Verbal Case Frame Acquisition from Bilingual Corpora

1 Introduction

For a practical research in natural language processing, it is indispensable to develop a large scale semantic dictio­ nary. It is especially important to devise techniques of compiling semantic dictionaries from natural language texts in existing human dictionaries and/or large cor­ pora. There are, however , at least two difficulties m analyzing real texts :

1. The problem of syntactic ambiguities
2. The problem of polysemy

Our approach to solving these difficulties is to make use of translation examples of two distinct languages that have quite different syntactic structures and word meanings (such as English and Japanese), and to com­ pare analyzed results of these languages [Utsuro et a/.,

·The authors would like to thank the editorial staff of Kodansha for permission to use the data of Japanese-English dictionary, and also thank Dr. Shouichi Yokoyama of ETL, and Prof. Hozumi Tanaka and Dr. Takenobu Tokunaga of Tokyo Institute of Technology, for providing them the data of Japanese-English dictionary. This work is partly supported by the Grants from Ministry of Education , #03245103.

1992]. In many cases, translation pairs of those lan­ guages have different types of syntactic ambiguities, and comparison of syntactic structures of both lan­ guages helps to resolve them. Multiple senses of a word may be reduced by the information of translated word of the opposite language [Dagan et a/., 1991; Tokunaga and Tanaka, 1991]. For example, in the case of the translation example given in Example 1, both syn­ tactic and semantic ambiguities are resolved.

Example 1

I hung my coat on the hook.

fl..

1;1: J:.if a-

'IJ· ~ 1:

'/J·It t.: o

I TOP coat ACC hook on hung

1. Syntactic disambiguation

The English sentence of Example 1 is syntactically ambiguous because the prepositional phrase "on the hook" can modify either of the verb "hung" and the noun phrase "my coat". On the other hand, in the Japanese sentence , the phrase "iJ''tJ, ~::"can modify nothing but the verb "iJ' It t::." . Thus , if knowledge about word equivalence pairs such as (J,fJ.... ), (hung , iJ'I::I' t::. ), (coat, J:!f ), (hook, iJ' ~·) are available from a bilingual dictionary, the am­ biguity of pp-attachment is resolved by matching the structures of the two sentences.

2. Semantic disambiguation

The verb "il' It 7.>" in the Japanese sentence is a typical Japanese polysemy. This verb has six sub-entries in a Japanese dictionary that has about 70,000 entries, and ten translatable English verbs ( "hang", "spend", "play", etc.) in a Japanese­ English dictionary. It is not easy to associate the surface word "iJ'It 7.>" with its exact sense. With the translation example, the corresponding English verb "hang" helps to find the sense of "iJ'It 7.>" .

From this viewpoint, we proposed the framework of Lexical Knowledge Acquisition [Utsuro et a/., 1992], as in Figure 1. In our framework, pairs of English and Japanese sentences are syntactically analyzed without any semantic knowledge and are translated into ( disjunc­ tive) feature descriptions, which represent dependency structures of the sentences. Then these feature descrip­ tions of the two languages are compared, or unified, us­ ing knowledge of word equivalence obtained from a bilin­ gual dictionary. In the process of unification, (some of)

syntactic ambiguities are resolved 1 . In the case of the translation example of Example 1, the bilingually unified feature description (henceforth bilingual feature descr·ip­ tion) is obtained unambiguously as follows :

In this bilingual feature description, the verb is repre­ sented as the pair of English and Japanese atomic val­ ues (hang, fJ' It o ) , and the feature labels are repre­ sented as pairs of English and Japanese feature labels like (subj, li ), (obj, ~ ), (on, 1: ). This is also the case with the nouns of the case elements. From the collection of these disambiguated data, lexical knowledge such as surface case frames of verbs is acquired.

The bilingual corpus we have collected includes 40,000 translation examples obtained from a machine read­ able Japanese-English dictionary (Kodansha Japanese­ English Dictionary [Shimizu and Narita, 1979]). In [Ut­ suro et al., 1992], we described the method of resolv­ ing the syntactic ambiguities of translation examples by unifying the feature descriptions of the two languages . Although the success ratio of unambiguous analysis of Japanese sentences alone are 22%, the success ratio of acquiring unambiguous case structure of each sentence goes up to 59% when the bilingual unification is em­ ployed. In this paper, we describe the process of acquir­ ing surface case frames of Japanese verbs in detail.

1 Recently we developed another method to find structural matching between parallel sentences of two languages, which is presented in [Matsumoto et a/., 1993]. The main idea is to find a set of pairs of minimal corresponding substructures of the dependency structures between parallel sentences.

When designing a framework of knowledge acquisi­ tion by computer program, we have to decide how much human-machine interaction is necessary. There are sev­ eral research activities of compiling semantic dictionar­ ies or knowledge bases by hand [EDR, 1989; IPA, 1987; Lenat and Guha, 1990]. Several difficulties arise in this approach. Huge amount of manual labor is necessary for compiling large scale dictionaries, and the expansion of the dictionaries is difficult once they are compiled by hand. Moreover, the results tend to be unstable because the quality of human writers is not uniform. On the other hand, there are other research activities that at­ tempt to compile semantic dictionaries automatically or semiautomatically by extracting lexical knowledge from natural language texts taken from human dictionaries or large corpora (e.g. [Brent, 1991; Tomiura et al., 1992; Tsurumaru et al., 1991]). Computer program may do away with most of the above difficulties of scale and sta­ bility. However, currently available knowledge resources are too poor to make it possible to extract lexical knowl­ edge from natural language texts in a fully automatic way. Thus, we consider that human interaction is nec­ essary to some extent though it should be kept to a minimum in order to make the results stable. We take an approach that the system asks questions to human only when it comes to a crucial point of acquiring lexical knowledge. In order to acquire stable lexical knowledge, we limit the interaction between human and the system to a selection among a few choices.

This paper introduces a method of acquiring surface case frames of verbs from the collection of bilingual fea­ ture descriptions. The acquisition process is depicted in Figure 2. We start with a set of bilingual feature de­ scriptions which contain the target verb. The acquisition process is divided into two sub-processes:

1. Discrimination of multiple senses of the target verb.
2. Construction of the case frame for each sense of the target verb, making use of a nominal thesaurus.

For discriminating multiple senses of a verb, the system detects the clue to the discrimination using some heuris­ tics, and asks the human instructor whether the detected clue is correct or not. When acquiring case frames of Japanese verbs, our heuristics for the clue detection be­ comes much more reliable compared with monolingual approaches, since Japanese words are paired with En­ glish information such as the corresponding English verb and English case labels. From the results of the heuris­ tics and human interaction, the set of bilingual feature descriptions is divided into subsets according to the dif­ ferent senses or usages of the target verb. Then, from each subset, a surface case frame of the target verb is constructed.

In Chapter 2, we introduce the data structure of the nominal thesaurus and the bilingual feature descriptions. In Chapter 3, we discuss how we can discriminate mul­ tiple senses of a Japanese verb. Chapter 4 describes the process of constructing a case frame for each sense of a verb. The results of some experiments and evaluation are shown in Chapter 5.

2 Data Structure

2.1 Nominal Thesaurus

We use semantic categories in a nominal thesaurus for describing the semantic restrictions of case slots . A nom­ inal thesaurus is regarded as a tree in which each node represents a semantic category. We define a nominal the­ saurus as a finite partially ordered set [SC, ~], where SC is the set of the semantic categories in the thesaurus, and ~ is the superordinate-subordinate relation of semantic categories . [SC, ~]has a single greatest element T which corresponds to the root of the tree . a V b represents the least upper bound of two semantic categories a and b. We assume that a unique least upper bound exists for all a and b.

At present we use an on-line thesaurus called 'Bunrui Goi Hyou'(BGH)[NLRI, 1964]. BGH has a six-layered abstraction hierarchy and more than 60,000 Japanese words are assigned at the leaves . In its nominal part, about 45,000 words are assigned at the leaves, and five semantic categories are allocated at the next level from T, which are abstract relations, agents of human activi­ ties, human activities, products and natural objects and natural phenomena. One word is assigned under several leaf semantic categories in the thesaurus when it has severa! senses. This thesaurus is not reliable enough for our target of acquiring case frames of verbs. It is, however, the most precise and broad covering Japanese nominal thesaurus obtainable to us currently.

2.2 Bilingual Feature Description

We acquire surface case frames of verbs from a collec­ tion of bilingual feature descriptions [Utsuro eta/., 1992]. A bilingual feature description is a feature description where the atoms and the labels are expanded to include paired symbols consisting of bilingual vocabularies:

V; and VE indicate the corresponding Japanese and English verbs, VoiceE denotes the voice of the English sentence, and Aux; denotes the auxiliary verb of the Japanese sentence. PJl , . .. ,PJn indicate the Japanese case markers and /E1 , .... , lEn indicate the corresponding English case markers. N Jl, . . . , N Jn indicate Japanese nouns of the case elements, and N El, . . . , N En indicate the corresponding English nouns. When a Japanese noun N; ; has several senses, it appears at several leaf semantic categories in the thesaurus. SEM; is a set of those semantic categories, and is referred to as a seman­ tic label.

When describing the semantic restrictions of case slots, this semantic ambiguity must be resolved and the correct semantic category Semi must be selected.

3 Discriminating Multiple Senses of A Verb

This chapter first discusses what can be the clues to dis­ criminating multiple senses of a Japanese verb, then pro­ poses a method for discriminating multiple senses of a verb.

3.1 Clues to Discriminating Multiple Senses of A Verb

One verb may have several senses and therefore may have as many case frames. In our framework, when discrim­ inating multiple senses of a Japanese verb, English in­ formation, such as the corresponding English verb and English case labels, as well as the semantic labels of Japanese nouns in case slots, are available . The clues are obtained by making use of such information.

Difference of the Corresponding English Verbs

When one Japanese verb has several senses, distinct senses are often translated into different English verbs. In such a case, the different English verbs can be the clue t.o discriminating senses of the Japanese verb. For example, Japanese verb "il'~t 7.> " has several senses and it is translated into different English verbs .

• -/J' tf ~~ _tift' ~ -/J' It 1.> hook on coat A CC hang I hang my coat. on the hook · n&. ~~ ~ ~ -/J'It '-> clothes on money A CC spend I spend money on clothes · f;f~ ~ i.PI:t .Q / play some music music A CC play

Difference of the Corresponding English Case Labels

When one Japanese verb has several case frames, it of­ ten happens that some case marker commonly occurring in distinct case frames play different roles . For exam­ ple, Japanese verb "!f <" has two case frames according to the different usages. Both case frames have a case marked by post positional particle "f;:".

• m.tli ~~ * ~ ~ < blackboa1·d on chamcte1· A CC w··ite I write a character on the blackboard

These "f::" cases have different senses and the differ­ ence of them can be the clue to discriminating these case frames of "If < ". As observed from this example, these "1:" cases are translated into different English prepo­ sitions, "on" and "to". Such an evidence appearing in English translation is available in our framework as the clue to discriminating these case frames.

Semantic Labels of Nouns in Case Slots

Sometimes, the corresponding English information is of no use to discriminate multiple senses of a Japanese verb. Although a Japanese verb has several case frames, it may be translated into the same English verb regard­ less of the difference of its senses . Also, it may happen that a Japanese case marker which appear in different case frames is translated into the same English case label, even when it plays a distinct role in each case frame . The following are two examples of case frames of a Japanese verb "IJ:f 7.>" .

• .UT- ~ .tl:f .Q I raise one's hat hat ACC raiSe

Both of them have the case marked by postpositional particle ":a-", and their corresponding English case label is "obj" for both of them. These case frames have no difference in their syntactic level. However, the nouns in these case slots, '11111-f-"(hat) and "IUj.[.F.fE"(cry), have completely disjoint semantic labels. So the disjointness of semantic labels of these nouns is useful as the clue to discriminating these case frames. Disjointness of seman­ tic labels are calculated after generalizing those semantic labels to a predetermined level in the thesaurus. In the case of BGH, currently the next level from T is regarded as that predetermined level, which consists of five cate­ gones.

3.2 Method for Discriminating Multiple Senses of A Verb

When acquiring case frames of a Japanese verb VJ, we start with a set S of bilingual feature descriptions that have VJ at 'pred' label. The problem of discriminat­ ing multiple senses of one Japanese verb VJ is regarded as the problem of discriminating two bilingual feature descriptions in which the corresponding senses of the verb VJ are different from each other. To illustrate the method for discriminating multiple senses of the verb VJ, we define the case description (CD) in the form of the feature description :

A case description is a part of a bilingual feature de­ scription that includes the information necessary for the subsequent acquiring process. For each pair of feature labels (le, PJ) in a bilingual feature description, a case description is extracted.

As we have shown in the previous section, the clues to discriminating two bilingual feature descriptions are found in the corresponding English verb Ve, English case labels le1, .. . , lem and semantic labels of the Japanese nouns in case slots. When extracting a case description from a bilingual feature description, the semantic label of the case slot is generalized to the predetermined level in the thesaurus. In the case of BGH, the predetermined level is the next level from T . This means that the se­ mantic label SEM in the case description is a subset of the set of five topmost semantic categories.

By comparing case descriptions , discrimination of case frames of bilingual feature descriptions is done through the following heuristics and human interaction. Then the set S of bilingual feature descriptions is divided into subsets 51 , ... , S1 .

Heuristics

The system finds the pair of case descriptions which can be the clue to discriminating multiple senses of the Japanese verb VJ by some heuristics. We take into con­ sideration two case descriptions of Japanese verb VJ that have the same Japanese case marker PJ . The clue is found from the case descriptions based on the discussion in the previous section, which are expressed in terms of case descriptions as follows .

• (a) Two case descriptions CD1 and CD2 have different English verbs.

• (b) CD1 and CD2 have the same English verb but have different English case labels for the Japanese case marker PJ·

• (c) CD1 and CD2 have the same English verb and the same English case label for PJ, but their semantic labels in case slots are mutually disjoint . In the case of BGH, this difference of semantic labels is detected among the five topmost semantic categories.

Human Interaction

After the system finds a pair of case description C D1 and CD2 that can be the clue, it asks the human in­ structor if the pair is really the clue. The question is presented with CD1 , CD2 and a few sample sentences involving bilingual feature descriptions. The human in­ structor selects one of the following three answers .

Ansl: Clue

CD1 and CD2 represent different cases and they must be included in different case frames of the Japanese verb V1 .

Ans2: Co-occur

Although C D 1 and C D2 represent different. cases, they can co-occur in a sentence.

For example, a Japanese verb "tf <" has two cases both of which are marked by a postpositional par­ ticle ""C'" and are translated into different English preposition "in" and "with ".

• "!Ji:.~fi "" -e:&< I write in English English !!! write
• ~ -e:&< I write with brush brush with write

These two cases are different. cases of "i!r <" and can co-occur in a. sentence. Thus they cannot be the clue to discriminating the case frames.

Ans3: Equivalent

Although C D 1 and C D2 are detected as the clue, they should be considered as equivalent cases. For example, the two case frames of the Japanese verb "jt 7.> " below should be equivalent although they have different English verbs.

• il!J!:i® ~ Je,J., I see a mov1e movie A CC see
• g;t,g. ~ Je,J., I watch a game game ACC watch

In heuristics (a),...., (c), the clues a.re found and asked to the human instructor in the order of( c), (b), (a) . This is because the clues are more likely to become equivalent in this order and consequently this order reduces the total number of questions .

'

Dividing the Set of the Bilingual Feature Descriptions into Distinct Subsets

From the results of the human interaction, some case descriptions are judged as equivalent. With this equiva­ lence relation, the set of all case descriptions is divided into equivalent classes , ECD1 , ... , ECDp . Besides, from the results of the human interaction, a set of the clues has been obtained. A clue is represented as a pair of case descriptions , (CD1 ,CD2 ) . A clue can be extended to a pair of ECDs , (ECD1 , ECD2) , where ECD1 in­ cludes CD1 and ECD2 includes CD2. Then the set of the clues is represented as a set of pairs of ECDs.

An expanded clue (ECD1 , ECD2 ) defines an incom­ patible relation between bilingual feature descriptions. The set S of all bilingual feature descriptions is then di­ vided into maximal compatible subsets sl, .. . , sl mak­ ing use of this incompatible relation 2 ·

4 Constructing Case Frames of A Verb

After S is divided into S1 , . .. , S1, the system con­ structs a case frame F1·; for each S; . In general, a surface case frame of a Japanese verb VJ is represented as the following feature description:

SR1, .. . , SR.n represent the semantic restrictions of the case slots. Each SRi represents the possible semantic categories of the case slot and is a set of the semantic categories of the thesaurus. However, in this section, we represent the case frame Fr; as a set of (ECD, SR), where EC D is an equivalent class of case descriptions which appeared in any of the bilingual feature descrip­ tions in S; and SR is the semantic restriction , that is,

Each SRi is calculated from SEM1 , .. . , SEMN, which are the semantic labels of all case elements of Pi con­ tained in S;, where Pi is the case marker which appears in ECDj. 3 Each SEMr(1 ~ r ~ N) is a set of these­ mantic categories of the nominal thesaurus. We assume that SRi is represented by semantic categories less gen­ eral than a predetermined upper bound, as described in Figure 3. In the case of BGH, currently the upper bound is the five topmost categories. SRi is meant to be the least general semantic category that covers at least one semantic category in every SEMr(1 ~ 7' ~ N). If the least general semantic category is not determined uniquely, SRi becomes a set of the least semantic cate­ gories.

2 Though this division is not always unique, bilingual fea­ ture descriptions that belong to more than one compatible sets are regarded as ambiguous and are neglected.

5 Experiment and Evaluation

We have obtained bilingual feature descriptions of 16 verbs and acquired surface case frames from them. The following example shows the process of acquiring case frames of a Japanese verb "Jt -J " from 48 bilingual fea­ ture descriptions. Case descriptions of "Jt -J " collected from the set of bilingual feature descriptions are shown in Table 1. Among them, the ~stem finds clues to discrim­ inating multiple senses of "Ji -J" by heuristics, and asks human if they are appropriate as in Table 2. In the case of "~ -J ", the English verbs "buy", "incur" and "ap­ preciate" become the clue to sense discrimination. An example of questions and answers is like below.

Then three case frames of "Ji -J" are acquired as in Ta­ ble 3. (The semantic restriction is described by the five topmost semantic categories in BGH for simplicity.) For the comparison of the acquired case frames with those of hand-made case frames, we show the case frames of ".lt-J" in IPAL dictionary (IPA Le2:icon of the Japanese Language fo1· compute1·s) [IPA, 1987] in Table 4. IPAL has case frames of 861 basic Japanese verbs and uses about 20 semantic markers to describe semantic restric­ tions . "Ji -J" has four case frames in IPAL. Case frame

3 Note that all case descriptions in an EC D have a com­ mon Japanese case marker.

1 and 2 represent the sense of "buy" and their difference is the "~::. (for)" case. In our corpus, the "~::. (for)" case does not appear. Although some cases in case frames of IPAL do not appear in the acquired case frames, case frames are discriminated almost in the same way.

In Table 5, results of experiments for 16 Japanese verbs are shown . Numbers in the column A and I are those of acquired case frames and case frames in IPAL. X 1\ -.Y represents the number of case frames in X that are not contained in Y . X > Y represents that sev­ eral case frames in X correspond to one case frame in Y . In this column, n-1 means that n case frames cor­ respond to one case frame . "Eg." is the number of ex­ amples, i.e., the number of bilingual feature descriptions used to acquire case frames . For some verbs, there are case frames which exist in IPAL but not acquired. This is because our corpus does not contain enough examples. (The poorest one is "il' It 1.>" , for which 28 case frames couldn't be acquired.) On the other hand, among ac­ quired case frames , 12 case frames don't exist in IPAL. In IPAL, idiomatic usages of verbs don't have their own entries and are included in the entries of literal usages. In our experiment, idiomatic usages are discriminated from literal ones and have their own entries. 6 out of 12 case frames are idiomatic usages.

Some acquired case frames correspond to one case

Table 4: Case Frames of "Ji -J " in IPAL

I

I I I

frame in IPAL. Observed from the results of these ex­ amples, it looks desirable to discriminate some of case frames in IPAL into finer-grained ones . For example, for the case frames of "-:::>~t ~", IPAL contains "]] 1: JJ 'S:-:::> It ~"(put an edge on a sword) and "J!!. EE 'S: -:::>It ~"(give some reason) in one case frame which has the meaning of " add or attach something to something". In our experiments, these two usages are discriminated. On the other hand, some case frames in IPAL corre­ spond to one acquired case frame . The reason is that the system couldn't find the clue to discriminating mul­ tiple senses of a verb by the heuristics. This is partly caused by the nominal thesaurus we are using. For ex­ ample, it is hard for the system to discriminate the two examples, "ilf (HUMAN ACTIVITIES) 'S: 920 ~ (know the rumor))" and "A~O)'i59J (HUMAN ACTIVITIES) :>S:-92Q ~ (know the bitterness of life)" . "920~" means "acquire some information or knowledge from some source" in the former and "understand the essence of things" in the lat­ ter. But the object nouns in both sentences belong to the semantic category HUMAN ACTIVITIES. And also for the verbs "t ~" and "~-:::>", it happens that some case frames in IPAL correspond to one acquired case frame much more often than other verbs. This is because the Japanese verb can be translated into one English verb in several senses (e.g. "take" for "t ~" and "have" for "~-:::>"). In these cases, it is hard for the system to find effective clues to sense discrimination of a verb.

Although some difficulties might exist, the results sug­ gest that it is quite possible to acquire case frames of verbs from bilingual corpora if enough translation ex­ amples and a suitable nominal thesaurus are available . We think it possible to construct a large scale bilingual corpus in the near future.

6 Concluding Remarks

We described a method of acquiring surface case frames of Japanese verbs from bilingual corpora. In the method English information such as translated English verbs and case ~abels are indispensably useful when discriminating multiple senses of a verb. The results of experiments show that it is quite possible to acquire case frames of verbs from bilingual corpora.

If the corpus contains some irregular sentences such as metaphoric expressions or erroneous sentences ac­ quired case frames may become incorrect. We guess 'that some metaphoric expressions are extracted as distinct case frames. But, it will be shown by our future work. It is necessary to develop techniques for acquiring the most probable case frames from noisy examples. The ex­ periments has shown that much larger bilingual corpus is necessary for case frame acquisition of enough number of verbs. We have collected translated newspaper editorial articles of about 21,000 sentences and are currently de­ veloping a robust system to analyze them and to extract corresponding verb phrases from which case frames are acquired. We believe that the proposed method is also applicable to acquiring case frames of English verbs and translation patterns of verbs between two languages.