MULTI-LEVEL MACHINE TRANSLATION METHOD FOR COMMUNICATION WITH TRANSLATION

	Satoru Ikehara NTT (Japan)
S. Shirai, K. Ogura, A. Yokoo, H. Nakaiwa and T. Kawaoka

Abstract

According to the progress of a borderless society, communication services with machine translation to facilitate social interaction are strongly needed. However, current MT technologies are not satisfactory to realize such services. This paper clarified the types of communication services with MT together with the quality of translation needed for such services. It also pointed out the difference of the requirements for MT methods based on the classification of language groups. And, taking up the case of Japanese to English translation as one of the pair of far different language groups, the problems of conventional MT technologies were clarified and Multi-Level Translation Method was proposed from the view point of Constructive Process Method of languages. This method is based on semantic analysis supported by large semantic dictionaries which are written by precise semantic attribute systems with 3,000 semantic categories. From me experiments performed to translate newspaper articles, it was shown that this method could improve the success rate of translations up to 80% from 30% of conventional methods which had placed importance on a syntactic processing. The 20% of translations that failed are difficult even for the technology of semantic analysis and require meaning understanding technologies based on the world knowledge. The achieved success rate satisfies the translation quality required to begin communication services with machine translation for information providing.

INDEX

Progress in the field of telecommunication technologies has overcome the communication barriers of distance and time facing mankind. The major remaining problem is the language barrier. To support the ever-increasing volume of international communications, the realization of communications with machine translatio services (CWMT services) has become the major issue ^(1-3) . CWMT services transcend the conventional scope of communication services which are limited to the transfer of signals. CWMT deals with the contents of a message and constitutes a new value added communication service.

Since the 1980s, MT technologies have rapidly progressed and the various systems has been developed ^(4,5) . Yet good quality machine translation between far different type of languages such as Japanese and English is still unrealized. Many actual MT systems require manual processes such as rewording the ohginal text into a form more readily translatable (pre-editing) and/or rewnting the translation results (post-editing). Thus, conventional MT systems are not satisfactory to realize the CWMT services that can be easily accessible by the general public.

When considering the difficulty of translation, it depends on the types of both the source and the target languages. Translating greatly different languages such as Japanese and English is much harder than translating similar languages such as Japanese and Korean.

This paper first discusses the success rate (the correctly understandable rate of translation for each sentence) required for the realizations of a MT system without pre-editing. Second, in order to perform the required translation quality for the translation between different type of languages, it proposes a new multi-lingual translation method based on the use of representative languages. In this method, languages close to each other are collected into discrete three language groups, namely an agglutinative language group, an inflection language group and an isolated language group. A representative language is selected for each group. Thus translations are classified into two types, translations between representative languages for each language group and translations between languages within the same group.

Third, it proposes a Multi-Level Translation method (MLT method) based on the Constructive Process Theory of languages (Tokieda Grammar) for Japanese to English MT as one of the translations between far different language groups. This method is based on the semantic analysis supported by large semantic dictionaries. These are written by precise semantic attribute systems which are composed of 3,000 attributes for common nouns, proper nouns and verbs.

Applying MLT method to translate newspaper articles, the success rate of translations will be discussed whether it is satisfactory or not to the realization of CWMT services between different types of language groups.

Depending on the media used, CWMT seIvices can be classified into three types: character code communications, character image communications (FAX communication) and voice communications. The last two types depend on the technologies of recognition and synthesis.

Synthesis technologies for voice and characters image are considered to have reached at the applicable level. However recognition technologies are yet far from such a level that is required for realizing comprehensive CWMT services. Therefore we consider CWMT on texts here.

When performing MT on texts, the type of source texts should be taken into account. The translation quality depends on the text types such as colloquial style and literary style and also the types of documents such as letters, manuals, theses, and newspapers. Here we limit the source texts to the literary style and examine the communication seivices which are expected to be popular in the near future

One of the most important area of international information services is business world. In this world, many kinds of information such as economics, exchange, industry, business results, goods, credits, stocks, banking, bonds, politics and so on are produced every day. Immediate distribution of these information are highly desired. And the correctness of translation is also important to put them into practical use.

Since strictly correct translation cannot be expected in MT, it is necessary to study the quality of usable MT for communication services with translation. In order to find the quality required for information providing services, two types of usages such as how a user can find the information necessary for himself (information gathering) and how correctly he can understand the details of them (information understanding).

(1)Translation for Information Gathering

In this case, how quickly information can be obtained is more important than how the translation is correctly performed. Concerning the qualify of translation, it can be said that at least the understandable outlines of the articles are required for deciding the value of them.

The zero to ten evaluation method ⁽⁶⁾ has been proposed for evaluating translation quality. In this method, a score of six or more points indicates successful translation in that the original meaning could be correctly understood. Experiments were conducted using this method to find the relation between me grade of text understanding and the translation quality for each sentence in a text. We found that when the translation score of at least 70% of the sentences was 6 or more, the text could be understood in general*1. These results indicate that CWMT services for newsflash (information gathering) type information must achieve the translation success rate of 70% to be successful.

(2)Translation for Information Understandlng

Since the perfect translation can not be expected in MT, post-editing of the translation results by human translators is required to get the strictly understandable results. From the studies ⁽⁷⁾ of the profitable conditions that the use of MT results have advantage compared with human translation, it was known that more than 80% of the success rates of translations are required for information providing to different language regions. Thus, it is considered that at least 80% of the translation quality will be required for information understanding.

Considering from these discussions, it can be said that the translation quality of 80 or more will satisfy the requirement for communication services of information providing type.

The interlingual translation method ⁽⁸⁾ which uses an artificial intermediate language has been proposed for multi-lingual translation. However, no successful system has been developed for widely different languages.

The difficulty of MT depends on the difference between the source and the target languages. Translating greatly different languages such as Japanese and English is much harder than translating similar languages such as Japanese and Korean. For example, it is considered that the limitation of translation quality will be 30 % in the case of Japanese to English MT based on syntax analysis. In contrast, in the case of Japanese to Korean translation ⁽⁹⁾ , it is known that over 90% of the translation quality can be obtained even based on the word to word translation method.

Natural language reflects mental processes in the act of communication. The way of thinking depends on social culture. Differences in culture have caused the difference in languages. It is difficult to unify widely disparate languages.

We propose a new multi-lingual translation method based on the use of representative languages. The method is shown in Fig. 1. In this method, translation is conducted as follows. First, languages close to each other are collected into discrete language groups. This paper introduces three language groups: the agglutinative language group, the inflection language group and the isolating language group. Next, a representative language is selected for each group. There are two types of translations.

<Translation Type-A> Translation between representative languages. Deep analysis is generally required. <Translation Type-B> Translation between languages within a group. Relatively shallow analysis leads to good results.

Conventional translation technologies appear suitable only for the type B translations. As mentioned above, the success rate of conventional Japanese to English MT systems is about 30 % for newspaper translations. Thus, more than twice the current translation quality is required. Progress can be made by considering why conventional translation methods fail to realize type A translations.

Up to now, researches into natural language processing (NLP) has been conducted using mainly the results of the computational linguistics ⁽¹⁰⁾ such as a generative grammar and a phrase structure analysis. Most MT systems that were developed in Japan in the 1980s were founded on these theories. The interlingual method mentioned before uses the dualism of syntax and semantics.

Natural languages have many aspects that does not fit well to logical methods. For instance, irregularity and incompleteness of natural language expressions cause difficulties to conventional theories. Especially, irregularity of the Compositional Semantics in natural languages can not be set aside. In the semantics of conventional logics, the principle of the Compositional Semantics is used to represent the meaning of the total expression. This principle assumes that the overall meaning is the combination of the meanings of the expression's components. Unfortunately this assumption does not hold in natural languages.

(1) The Concept of the ConstructIve Process Theory

Starting with the Constructive Process Theory ^(11,12) of languages, we propose a new approach to language processing with the goal of practical MT. The language model of the Constructive Process Theory is shown in Fig.2.

Generative grammar considers that linguistic expressions are generated by transformation from a deep structure which is meaning. On the contrary, the Constructive Process Theory accepts that languages are complex bodies similar to the nature and are composed of three components: objects, the speaker's conception of them, and expressions. This theory explains a grammar as the set of rules spontaneously developed by a society. Grammar is used to link a speaker's conception to expressions. The meaning of an expression is the relationship between the speaker's conception of the object and the expression used. From this theory, we propose the following two ideas.

(2) Distinction of Rules and Meanings

Past research often misunderstood the relation between rules and meanings. The so called word meanings usually registered in dictionaries should be regarded as word usages, namely the linguistic rules. Meanings appear only when the rules are used in expressions. Here, we separate the conventional meaning processing into two processes as shown in Fig.3. One is the process of semantic analysis which specifies the rule used in constructing the actual expression. The other process is called meaning understanding which specifies the relations between expressions and objects in the world.

(3) Introduction of Structural Meaning Unit

The syntax rules are not necessarily independent from those of words. There are also many cases in which the expression structure have meaning in cooperation with word meanings. From the view point of abstracting the expression structure, we extract expression units as the structural units of meaning.

Based on the idea mentioned above, we proposed a Multi-Level Translation Method (MLT method) supported by semantic analysis. And the Japanese to English MT System, ALT-J/E ^(13,14), was developed to evaluate the effects of the method. The relation between the theoretical background of ALT-J/E, translation method, linguistic knowledge systems and prominent functions achieved by them are described in Fig.4.

The structure of the Multi-level Translation Method is shown in Fig.5. It is mainly featured by the following two points.

(1) Separation of Expressions

Expressions in the source text are separated into two kinds of expressions; subjective expressions and objective expressions. Subjective expressions express the speaker's emotions and intentions directly. On the other hand, objective expressions express the conceptualized object world. Subjective expressions are translaled into the target language using reference tables.

(2) Abstraction of Patterns

Objective expressions are translated into the target language through transfer rules. Transfer rules are prepared in advance so as to abstracting patterns without changing their meanings. Pattern abstraction is performed in accordance with the strength of the structure. Currently these patterns are classified into 3 groups and are registered into dictionaries that reflect differing degrees of abstraction.

The Multi-level Translation Method is mainly supported by the semantic analysis not by the meaning understanding. In this section, the necessity of semantic knowledge for the semantic analysis will be discussed and the outline of knowledge systems prepared for ALT-J/E will be described.

(1) Linguistic Knowledge for Semantic Analysis

The conventional natural language processing has been developed on the premise of the Compositional Semantics and on the assumption that linguistic rules can be represented by a relatively small number of rules except for word dictionaries. Researchers avoided considering the detailed contents of each linguistic rule and concentrated on trying to build a processing framework. This approach seems not appropriate for natural language processing; it is necessary to build a linguistic knowledge system for the realization of the semantic analysis mentioned beforehand. Both the general knowledge such as common sense and professional knowledge based on a world model are necessary to understand the meanings.

Each word and every expression have historically disparate backgrounds. Moreover, there is no set of relations that is common to all nations. We need to consider the contents of each rule as the research target to create a linguistic knowledge system^*1 that realizes natural language processing applicable to the real world.

(2) Linguistic Knowledge in ALT-J/E

Within this system the linguistic knowledge systems shown in Fig.6 were developed.

• Description Language for Dictionaries
  
  In order to describe semantic usages of words, semantic attribute systems (3,000 attributes) were developed. The semantic attribute systems are comprised of three subsystems: the common noun semantic attribute system (2,800 attributes), the proper noun semantic attribute system (200 attributes), and the verbal semantic attribute system (100 attributes). Semantic features of words are defined by sets of attribute names and their values. Semantic features of words are defined by attribute names.
  
  
• Syntactic and Semantic Dictionaries
  
  Linguistic knowledge needed for Japanese to EngliSh MT is compiled in semantic dictionaries. The semantic dictionaries are composed of a semantic word dictionary (400,000 words) and a semantic structure dictionary (15,000 patterns). Abstracted Japanese sentence patterns are written using the semantic attribute system and registered in the latter dictionary with corresponding English sentence patterns.

Language processing rules were also written by using the above mentioned syntactic and semantic attribute systems. The rules were applied to translations in cooperation with the linguistic knowledge described by the same attributes. This framework makes it easy to develop various new translation functions.

This section shows some of the new translation functions realized by the semantic analysis based on the semantic dictionaries.

(1) Differentiating Translations for Verbs

One Japanese verb usually corresponds to more than one expression in English. For example, the Japanese verb "掛ける(kakeru)" can be translated in more than 80 ways. Experiments showed that 2,000 more classes of semantic attributes are needed ⁽¹⁷⁾ if we are to successfully differentiate the translation of Japanese verbs. This problem has been solved by constructing the semantic attribute system with the attributes of 3,000.

(2) Supplementation of Elements Ellipsis

Japanese writers refrain from writing what readers are assumed to understand. Specifically, subjects and objects are often omitted. Successful translation, therefore, demands that these elements are recovered from the context. Our experiments ⁽¹⁵⁾ into translating newspaper articles by ALT-J/E showed that 95% of subject and object ellipsis could be coneedy supplemented automatically.

(3) Automatic Rewriting of Source Text

Although the automation of pre-editing has been desired for a long time, it was difficult because of undesirable side effects. With ALT-J/E, it is possible to judge whether a expression can be rewritten or not without changing the meaning using the automatic rewriting rules developed by the semantic attribute systems. Experiments ⁽¹⁶⁾ showed that if me original success rate is about 50%, the improvement in translation success rate is about 20% with automatic rewriting perfonned by ALT-J/E.

The experiments performed to translate newspaper articles found that originally omitted translation rules can easily be added without causing conflicts among the rules in the system such that the success rate of translations can be improved up to 80%. The 20% of translations that failed are difficult even for the technology of semantic analysis and require meaning understanding technologies based on world knowledge.

The suceess rate obtained in this experiment (80%) is more man the twice of that of the conventional method (30%). We are now able to realize CWMT services for information providing services mentioned in section 2.2.

This paper has proposed the Multi-Level Translation(MLT) method to realize communication with machine translation (CWMT) services .

First, it was shown that the success rate of 70% or more was necessary for realization of such services. in order to perform this quality of translation between languages from far different language groups, the MLT method based on the Constructive Proeess Theory of languages was proposed. This method was based on the semantic analysis which is highly supported by the large and precise semantic dictionaries. Some of the prominent functions have been shown from the many new functions which had been realized in this system using the semantic analysis technologies.

The experiments were conducted applying ALT-J/E system which was developed by this method to the translation of newspaper articles. The experiments showed that the success rate of translations was improved up to 80% by the new method, This rate is more than the twice of that of the conventional method (30%) which place emphasis on the syntactic analysis developed from the Compositional Semantics. It exceeds the translation quality of 70% required to begin CWMT services for information gathering and satisfies 80% required for information providing services.

The 20% of translations that failed are difficult even for the technology of a semantic analysis and require meaning understanding technologies based on the world knowledge.

(1)

Y. Sakama, S. Ikehara: Communication with Translation (in Japanese), NTT Technical Journal, pp.69-71 (1991)

(2)

S. Ikehara: Toward thgh Quality Machine Translation(In Japanese), NTT Teehnieal Journal, pp.44-47 (1992)

(3)

Y. Sakai: Applications of Natural Language Processing to Communications (in Japanese), JTT 15-3, pp.2-4(1992)

(4)

Proceedings of Machine Translation SUMMIT 1-4 (1987, 1989, 1991 and 1993)

(5)

J. Carbonell et.al.: JTEC Panel Report on Machine Translation in Japan, Loyala College in Maryland (1992)

(6)

S. Ikehara: Language Characteristics in MT Evaluation, MT Evaluation Workshop (1992)

(7)

T. Ehara: Study of Conditions for Usable Machine Translations, AAMT Journal, No.10, pp.17-19 (1995)

(8)

H. Uchida and M. Zhu: Interlingua for Multilingual Machine Translation, Proeeedings of MT-SUMMIT IV, pp.157-169 (1993)

(9)

T. Kim and S. Ura: Korean Expression Generation Method using Semantic Connection in Japanese to Korean Machine Translation (in Japanese), Trans. IPS Japan, Vol.33, No.12, pp.1578-1588 (1992)

(10)

N. Chomsky: Aspects of Theory of Syntax, MIT Press, Cambridge, Mass. (1965)

(11)

M. Tokieda: Basic Theory of Language (in Japanese), Iwanami Bookstore, Tokyo (1941)

(12)

T. Miura: Theory of Languages and Recognitions, Vol.1-3 (in Japanese), Keiso-Bookstore, Tokyo (1967)

(13)

S. Ikehara: Multi-Level Machine Translation System, Future Computer Systems, Vol.2, No.3, pp.261-274 (1989)

(14)

S. Ikehara, M. Miyazaki, S. Shirai and Y. Hayashi: Speaker's Recognition and Multi-Level Translation Method based on It (in Japanese), Trans. IPS Japan, Vol.28, No.12, pp.1269-1279(1987)

(15)

N. Nakaiwa and S. Ikehara: Zero Pronoun Resolution in a Japanese to English Machine Translation System Using Verbal Semantic Attributes, Proeeedings of the 3rd Conference on Applied Natural Language Proeessing, PP.201-208 (1992)

(16)

S. Shirai, S. Ikehara and T. Kawaoka: Effects of Automatic Rewriting of Source Language within a Japanese to English MT System, Proeeedings of the Fifth International Conference on Theoretical and Methodological Issues in Machine Translation, pp.226-239 (1993)

(17)

S. Ikehara, M. Miyazaki and A. Yokoo: Semantic Dictionaries for Maehine Translations, Trans. IPS Japan, VoL.34, No.8, pp.1692-1704 (1993)

S.Ikehara received Master's degree and Ph.D. from Osaka University in 1969 and 1983 respectively. He joined ECL (Electrical Communication Laboratories) of NTT (Nipponn Telegraph and Telephone Public Corporation) in 1969 and he is currently the leader of Ikehara Research Group of NTT-CS Labs (Communication Science Laboratories).

S.Shirai received Master's degree from Osaka University in 1980. He joined ECL of NTT in 1980 and he is currently a senior researcher of the NTT-CS Labs.

K.Ogura received Master's degree from Keio University in 1980. He joined ECL of NTT in 1980 and he is currently a senior researcher of the NTT-CS Labs.

A.Yokoo received Master's degree from the University of Electro-Communications in 1982. He joined ECL of NTT in 1982 and he is currently a senior researcher of the NTT-CS Labs.

H.Nakaiwa received Master's degree from Nagoya University in 1987. lIe joined ECL of NTT in 1987 and he is currently a senior researcher of the NTT-CS Labs.

T.Kawaoka received Master's degree and Ph.D. from Osaka University in 1968 and 1984 respectively. He joined ECL of NTT in 1968 and he is currently a professor of Doshisha Universify.