We propose a thesaurus of predicates that can help to resolve pre-editing and/or post-editing problems in machine translation environments. It differs from earlier approaches such as conventional dictionaries in that we are aiming to link a wide range of near-synonyms and paraphrases. We are compiling such similar examples through both introspection and the use of translation data, giving us a large collection of monolingual and bilingual equivalences. This thesaurus enables the following machine translation techniques.
| (a) | Unification of synonymous expressions in the source language (source language paraphrasing). | |
| (b) | Conversion of homonymous expressions to more easily translated ones (source language rewriting). | |
| (c) | Development of expressions appearing in the target language into various expressions (target language paraphrasing). |
If we view machine translation from the user's viewpoint, the utilization of this technology can be classified into information assimilation, information dissemination, and computer aided translation, each with different requirements (Boitet 2001). To date, researchers of machine translation have maintained the position of uniformly requiring pre-editing and post-editing. Users of machine translation, however, have considered this to be a serious obstacle. We first attempt an easy adjustment of these requirements and tolerances since they are quite dependent on the utilization of the technology.
An example of information assimilation is the case of viewing a thesis or homepage in a foreign language. Machine translation can be used here in place of directly reading the foreign language if the user's burden is mitigated, but it is desirable for the translation quality to be as high as possible. With information assimilation, however, it is impossible to require pre-editing. In addition, in a number of cases, no post-editing is performed; however, since the user might act spontaneously when the situation calls for action, the tolerance level is wide. Many of today's commercial translation systems are considered to fit this description to some extent.
As information dissemination, we can think of using machine translation to reduce the translation time in order to accommodate, for instance, flash news reports. There are several system implementation examples corresponding to this case, but the number is small. Naturally, if we were to presuppose post-editing for this type, the translation time could not be reduced; we therefore need a high translation sentence quality that does not require just post-editing. A small amount of pre-editing may be tolerable. However, the system construction generally adopted by the development side, i.e., a structure that allows the user to give feedback in the pre-editing process depending on the accuracy of the translation result, might not be well received from the viewpoint of shortening the translation time. We believe that the actual system construction should allow the system to point out hard-to-accept expressions in advance.
This problem might be caused by the fact that the requirements for computer aided translation have not been fully examined. MT researchers dislike having to set tacit presumptions on the utilization of the user well informed about both the input language and the output language. Because of this, it is the user who is required to carry out both the pre-editing and post-editing. However, a user well informed about both languages, like a system that a translator employs, is almost completely limited to the translation memory type or the dictionary search tool. When machine translation can be used, the translation unification of special terms, etc., can be considered as the substitution of a dictionary search tool.
There are several reasons why translators do not use machine translation, but from the viewpoint of reducing labor, at minimum text should be translated correctly in paragraph units without pre-editing. If this requirement were satisfied, labor savings might be possible in terms of the entire translation process, even if word replacement or word paraphrasing as post-processing were carried out toward synonymous expressions (in order to prevent the translations from becoming monotonous) and the translations were rearranged. However, if this requirement cannot be satisfied (i.e., the translation quality requirement is higher than that of information dissemination), it is more efficient at this point to translate everything by oneself. Contrary to intuitive expectations on the development side, this may be the most difficult research problem.
Incidentally, if we cross examine these forms of utilization and, for example, use a thesaurus of these expressions, we believe we can solve one of the problems.
Since the start of research in this field, studies have been done on improving the efficiency and automating the pre-edit and post-edit. Methods tested on the pre-editing side include methods employing specialists, methods creating manuals from know-how and utilizing them toward restricted language targets, and methods that automatically carry out pre-editing. A well-known method on the post-editing side attempts to systematically reflect changes in the translations and expressions in the whole document. Although these methods differ in the steps and procedures that they employ, we can say that as a rule they do try to extract the peculiarities of the translation system.
Under this strategy, individual correspondences are indispensable for every translation system. That being the case, however, it is difficult to find an all-purpose technique. Automation, as the opposite idea, might be possible provided we had a method to prepare various synonymous expressions beforehand, and carry out matching and translation on everything that the translation system can accept.
If we think about post-editing by translators from the viewpoint of computer aided translation, although the contents would be treated variably, computer aided translation by computer could still be used in many cases if the translations were handled one item at a time. As mentioned above, in order to prevent a sentence from becoming monotonous, a translator might perform replacement or paraphrasing toward synonymous expressions. A lot of translators have independent thesauri (much-treasured) prepared individually based on the accumulation of experience, and these translators use these thesauri and replace or paraphrase words and phrases. Computer aided translation would allow the influence on expressions by noun substitution to be done locally and more easily, although verb substitution would often take a lot of time since syntax would also need to be corrected. Also, even though thesauri of words are substantial, there are only fragmentary thesauri that include syntax of verbs.
Viewed in this light, we might be able to give solid replies to the above demands if we could cover as many diverse predicate expressions as possible and use a thesaurus containing these synonymous relations. The authors view the real problem as the lack of thesauri covering structures as opposed to the numerous thesauri covering words. As one means of achieving sentence structure replacement, we consider directing our effort toward the development of a thesaurus containing sentence structures. WordNet (Fellbaum 1998) has a very useful thesaurus of English verbs, and most of the verbs have example sentences, but there is not a lot of syntactic information: words are linked, not structures.
As a realistic approach, we use the Japanese-English Paraphrase Corpus (Shirai, et al., 2001), and we test our reconstruction by using sets of expressions by assuming the corpus' predicate expressions as indices. Below, we explain how we collected comprehensive and various example sentences, and the thesaurus is generated from them by adding the predicate index for each sentence.
Lexical resources already exist where basic Japanese-English predicate frames are paired together. For Japanese and English, 14,000 Japanese-English basic patterns are given in "Goi-Taikei: A Japanese Lexicon" (Ikehara, et al., 1997). This has been motivated by the improving Japanese-to-English machine translation quality, and Fujita & Bond (2002) are expanding them in their ongoing project. Here, the term pattern describes a verb, adjective, or noun-copula, along with its arguments (mainly noun-postposition combinations). Even so, problems still remain that need to be addressed, such as the coverage of the types of expressions and the restrictions on the use of each expression, the diversity in the types of expressions able to express the same meanings, and the description of the pattern constraints (Shirai, et al., 1998).
The coverage problem is caused by characteristic differences between dictionaries designed for human use and dictionaries designed for machine use. A somewhat limited Japanese-English dictionary uses words and usages having above-average usage frequencies. This is a common design measure for a human use dictionary. However, such a dictionary perhaps intentionally excludes words and usages of comparatively lower usage frequencies. When humans use the dictionary, they obtain words and usages suitable for their purposes by performing trial and error, i.e., they change and reword what they want to say using different expressions until the target words or usages are suitable. It is very difficult to achieve the same kind of mechanism in computer processing, and accordingly, attempts have been made to comprehensively record the words and usages for machine-targeted dictionaries. Examinations are continuing in order to improve the coverage of basic patterns by the collection and abstraction of examples (Shirai, 1999).
The problem of diversity lies in the uniformity of translations: the same expression will always be translated in the same way. This is both an advantage and disadvantage of machine translation. Another cause is that the correspondences of Japanese-English basic patterns are normally limited to one-to-one correspondences. This may be a result of trying to produce initial results quickly. The uniformity of translations can be a disadvantage because of the monotomy of the translated sentences. When machine translation is used as a tool, one of the post-editing processes is to diversify expressions using a thesaurus. The influence of doing so is great when substituting verbs in many cases while the influence is significantly less when substituting nouns. We believe that it would be very useful if there were a thesaurus for patterns (like thesauri of regular words) and if it also corresponded with the sentence pattern substitution in machine translation.
A proposal has been made to separate the Japanese parts and the English parts, resulting in two monolingual lexicons with a smaller linking lexicon (Baldwin, et al., 1999). In this case, the selectional restrictions on the source language would cease to be influenced by the target language equivalent, making for more natural monolingual dictionaries. This architecture makes it far easier to add more potential translations, as each new pair would just be a link, rather than a full pair of Japanese and English patterns.
The cause of the condition description problem is assumed to be that the original valency dictionary was designed for analysis purposes, and the description of the conditions was done by hand. The former, for example, abstracts a noun based on a semantic system such that "Musume-ga mago-wo umu [The daughter has given birth to a grandchild]" becomes "<person> has given birth to <person>", and "Inu-ga koinu-wo umu [The dog has given birth to a puppy]" becomes "<animal> has given birth to <animal>". Then, if we integrate both, we get "<person or animal> gives birth to <;person or animal>". Obviously, the mutual relationship between a noun of the ga case and its corresponding noun of the wo case is lost. Although there are very few problems in the acceptance of linguistic expressions with the "typical" (assuming correct sentences) analysis processing, unsuitable combinations are produced in the language generation besides the emergence of detection problems when an attempt at use is made in the detection of errors.
In terms of preparing a valency dictionary as a basic dictionary of Japanese language analysis and Japanese-English translation, importance had been placed on coverage until now. At present, we believe that the utilization of human soul-searching is effective, where analysts try to invent as many possible paraphrases as they can. This is necessary because it is not easy to obtain a large-enough corpus to obtain low-frequency words and information related to their usages. In other words, at present, such utilization seems to be an appropriate step for accumulating data since there is only fragmentary information on the diversity of expressions. Accordingly, focusing on sentences for Japanese-English translation collected as a part of improving the coverage of Japanese-English basic patterns (Shirai, 1999), we found that sufficient results are possible by assuming constrained semantic correspondences between Japanese and English sentences and attempting to collect sentences spoken in other ways for Japanese sentences. The same results were also obtained for English sentences. Below, we explain an outline of the collection method and the trial and error we used to refine this method. Then, we continue on the collection methods of the paraphrased sentences.
In the past, we aimed at improving the coverage of the valency dictionary and used example sentences by soul-searching. In the soul-searching, we often considered the possibility that the arbitrariness of the created example sentences would become problematic. However, we also believed that this arbitrariness problem would not easily occur, since our problem setting was where usages were enumerated and not where a small number of example sentences matching specific scenes were created. There was the occasional problem concerning whether or not it was possible to call a generated example sentence a natural expression. For this problem, the same person reconsidered the problematic sentence after a certain amount of time had elapsed or exhaustively carried out the work with others through mutual checking.
Below, we first show the method when carrying out implementation aiming at improving the coverage, and then show the current method aiming at improving the diversity.
First, we covered various usages by soul-searching in the form of example sentences and decided to consider them in two steps to abstract the example sentences. This was because our final aim was to improve the coverage of the valency dictionary despite the fact that it is not easy to collect abstracted sentence patterns. As a criterion of selecting a terminology for the creation of an example sentence, we separately judged whether the terminology was suitable as terminology of the modern language. Here, we chose only one dictionary and created a policy that it be used as a rough standard. At times, it was problematic to judge whether or not a generated example sentence was a natural expression. Concerning this point, the same person reconsidered the problematic sentence after a certain amount of time had elapsed or exhaustively carried out the work with others through mutual checking. We set the following conditions based on our work experiences1.
| (1) | If the predicates can be found in "Gendai Kokugo Rêkai Jiten"
(Hayashi, 1985; 1997), consider the existing words and example sentences, and
then create example sentences from imagination. Comments: While creating the example sentences, we excluded those that posed difficulty in the sentence creation process based on discussions with other example sentence composers. | |
| (2) | If there are differences in opinions between the analysts, try to
make as many example sentences as possible. Use nouns with broad meanings as
much as possible. Comments: This work was carried out by the people creating the Japanese expressions. In other words, we did not require any work where corresponding English translations differed. As a result, we allowed translated words to be the same. | |
| (3) | In creating the example sentences, look at differences in nuance
between adverbial forms and adnominal forms, i.e., do not only look at example
sentences where predicates are of the finite form. Comments: This was based on the consideration that there are idiomatic usages in the adverbial usages of adverbial forms and attributive usages of adnominal forms, and we dealt with their sets too. For example, when we make sentences for manzoku-da "be satisfied", we also add examples for manzoku-na "satisfied" (attributive) and manzoku-ni "properly", where necessary. | |
| (4) | Aim for at least two example sentences per predicate. Here, create
sentences until no more example sentences can be conceived after a certain
degree of consideration. Comments: Based on our experiences to date, if we assume the creation time of n sentences to be t, t is approximately proportional to n2. We therefore decided to stop work for a predicate if after 10 to 15 minutes no new usages could be thought of. | |
| (5) | For the example sentences that are collected, have them made into
English translations by translators so that the results are true to the
originals as much as possible and also that they are sufficiently fluent as
English (free translations are allowed to a limited extent). Comments: Based on our experiences, we asked for the cooperative work of native English translators and native Japanese translators. |
The most direct motive here is to get more than one English translation for one Japanese expression. This can also be called the paraphrasing of English expressions. However, this does not necessarily mean that several English expressions absolutely must be generated from a specific Japanese expression. Considering this, we decided to implement Japanese paraphrasing and English paraphrasing in parallel.
The concept of collecting paraphrased cases should perhaps include the collecting of synonymous expressions within the same language from some viewpoint. However, there are many cases in which other expressions cannot be easily thought of after example sentences are shown and people become dazzled by them. It is also difficult to enumerate types of viewpoints beforehand. Accordingly, here we presuppose the existence of Japanese-English translation pairs, and while we use these Japanese-English sentence pairs under constraints, that is, as we make various translation example sentences, we also collect paraphrased cases.
The paraphrasing we mention here (for example, in the case of an English sentence) is something that imitates the generation of a synonymous expression in Japanese and a re-extraction from a Japanese-English dictionary, when the system comes across a word or expression that is not in the Japanese-English dictionary. Accordingly, it is possible that a single language speaker who is not familiar with the target translation language can also help in the work. As a real problem, however, there is the fear that the synonymous agreement gradually becomes broader, i.e. the difference in the meaning expands, when different expressions that are thought of one after another are not recorded in the Japanese-English dictionary. In fact, in our current attempt, such trials were present, and the people responsible for the comprehensive example collection had to make special requests to the translators responsible for the example sentences. Our experiences have shown that it is not easy to judge subtle Japanese-English correspondences when securing coverage. In the future, we hope to improve the following condition settings based on an analysis of the current problems.
| (1) | To deal with example sentence pairs of Japanese-English translations for Japanese predicates as described in the preceding section. | |
| (2) | To have Japanese paraphrasing be carried out with the intent of attaching various Japanese translations to English, and vice versa (to have English paraphrasing be carried out with the intent of attaching various English translations to Japanese example sentences). | |
| (3) | As a principle, to create neutral expressions where special scene settings are unnecessary. |
Based on the idea of valence by Ishiwata (Ishiwata & Ogino, 1983), we began to construct a semantic valency dictionary as the base of a valency dictionary by abstracting example sentences of a somewhat limited Japanese-English dictionary. In the early version, we collected 10,000 general sentence patterns and 3,000 idiomatic sentence patterns. However, we immediately found that the frequent lack of sentence patterns was problematic in experimental evaluations. Therefore, we searched for a way of covering sentence patterns automatically. Realistically, it is not easy to obtain a sufficiently large corpus in collecting low frequency usages. Accordingly, we decided to collect various usages as example sentences by "soul-searching".
| No. of Words | Created Sentences | Paraphrases | w/o Paraphrase | Work Order and Contents | |||||||
| Jpn | Eng | Jpn | Eng | ||||||||
| Japanese verb /IPAL | 849 | 16,713 | 7,043 | 4,096 | 12,020 | 13,748 | 0 (IPAL), 1 (add), 3 (modify), 8 (paraphrase) | ||||
| Japanese verb /others | 936 | 1,883 | 0 | 0 | -- | -- | 7 (collect) | ||||
| compound Japanese verb | 2,101 | 3,701 | 1,212 | 480 | 2,487 | 3,220 | 4 (collect), 9 (paraphrase) | ||||
| -i type adjective/IPAL | 136 | 2,156 | 530 | 219 | 1,626 | 1,937 | 0 (IPAL), 2 (add), 6 (modify), 11 (paraphrase) | ||||
| -i type adjective/others | 522 | 830 | 1,561 | 1,584 | 1 | 0 | 12 (collect & paraphrase) | ||||
| -na type adjective | 1,296 | 2,356 | 621 | 440 | 1,735 | 1,915 | 5 (collect), 10 (paraphrase) | ||||
| Verbal noun (in progress) | (885) | (1,550) | (4,448) | (4,245) | (6) | (3) | 13 (collect & paraphrase) | ||||
| Total | 5,840 | 27,639 | 10,967 | 6,819 | 17,869 | 20,820 | Note: Not including verbal nouns. | ||||
We focused our attention on the IPAL dictionary (Technical Center of IPA, 1987; 1990) in which various usages for individual predicates are recorded as example sentences. We added usages of different-nuance predicates as example sentences. Next, we decided to raise the coverage of the predicates based on a Japanese dictionary, and sought standards for the selection of these words in a modern example dictionary (Hayashi, 1985). We are now continuing with the creation of example sentences targeting predicates (i.e. not recorded in the IPAL dictionary), and are working on verbal nouns2. We are also doing paraphrasing work, which was started midway through our research.
Table 1 shows the collected data as of March 2002. "Japanese verb/IPAL" deals with words among the Japanese verbs recorded in the IPAL dictionary, and "Japanese verb/others" deals with all others. The order of the work and contents of the work are shown in the comments section. Each item is equal to an amount of work of one to three years. Some of the parts were implemented in parallel. Paraphrasing verbal nouns was easier in comparison with the others because of the more specific meanings.
We compiled a thesaurus of predicates by adding the predicate index of each sentence. Sample of our thesaurus is shown in Appendix.
In this section, we explain the work history and problems in our creation of example sentences based on the impressions of the people carrying out the work.
This work was started with the aim of covering the usages of predicates. That is we were trying to create at least one example sentence for every sense of every Japanese predicate along with its English translation.
At the start, we found a lot of words to be deeply familiar in "Japanese verb/IPAL" and "-i type adjective", and we understood that colorful example sentences, i.e., 10 or more example sentences (on average) per predicate, could be created if we excluded rare exceptions. Initially, there was a delay since we had to confirm the IPA dictionary set (due to the amount of example sentences) and its usage overlaps with the created example sentences. In particular, we needed time to confirm that the IPAL adjective dictionary was thoroughly classified in terms of the meanings of words in comparison with the IPAL verb dictionary, and that the recorded example sentences dealt with detailed differences in nuance. Because of this, we could improve the degree of allowing example sentences of similar usages to overlap.
There were a lot of words with restricted usages under "compound Japanese verb" and "-na type adjective", and we therefore decided to stop at two (or even one) example sentences per predicate. On the flip side, the necessity arose to add background explanations for better conciseness, since the expressions became unnatural when we attempted to gather the reduced usages. Obviously, when an analyst feels unnaturalness, it is typical for his/her degree of sharpness to be diminished when carrying out repetitive reading, and for the resulting judgment to gradually become more difficult. In consideration of this, everyone worked to eliminate unnaturalness by carrying out mutual checking, and rechecking after intervals.
Opinions were sometimes divided on whether or not a word (before usage under "Japanese verb/other") was a modern word. For such words, we contrasted ways of speaking (something) using similar words and judged the validity by mutual checking, and we also made efforts to create example sentences within the possible ranges. In spite of this, however, we allowed exclusions due to judgments made by the people carrying out the work, since there were cases where they were not confident in the results.
We warmed to the basic idea of creating paraphrased example sentences even while performing the above work to create example sentences. However, this resulted in example sentences of "Japanese verb/others" and the work efficiency appeared higher on the side working to keep pace with comparisons to similar expressions. In addition, because we did not have concrete condition settings in terms of what standards should be used to implement the paraphrasing (which are not easy to determine), we had to assume for the time being each of the Japanese-English translation pairs to be the target of translation and then had to establish basic measures to create expressions suitable for the translation.
Under these conditions, we tested paraphrasing for "Japanese verb/IPAL" (where the example sentence creation was comparatively easier) and "Compound Japanese verb" (where the example sentence creation was comparatively more difficult). Then, we assumed the situation where Japanese natives consulted a Japanese-English dictionary once more for the Japanese-English translations and dealt with the creation of synonymous expressions close to the predicates. In this step, strict synonymy was made a requirement. This work resulted in the creation of paraphrased sentences for 1/2 to 1/3 of the target sentences.
When we identically tested the paraphrasing with "-na type adjective" and "-i type adjective/IPAL", we found that the work became more difficult as only about 1/4 could be paraphrased. The cause of this might have been the lack of a sufficient analysis, but one of the more plausible causes of this was the difficulty in paraphrasing only nearby predicates. For the Japanese kare-wa jôzu ni oyogu, "He is a good swimmer." might be more appropriate than "He swims well.", but the former translation is almost never created since considerations center on the true translation for an original sentence in Japanese to English translation. Accordingly, we decided on an expanded interpretation of the basic measures targeting Japanese-English translation pairs, preferably to create paraphrased example sentences with the intent of creating translated sentences.
With "-i type adjective/others", we created Japanese example sentences, gave multiple English translations to them, and by looking at the results, created more (other) Japanese example sentences. In this work, we created paraphrased example sentences of about two-fold the number of example sentences for basic translations. At present, we are proceeding with the creation of example sentences using "verbal nouns" under the same conditions as those of "-i type adjective/others", and are seeing about the same example sentence results as those of "-i type adjective/others".
The objective of illustrative sentence creation is as described in 3., but it has been established as a result of the above-implied trial and error. From now on, we can expect problems like the ones below to add complexity. It is particularly important now to examine validity since we have finally reached a stage with fixed condition settings in terms of the collection of paraphrased illustrative sentences. We believe that it might also be a good time to reexamine past views on valency, based on recent research results (Ishiwata, 1999).
Dissatisfaction remains with respect to the small amount of created illustrative
sentences and lack of diversity in the initial work (there is a large number of
dissatisfied workers). Although reconsiderations are being made in work
targeting the verbs and adjectives of IPAL, it is due to this work that workers
now subjectively believe that the quality of the illustrative sentences can be
improved. When they first began this task, they also felt the need to reexamine
the concept of carrying out collection by placing limits on the correct
expressions. Naturally, placing limits is meaningless even if incorrect
expressions are collected, as there are a number of misuses that are abundantly
utilized in commonly used Japanese expressions like "(?)
(mato wo
eta, literally, you got the target)" (which is considered as the mixed use of
"
(mato wo ita, literally, you shot the target)" and "
(tou wo eta, literally, you obtain the hit)". These uses should be
collected (with notes) from the standpoint of the practical use of machine
translation.
There are only a few continuous adverbial usages for verbs (e.g., "tsuide"); in many cases, neither their use as an attributive form nor their use as an inflection form has a difference semantically. In contrast, there are many commonly used relationships for adjectives including not only many continuous adverbial usages but also a large variety of attributive deterministic usages. Some people might also think that there are no usages of adjectives as inflections, but in fact confusion has reigned when illustrative sentence creation has been looked at carefully. This confusion has been brought about by the difficulty in objectively showing how a usage is not general. Because of this, reference to the Internet (along with exchanges of opinion among workers) has been used to investigate and subsequently address the existence of applicable expressions. Unfortunately, there have been a number of cases reconfirming the diversity of the expressions. The major flow of work to date is as follows: 1) selecting Japanese predicates, 2) creating a Japanese example, and 3) creating an English example. For our part, however, we want to set the Japanese translation of an English adjective or an adverb as reference, and then consider the complement of a Japanese illustrative sentence.
In the early stages, we placed emphasis on the collection of general expressions, and as a result took in expressions that should have been commonly used expressions. Comprehensively collecting commonly used expressions is by no means an easy task (including with what standard to judge common use). However, there have been many possible literal interpretations among the expressions considered to be commonly used expressions, and conversely, we have seen cases where we were not aware about the interpretations on commonly used expressions while being quite conscious about general expressions. From the viewpoint of collecting various expressions toward actual use, we believe it might be better not to establish standards on general common usages.
The number of created illustrative sentences is increasing due to largely
polysemous words, and so it is not easy to examine how comprehensive the usages
are while looking over all of the illustrative sentences. On the other hand, not
only is it not easy to create natural illustrative sentences for individual
words in itself, but the work efficiency is also poor, e.g., judgment must be
made on whether to add a background explanation depending on the situation. In
such cases as those that are at the diametrically opposite end, we believe it
might be effective to carry out some individual work support.
In Fujita, et al. (2000), a support environment targeting the paraphrasing of
nouns is proposed. We hope to reference this environment in the future.
In the early stages, we selectively proceeded with the paraphrasing of predicate portions from the viewpoint of extending our sentence construction system. However, we found that various expressions could be formed when we paraphrased groups of rank elements and words as units; we continue to gradually loosen the conditions. Recently, we have been considering that people might not really care about the restriction to only assure the correspondence of the translation, which is a major premise. We have also been thinking that it might be effective for a person working with Japanese-English translation to mutually exchange paraphrased results, and then attempt a reexamination. Our aim is to carry out the paraphrasing of already created illustrative sentences by mobilizing a number of Japanese native speakers or English native speakers (while not necessarily requiring bilingual speakers). This approach increases the degree of objectivity by a majority rule based method.
We proposed a thesaurus of predicates and introduced problems related to the present state and the present work on the illustrative sentence collection of Japanese predicates (used as the data of the material in the thesaurus). We reported that reflection based on "questionnaires" is effective when comprehensively collecting illustrative sentences applicable to various usages. We also showed that creating various translations in order to create paraphrased illustrative sentences is a powerful method in the domains of Japanese-English translation or English-Japanese translation.
Because the method introduced in this paper has been achieved while incrementally evolving how the work is performed through the accumulation of experience, there are still a number of problems that should be looked at again in the illustrative sentences created in the early stages. In addition, although there is very little collection taking place in cases like nouns serving as predicates, we want to include into the viewpoints not only cases that function like attributes but also measures toward diversity in the spoken language (Takezawa, et al., 2001), and to begin examinations from how we can narrow down target words.
Although the initial aim of translation-based illustrative sentence creation work was to improve the comprehensibility of sentence construction systems, i.e., reduce the number of unknown words in machine translation, we can expect a wider range of utilization of the illustrative sentence sets themselves by adding a diversity of viewpoints. We would also like to consider the effective use of illustrative sentence sets.
| No | Predicate | Example sentence | Remarks on J0 | |||
| J0 |  |     | Japanese verb | |||
| J1 |  |  | ||||
| E0 | success | His plan was a success. | ||||
| J0 | |     | Japanese verb | |||
| J1 |  |   | ||||
| E0 | look up | He looked up that character in the dictionary. | ||||
| J0 | |     | Japanese verb | |||
| J1 |  |  | ||||
| E0 | ask | I asked his friends about his destination. | ||||
| E1 | question | I questioned his friends about his destination. | ||||
| J0 |  |   | Japanese verb | |||
| J1 |  | | ||||
| E1 | affect | He was affected by the heat. | ||||
| J0 | |   | Japanese verb | |||
| E0 | right | My prediction was right. | ||||
| J0 | |  | Japanese verb | |||
| E0 | poison | He was poisoned by eating blowfish. | ||||
| J0 |  |     | compound Japanese verb | |||
| J1 |  |  | ||||
| E0 | swap | The athletic field was swamped with spectators. | ||||
| J0 | |    | compound Japanese verb | |||
| J1 | swell up | The place where I was stung by the bee has swollen up. | ||||
| J0 | |       | compound Japanese verb | |||
| J1 | |  | ||||
| E0 | double | The population of this city is double what it was 10 years ago. | ||||
| E1 | double | The population of this city has doubled in the last 10 years. | ||||
| J0 |  |  | -i type adjective | |||
| E0 | favorable | His attitude is favorable. | ||||
| J0 | |    | -i type adjective | |||
| E0 | want | He is not the kind of person we want in our company. | ||||
| J0 | |   | -i type adjective | |||
| J1 |  | | ||||
| E0 | desirable | Formal attire is desirable for dinner. | ||||
| J0 | |    | -i type adjective | |||
| J1 |  |  | ||||
| E0 | best | It is best to keep potatoes at room temperature. | ||||
| E1 | should | Potatoes should be kept at room temperature. | ||||
| J0 |  |   | -na type adjective | |||
| E0 | satisfy | I am satisfied with my present position. | ||||
| J0 | |     | -na type adjective | |||
| J1 |  | | ||||
| E0 | proper | I have not had a proper meal since yesterday. | ||||
| E1 | proper | I have not eaten a proper meal since yesterday. | ||||
| J0 | |   | -na type adjective | |||
| J1 |  | | ||||
| E0 | properly | He cannot even properly write the alphabet. | ||||
| J0 |  |     | verbal noun | |||
| J1 |  |  | ||||
| J2 |  |  | ||||
| E0 | overwhelm | Their attack overwhelmed the opposing team. | ||||
| E1 | overpower | Their attack overpowered the opposing team. | ||||
| E2 | swamp | Their attack swamped the opposing team. | ||||
| J0 | |    | verbal noun | |||
| J1 |  | | ||||
| J2 |  | | ||||
| E0 | overwhelm | I was overwhelmed by the scale of Niagara Falls. | ||||
| E1 | thunderstruck | I was thunderstruck by the magnificence of Niagara Falls. | ||||
| E2 | awe | I was awed by the scale of Niagara Falnnls. | ||||