of Target Side of Rules </SectionTitle> <Paragraph position="0"> Lexical-forms written in the target side may be different from translation words of matched input word, since the fuzzy-matching is used. Therefore, we must modify the target side before constructing a target structure.</Paragraph> <Paragraph position="1"> Suppose that a RHS term tt in the target side of a pattern has a lexical-form wt, tt has a corresponding term t, in the source side, and G matches an input word wi. If wt is not a translation word of wi, then wt is replaced with translation words of wi.</Paragraph> </Section> <Section position="4" start_page="1371" end_page="1371" type="sub_section"> <SectionTitle> 4.4 Translation Example </SectionTitle> <Paragraph position="0"> Figure 3 shows a translation example by using example-based processing described above.</Paragraph> <Paragraph position="1"> In this example, the following translation patterns are used.</Paragraph> <Paragraph position="2"> (p2) NP:I VP:2 =~ S:2 S:2 C/= NP:I ha VP:2 (p3) PRON:I =~ NP:I NP:I C/: PRON:I (p4) take:VERB:l a bus:2 =~ VP:I VP:I C/= basu:2 ni noru:VERB:l The pattern (p4) matches a phrase &quot;take a taxi,&quot; since &quot;taxi&quot; and &quot;bus&quot; are semantically similar. By combining target parts of these translation patterns, a translation &quot;PRON ha basu ni noru&quot; is generated. In this translation, since &quot;basu(bus)&quot; is not a correct translation of a corresponding source word &quot;taxi,&quot; it is changed to a correct translation word &quot;takusi(taxi).&quot; Further, PRON is instantiated by &quot;watashi&quot; which is a translation of &quot;I.&quot; Then a correct translation &quot;watashi ha takusi ni noru&quot; is generated.</Paragraph> </Section> </Section> class="xml-element"></Paper>