<?xml version="1.0" standalone="yes"?>
<Paper uid="C88-1059">
  <Title>Completion of Japanese Sentences by Inferring Function Words from Content Words</Title>
  <Section position="2" start_page="0" end_page="291" type="metho">
    <SectionTitle>
2 Basic assumptions
</SectionTitle>
    <Paragraph position="0"> In this study the following restrictions relevant to the interface problem are assumed:  (1) A Japanese sentence usually consists of a certain number of noun phrases followed by a verb phrase at the end. The basic unit of speech recognition is assumed to be a continuously uttered phrase, so that any input to the machine translation module is a 'phrase lattice', i.e., a set of phrase candidates hypothesized by the speech recognition module.</Paragraph>
    <Paragraph position="1"> (2) The range of telephone conversation tasks is  restricted to inquiries from a researcher to a clerk about an international conference concerning the main topic of the conference, deadlines for paper submission, exhibitions, social events, accommodation, payment, cancellation, etc..</Paragraph>
    <Paragraph position="2"> utterance results of</Paragraph>
    <Paragraph position="4"> An example of a phrase lattice.</Paragraph>
    <Paragraph position="6"/>
  </Section>
  <Section position="3" start_page="291" end_page="291" type="metho">
    <SectionTitle>
3 Research goal
</SectionTitle>
    <Paragraph position="0"/>
    <Section position="1" start_page="291" end_page="291" type="sub_section">
      <SectionTitle>
3.1 A phrase lattice as the result of speech recognition
</SectionTitle>
      <Paragraph position="0"> Consider a Japanese sentence consisting of two noun phrases and one verb phrase: 'genkou-no shimekiri-wa itsu-desuka'.</Paragraph>
      <Paragraph position="1"> ('When is the deadline for a manuscript?') Usually a Japanese phrase begins with a jiritsugoword (J-word for short) such as a noun or verb, and ends with a sequence of fuzokugo-words (E-words for short) such as postpositional particles or auxiliary verbs. In the above notation, boundaries between J-words and F-words are explicitly indicated by hyphens, and all F-words are italicized.</Paragraph>
      <Paragraph position="2"> Figure 1 shows an example of a phrase lattice for this sentence obtained as the result of speech recognition. Notice that there are candidates for both J-words and F-words together with a recognition score of the probability that the word is correct. The problem is to select the most appropriate candidate from this phrase lattice.</Paragraph>
    </Section>
    <Section position="2" start_page="291" end_page="291" type="sub_section">
      <SectionTitle>
3.2 Selection-by-generation approach
</SectionTitle>
      <Paragraph position="0"> Attention is focused on candidates for F-words, assuming that J-words have already been correctly selected by a suitable method.</Paragraph>
      <Paragraph position="1"> The assumption that J-words have been correctly selected is realistic if the task domain is limited enough to allow a high recognition rate for J-words and a knowledge-base, etc. is available for the limited task domain. Techniques related to this procedure are now being studied. Of the J-words, the predicate atthe end of a sentence is less accurately recognized by the speech recognition module than nouns. A method to solve this problem will be discussed in the second half of this paper. In Figure 1, for instance, it is assumed that a sequence of J-words: 'c)enkou' 'shimekiri' 'itsu' (manuscript') ('deadline') ('when') has been correctly selected according to the recognition scores. Corresponding to these J-words, there are three sets of candidates for F-words in the phrase lattice: 'wo' 'wa' 'desuka' 'too' ' ga&amp;quot; &amp;quot;deshita ' 'no' &amp;quot;bekika'.</Paragraph>
      <Paragraph position="2"> Our major concern here isthe subproblem of selecting the most appropriate one in each of these sets.</Paragraph>
      <Paragraph position="3"> This sub-problem is characteristic of the Japanese language. In fact, as easily seen in the above example, frequently used F-words, specifically those indicating grammatical cases such as 'ga', &amp;quot;wo', &amp;quot;ni', etc., are too short to be recognized correctly. Their recognition scores are much lower than those of J-words. But in Japanese it is often possible to infer the meaning of a given sentence from the sequence of J-words when the task domain is narrow.</Paragraph>
      <Paragraph position="4"> 2P)2 Our method of selecting the correct F-words is composed of two steps: 1) generate a meaningful sentence by inferring suitable F-words for a given sequence of J-words, and 2) compare these inferred F-words with the candidates in the phrase lattice to select those most appropriate.</Paragraph>
      <Paragraph position="5"> This idea of 'selection-by-generation' distinguishes this approach from previous ones: Hayes et al. \[1\] for English or Niedermair \[2\] for German. In this paper only Step 1, which is considered the key step, will be discussed. 4 Generating a sentence by inferring F-words The task domain is restricted to inquiries about an international conference, and therefore the dialogue is basically a repetition of simple questions and answers. This increases the probability of inferring the correct F-words for each phrase.</Paragraph>
    </Section>
    <Section position="3" start_page="291" end_page="291" type="sub_section">
      <SectionTitle>
4.1 Key information for the inference
</SectionTitle>
      <Paragraph position="0"> The following types of information are used to infer F-words. The information is described in a lexicon of Jwords. null  (1) Semantic features of nouns and valency patterns  First, two types of semantic features are set up for nouns appearing in the restricted task domain. One is a general type of semantic feature, independent of the task domain, such as abstract, action, concrete, human, location, time, number and diversity. The other is a specific type of semantic feature dependent on the task domain. Table 1 shows examples of such features. Using these semantic features valency patterns of the basic predicates necessary in the task domain are defined. As an example, the predicate 'okuru' ('send' in English) is given the following valency patterns:</Paragraph>
      <Paragraph position="2"> etc..</Paragraph>
      <Paragraph position="3"> The first valency pattern in this list, for instance, specifies that the predicate V ('okuru') can take one noun phrase consisting of a noun with the general semantic feature 'concrete' / specific semantic feature non-'transport' and F-word 'we'.</Paragraph>
      <Paragraph position="4"> In this way the valency patterns summarize the basic J-word and F-word relationships, and thus give the most essential information for inferring F-words from a given  sequence of J-words.</Paragraph>
      <Paragraph position="5"> gener___a/ abstract Table 1 Semantic features specific example logic 'riron'(theory), 'houhou'(method) state 'yousu'(state), 'baai'(case) 'language 'nihongo'(Japanese), 'eigo'(English) learning 'bunya'(field), 'senmon'(specialty) intention 'kyoumi'(interest), 'kibou'(hope) value 'hitsuyou'(necessity) sign 'namae'(name) labor 'youken'(business) concrete document 'genkou'(manuscript),'youshr(form) transport 'basu'(bus), 'tikatetsu'(subway) article 'syashin'(photograph) .___money. 'okane'(money),'kado'(cash card)  These valency patterns are obtained from the valency patterns of predicates (obtained from dialogues collected for the task of inquiries about an international conference). If necessary, certain modifications such as omission of the nominative case, modification of the word sequence or of the F-words, and addition of interrogative pronouns are carried out. In dialogue sentences, the nominative case such as 'watashi' (T) or deganata' ('you') is seldom used; hence, the nominative case is usually not included in the valency patterns. To describe the modification of the two valency patterns for 'okurU' ('send'):</Paragraph>
      <Paragraph position="7"> If the noun N\[con/-tra\] of the valency pattern (a) becomes the subjec% the F-word is replaced with 'wa'and the word sequence is changed, often resulting in the valency pattern (b). Interrogative pronouns are added to produce valency patterns specific to interrogative sentences because a large number of questions occur in this task domain.</Paragraph>
      <Paragraph position="8"> In a \]imited task domain, even individually optional cases behave in a similar way to the obligatory case for each predicate. Therefore, the optional cases are described in these valency patterns. When valency patterns were prepared for 65 words working as predicates, an average of about 11 valency patterns were produced for each predicate. Details will discussed in Chapter 5.</Paragraph>
      <Paragraph position="9"> (2) Connection of two nouns by F-word 'no&amp;quot; it is inferred that nouns which cannot be processed through valency patterns are likely to be connected with the F-word 'no' (roughly corresponding to 'of' in English) in the form 'A no B', where A and B denote nouns. For a given noun A, the other noun B can also be specified through the semantic features. For instance, the noun 'kaigr ('conference') can be joined with other nouns as follows:</Paragraph>
      <Paragraph position="11"> As shown above, whether or not to insert the F-word &amp;quot;no&amp;quot; is automatically determined by presetting which nouns are to be connected with the F-word &amp;quot;no'.</Paragraph>
      <Paragraph position="12">  (3) Syntactic information Pure syntactic knowledge is also useful in this process. It is known that, in Japanese, no F-word can be attached to an adverb or a conjunction, and that a verb in  conditional form can be connected with an adjective via F-words such as'ba'.</Paragraph>
      <Paragraph position="13"> In addition, the following rules are used, for  (May I record the speech of the conference atthe hall?).</Paragraph>
      <Paragraph position="14"> In this case it is assumed that J-words 'kaigi', 'yousu', 'rokuonshi', and 'ii' are correctly recognizable. The inference proceeds as follows: 1) syntactic information can connect 'rokuonshi' and 'ii' with F-words 'te&amp;quot; or &amp;quot;temo' and 'desuka' to generate the phrases 'rokuonshi-te' or 'rokuonshi-temo' and &amp;quot;ii.desuka',  respectively, 2) considering the semantic features of the first three J-words, and taking the fourth J-word 'rokuonsuru' as V, the valency pattern: N\[Ioc\]'de' + N\[act,abs\]'wo' + V, can be applied to them, 3) there are two possible connections: 'kaigi-no yousu' and 'kaijyou-no kaigi', and 4) considering both 2) and 3) together, sentences: 'kaijyou-(de, no) kaigi-no yousu-wo rokuonshi-(te, temo) ii-desuka', can finally be derived.</Paragraph>
      <Paragraph position="15"> In a similar way, (b) shows how the following sentence is to be processed: 'genkou-wa itsu-madeni okure.ba yoroshii-desuka ', (By what time may I send the manuscript ?).</Paragraph>
      <Paragraph position="16"> Here, 1) 'okure' is combined with 'yoroshii' by the F-words 'ha' and 'desuka', to yield 'okure-ba yoroshii-desuka', 2) analyzing the semantic features of the nouns 'genkou' and 'itsu' and the presence of the verb 'okuru', the following valency pattern is applied:  Using the valency patterns obtained from collected dialogue sentences, we carried out an experiment of producing sentences. Of the total of 256 interrogative sentences, 146 were used in determining valency patterns. The number of verbs was 65 and that of nouns was 229. In total, 669 valency patterns were prepared (10.7 patterns for each verb on the average).</Paragraph>
      <Paragraph position="17"> In addition to the collected dialogue sentences, we prepared 70 test questions. For these interrogative sentences, we carried out a sentence-producing experiment. The results of this experiment are shown in  (b) Or_~nal utterance 'genkou-wa itsu-madeni okure-ba yoroshii.desuka'-- I ~F-words ~to be inferred Sequence of J-words correctly recoqnized L'o edeg-u' .,,so. iiill-- ,u,e' iiilI ,- .............. ......... , r- ................... r-- ~ ..... ~ ', Lexicon of J-word=s noun ', noun ~ verb adjective \] part of speech ,m~ con/doc '1 tim/pro. '~ conditional conclusive -~..-~ conjugation \ I _~_ = _~. I semantic features ...... ~ ........... T--_ ....... f ..... ~ ...... -~-'.-. ...... 2)Valency pattern - LN\[c_deg_n/ddegc\]:wa_'_+N\[ti.m_\]:ma_den!_+V__ i l 3)Connection.of..n. oun.s. ...................... ; ~re- ba yoroshii.desuka'l)Syntactic informat!onj 'genkou no itsu' / ,,Sente=e ........... .... / \[ '..n.ou-wa ...-ma,e.. o,.=:~a .oro.,,-,.uka' I  each test sentence was input, and a complete sentence including F-words was output. The correct answer rate in Table 2 is the percentage of all the output sentences that were consi~.~tent with the input sentences. At the first trial, 64.3% coirect sentences were produced from the prepared valency patterns.</Paragraph>
      <Paragraph position="18">  The !lpper hatf of Table 2 shows the number of candidates in the trials where some of the output sentences were correct. For example, the number 7 shown for the first trial in the line corresponding to 5 candidate sentences means the number of candidate sentences was 5~ and that 7 of the 70 test sentences were correct ones. The lower half of Table 2 shows the number of test sentences in'trials where no candidate sentence was produced or where none of the candidate sentences produced were correct.</Paragraph>
      <Paragraph position="19"> The figures for the second and subsequent trials in Table 2 show the change' in the correct answer rate when additional valency patterns were used to increase the incidence of correct sentences. In this experiment, enough valency patterns were added so th'at the sixth trial always produced correct sentences.</Paragraph>
      <Paragraph position="20"> AI~ elf the test sentences used in the above experiment were simple interrogative sentences. As shown in Table 2, tile Use of ~alency patterns allows easy production of a complete sentence from a given sequence of J-words.</Paragraph>
    </Section>
  </Section>
  <Section position="4" start_page="291" end_page="291" type="metho">
    <SectionTitle>
6 Inferring an omitted verb
</SectionTitle>
    <Paragraph position="0"> The verb in a given sequence of J-words has an important role in this method because it allows the selection of a correct valency pattern. It would he difficult to proceed by this method if the verb is omitted for some reason, such as speech recognition failure, or if it were originally omitted as is often the case in Japanese dialogue.</Paragraph>
    <Paragraph position="1"> However, in this restricted domain, nouns with particular semantic features are often related to particular verbs. For example, as shown in Figure 3, in sentences which contain a noun with the semantic feature of concrete/document, the noun + F-word 'wo' tends to be accompanied by the verb 'okuru'('send'), 'kaku'('write') or 'motsu'('have'), and the noun + F-word 'hi' tends to be accompanied by the verb 'kinyuusuru'('enter').</Paragraph>
    <Paragraph position="2"> This suggests the possibility of inferring an omitted verb from the nouns by inversely applying a suitable valency pattern. In fact, the definition of a valency pattern can be generalized as follows: N\[sem Iglsem Is\] + N\[sem2glsem2s\] -t .... + V\[v-class\], where V\[v-class\] denotes a verb belonging to verb class 'vclass'. This valency pattern can be used to infer a verb V\[v-class\] associated with a set of nouns N\[semlglsemls\], N\[sem2glsem2s\], etc..</Paragraph>
    <Paragraph position="3"> This is schematically illustrated in Figure 4. When there is a verb group A (consisting of verbs which are inferred when a certain F-word is added to noun A) and a verb group B (consisting of verbs which are inferred when a certain F-word is added to noun B),the common area of these two groups indicates the verbs which are inferred from the valency pattern containing noun A and noun B.</Paragraph>
    <Paragraph position="4"> For example, in a sentence which contains a noun  with the semantic features of concrete/document and a noun with the semantic features of location/position, verbs such as 'aru'('be') and 'motsu'('have') tend to be selected, and F-words specific to these verbs are chosen. Table 3 shows the number of verbs which are inferred from a given sequence of nouns using the valency patterns described in Chapter 5. Since valency patterns were prepared for 65 verbs, the verbs are inferred -from these 65.</Paragraph>
    <Paragraph position="5"> The columns in Table 3 Show the number of verbs inferred. The lines in the same table show the number of nouns in the valency patterns. For example, when the number of inferred verbs is 5, there are 6 valency patterns where 5 verbs are inferred; and of these 6 patterns 1 has one noun, 4 have 2 nouns, 1 has 3 nouns and none have 4 nouns.</Paragraph>
    <Paragraph position="6"> In counting the number of inferred verbs, only verbs having a valency pattern consistent with given valency patterns are counted. When the noun of a valency pattern bears a specification as to the upper-level general semantic features but no specification as to the lowertask~dependent semantic features, the verbs of the valency patterns bearing a specification as tothe lowerevel semantic features are counted. Conversely, for valency patterns where the lower-level semantic feature is specified, the verbs bearing no specification as to the Iower.-levei semantic features are not counted. For example, in the followingtwo valency patterns, the #erbs Vl and V2 are inferred from the pattern (a), while only the verb V2 is inferred from the pattern (b).</Paragraph>
    <Paragraph position="7"> N\[con\] + N\[Ioc\] + Vl (a) N\[con/doc\] + N\[Ioc\] + V2 (b) As can be seen in Table 3, only one verb was inferred in more than 50% of the valency patterns. Irl 90% of the remaining valency patterns where multiple verbs were in~erred, the number c~f verbs i~fe~rred was 6 o~ le~s. ~hese Table 3 Verbs inferred using valency patterns  results indicate that, in a restricted task domain, the semantic features of the preceding nouns and valency patterns allow a fairly restricted number of candidate verbs to be inferred.</Paragraph>
  </Section>
  <Section position="5" start_page="291" end_page="291" type="metho">
    <SectionTitle>
7 Conclusions
</SectionTitle>
    <Paragraph position="0"> As a first step toward a better interface between speech recognition and machine translation, a method which is particularly useful for Japanese sentences was proposed to infer F-words for a given sequence of Jwords. null In a restricted task domaia., the most appropriate F-word can be inferred from a given sequence of J-words if the task-dependent semantic features of nouns are preset and the information of valency patterns is utilized.</Paragraph>
    <Paragraph position="1"> In addition, the results of this study suggest that correct verbs can be inferred from valency patterns.</Paragraph>
    <Paragraph position="2"> The authors are now evaluating the effectiveness of the procedures proposed in this paper by applying them to actual results of speech recognition.</Paragraph>
  </Section>
class="xml-element"></Paper>