SUPER-FUNCTION BASED MACHINE TRANSLATION SYSTEM FOR BUSINESS USER. Xin Zhao

Transcription

1 SUPER-FUNCTION BASED MACHINE TRANSLATION SYSTEM FOR BUSINESS USER by Xin Zhao A Dissertation Submitted to the Factulty of Engineering at Tokushima University in conformity with the requirements for the degree of Doctor of Engineering. Doctoral Course for Information Science and Intelligent System Graduate School of Engineering Department of Information Science and Intelligent Systems Tokushima, Japan September, 2005

2 Abstract In today s increasingly networked world, there is an increased need for language translation. Attempts of language translation are almost as old as computer themselves. Machine Translation (MT) is the attempt to automate all, or part of the process of translating between human languages and is one of the oldest large-scale applications of computer science. Developing a system that accurately produces a good translation between human languages is the goal of MT systems. Machine translation between closely related languages is easier than between language pairs that are not related with each other. The present work reports my attempt in developing a Super-Function (SF) based machine translation system which translate Japanese into Chinese, and proposes a shallow method to translate Japanese compound nouns by using word-level translation dictionary and target language monolingual corpus. The machine translation system consists of three main parts, responsible for analysis, SF matching and translation generation. A SF is a functional relation mapping sentences from one language to another. The core of the system uses the SF approach to translate without going through syntactic and semantic analysis as many MT systems usually do. This work focuses on business users for whom MT often is a great help if they need an immediate idea of the content of texts like messages, reports, web pages, or business letters. The translation of compound nouns is a major issue in MT due to their frequency of occurrence and high productivity. Several aspects of compound nouns make them particularly difficult to be handled in a system performing automatic translation. In this work, I also discuss the challenges of automatic translating Japanese compound nouns into Chinese in the Super-Function Based Machine Translation (SFBMT) system to address this ii

3 issue. A major design goal of this method is that it can act as a standalone module and can be very well integrated with the machine translation system for Japanese sentence. My experiment is not yet completed, but already it has displayed interesting results. By using the SF in the translation system to translate without syntactic and semantic analysis as many MT systems usually do, will enhance the translation quality and reduce the glossary quantity. Even if the method for handling compound nouns is still a shallow method, but my experimental results show that the method, if improved can perform the task of automatic compound noun translation quite well. iii

4 Acknowledgements I gratefully acknowledge many people who supported me with deepest appreciation. I would like to take this opportunity to express my gratitude to my supervisor Professor Fuji Ren, who introduced me to the fascinating world of natural language processing, for suggesting the research topic, for his generous guidance and patience given to me. His numerous support and encouragement, as well as inspiring advice are extremely essential and valuable in my research papers and this thesis. I would like to express my deepest gratitude to Professor Junichi Aoe and Professor Kenji Kita, they gave me the opportunity to entered the department of Information Science and Intelligent Systems, Faculty of Engineering, Tokushima University, and introduced me to the A-1 group, to work on the topic of machine translation and to find my own way as a researcher. I cordially thank my other supervisor, Associate Professor Shingo Kuroiwa, for his advice, patience, and linguistic corrections during my entire work. Without his effort, I will not be able to strengthen and improve my research project and papers. I would also like to show my gratitude to the Department of Information Science and Intelligent Systems, for the provision of the best equipment and pleasant office environment required for high quality research. Special thanks should be given to Professor. Dr. Stefan Voss who has given me valuable suggestions, encouragement and supports. I can not list the names of all people who I am indebted to. But I particularly would like to give my thanks to my research group members, HaiQing Hu, Manabu Sasayama, Yuhsuke Konishi, Ippei Fukuda, Kazuyuki Matsumoto, Keisuke Ueta, DaPeng Yin, Mohamed Fattah, Tomohiro Yagi, Takahiro Kuroda, Jiajun Yan, Min Shao, Yu Zhang, Jing Wang. They have given me support, and a joyful iv

5 and wonderful university life. I was also blessed with wonderful friends; in many ways, my successes are theirs, too. They gave me a lot of support and encouragement through the hard times. Thank you Yoshinori Matsuura, Tomoyo Matsuura, Qing Liu, Xin Lu, ZhenYu Jin, Furthermore I would like to thank my parents and my husband, for all their support and unconditional love, sacrifice and consideration throughout my Ph.D. v

6 Contents Abstract Acknowledgements List of Tables List of Figures ii iv ix x 1 Introduction Overview Machine Translation The Need for Translation Technology History of Machine Translation Natural Language Processing Standard Paradigm for NLP The Source Language Analysis Target Language Generation Aim of The Work Structure of The Thesis Machine Translation Automatic Translation Machine-Assisted Translation Fully-Automated Machine Translation Human-Aided Machine Translation Machine-Assisted Human Translation Translation Tripod Source Text: Specifications: Terminology: Why Machine Translation is Not Successful? Traditional Approaches Direct or Transformer Approaches Transfer-Based Approaches vi

7 2.3.3 Interlingual Approach Other Approaches Rule-Based Approach Knowledge-Based Approach Corpus-Based Approach Example-Based Approach Statistical Approach Summary Super-Function Based Machine Translation How to Do Translation? Pre-Editing Translation Process Post-Editing Super-Function SF Definition SF Format SF Architecture Directional Graph Transformation Table Robust Super-Function Process of SFBMT Morphological Analysis Super-Function Matching Morphological Agreement Comparing with Other Related Research Pattern-Based Machine Translation Glossary-Based Machine Translation Example-Based Machine Translation Summary Japanese-Chinese SFBMT for Business Users Proposed MT System Review of Business Letter Outline of The System Experiment and Evaluation Examples Evaluation Experiment Summary Translating Compound Nouns in SFBMT Overview The Problems and Proposed Solution Intractable Problems of Compound Nouns Proposed Shallow Solution vii

8 Generation Stage Selection Stage Used Resources and Evaluating Results Used Resources Evaluating Results Related Works Summary Conclusions and Future Works Conclusions Future Works References 116 viii

9 List of Tables 3.1 Example of a Node Table Example of an Edge Table Examples of Translation Pattern United Table Result of the Evaluation Experiment Example of Translation Templates Results of Alignment Result of Aligned Case Examples of Aligned Case Examples of Partially Aligned Case Examples of Unaligned Case Examples of No Template Found ix

10 List of Figures 2.1 Human-Aided Machine Translation Translation Tripod Kind of Translation Approaches Syntax Structure Translation and Semantic Structure Translation Traditional Approaches of MT Direct Approach Transformer Architecture Components of Transfer system Interlingual Architecture Component of Interlingual System Flow of Translation Flowchart of Translation process Example of DG Architecture of PBMT Process of GBMT Architecture of EBMT Outline of The System Translation Interface Flowchart of The Processing Example of The Processing x

11 Chapter 1 Introduction Machine Translation, or Automatic Translation is the attempt to automatic all, or part of the process of translating from one human language to another. This chapter will briefly sketch some background on machine translation, a brief history of machine translation, and some basic concepts of natural language processing. The idea is to give a clear basic understanding of the state of the art. Then the aim of the current work and the structure of this thesis are summarized. 1.1 Overview Machine translation of natural languages, commonly known as MT, has multiple personalities. First of all, it is a venerable scientific enterprise, a component of the larger area of studies concerned with the studies of human language understanding capacity. Indeed, computer modeling of thought processes, memory and knowledge is an important component of certain areas of linguistics, philosophy, psychology, neuroscience and the field of Artificial Intelligence (AI) within computer science. MT promises the practitioners of these sciences empirical results that could be used for corroboration or refutation of a variety of hypotheses and theories. But MT is also a technological challenge of the first order. It offers an opportunity for software designers and engineers to dabble in constructing very complex and large-scale non-numerical systems and for field computational linguists, an opportunity to test their understanding of the syntax and semantics of a variety of languages by encoding this vast, though rarely comprehensive, knowledge into a form suitable for processing by computer programs. The study of natural language has been an important area of artificial intelligence almost since the beginning of the field. Two main goals motivate AI work on natural language. One is 1

12 the theoretical goal, and close to that of the linguist, namely, to discover how we use language to communicate. The other is technological goal, namely, to enable the intelligent computer interfaces of the future, where natural language becomes an important means for man-machine interaction. Luckily, progress toward one of these goals often is progress toward the other a better theoretical understanding leads to more robust systems, and a better understanding of processing issues in actual applications suggests new goals and techniques of theoretical interest. The ultimate solution to language understanding must wait until we can effectively model almost all aspects of human intelligence. Many applications, however, do not require full conversational capabilities or encyclopedic knowledge. For instance, a natural language interface that serves as a query language to a database need only focus on questions and can limit the language it understands to concepts that arise in the database. On the contrary, MT is the attempt to automate all, or part of the process of translating from one human language to another. It is one of the oldest large-scale applications of computer science. 1.2 Machine Translation People who need documents translated often ask themselves whether they could use a computer to do the job. When a computer translates an entire document automatically and then presents it to a human, the process is called Machine Translation. When a human composes a translation, perhaps calling on a computer for assistance in specific tasks such as looking up specialized words and expressions in a dictionary, the process is called Human Translation. There is a gray area between human and machine translation, in which the computer may retrieve whole sentences of previously translated text and make minor adjustments as needed. However, even in this gray area, each sentence was originally the result of either human translation or machine translation. We will reserve the label machine translation for the case when both the initial translation of the sentences and subsequent manipulations are performed by a computer[42]. The original proposals covered only the making of a straightforward dictionary translation from the Source Language (SL) to the Target Language (TL). It is convenient to start by seeing how this simple objective may be achieved on a machine whose primary purpose is only the most rudimentary machine functions in order to perform machine translation: 1. The machine has a large memory. 2. The input typewriter sends data, either direct to the memory, or to a register provided with 2

13 subtraction facilities, the accumulator register. 3. The machine contains a conditional transfer order which enables the machine to select between alternative courses of action according to the sign of the number held in the accumulator register. 4. The contents of the accumulator can be typed at the output. The reader familiar with modern automatic digital computers will see that all of the above functions are present in all such computers existing, with the exception in many cases of the large memory. There are some popular misconception about MT, we will discuss them in turn: MT is a waste of time because you will never make a machine that can translate Shakespeare. The criticism that MT systems cannot, and will never, produce translations of great literature of any great merit is probably correct, but quite beside the point. It certainly does not show that MT is impossible. First, translating literature requires special literary skill it is not the kind of thing that the average professional translator normally attempts. So accepting the criticism does not show that automatic translation of non-literary texts is impossible. Second, literary translation is a small proportion of the translation that has to be done, so accepting the criticism does not mean that MT is useless. Finally, one may wonder who would ever want to translate Shakespeare by machine it is a job that human translators find challenging and rewarding, and it is not a job that MT systems have been designed for. The criticism that MT systems cannot translate Shakespeare is a bit like criticism of industrial robots for not being able to dance Swan Lake. Generally, the quality of translation you can get from an MT system is very low. This makes them useless in practice. Far from being useless, there are several MT systems in day-to-day use around the world. Examples include METEO (in daily since 1977 use at the Canadian Meteorological Center in Dorval, Montreal), SYSTRAN (in use at the CEC, and elsewhere), LOGOS, ALPS, EN- GSPAN (and SPANAM), METAL, GLOBALINK. It is true that the number of organizations that use MT on a daily basis is relatively small, but those that do use it benefit considerably. For example, as of 1990, METEO was regularly translating around words of 3

14 weather bulletins every day, from English into French for transmission to press, radio, and television. In the 1980s, the diesel engine manufacturers Perkins Engines was saving around 4000 on each diesel engine manual translated (using a PC version of WEIDNER system). Moreover, overall translation time per manual was more than halved from around 26 weeks to 9-12 weeks this time saving can be very significant commercially, because a product like an engine cannot easily be marketed without user manuals. Of course, it is true that the quality of many MT systems is low, and probably no existing system can produce really perfect translations. 1 However, this does not make MT useless. First, not every translation has to be perfect. Imagine you have in front of you a Chinese newspaper which you suspect may contain some information of crucial importance to you or your company. Even a very rough translation would help you. Apart from anything else, you would be able to work out which, if any, parts of the paper would be worth getting translated properly. Second, a human translator normally does not immediately produce a perfect translation. It is normal to divide the job of translating a document into two stages. The first stage is to produce a draft translation, i.e. a piece of running text in the target language, which has the most obvious translation problems solved (e.g. choice of terminology, etc.), but which is not necessarily perfect. This is then revised either by the same translator, or in some large organizations by another translator - with a view to producing something that is up to standard for the job in hand. This might involve no more than checking, or it might involve quite radical revision aimed at producing something that reads as though written originally in the target language. For the most part, the aim of MT is only to automate the first, draft translation process. 2 MT threatens the jobs of translators. The quality of translation that is currently possible with MT is one reason why it is wrong to think of MT systems as dehumanizing monsters which will eliminate human translators, or enslave them. It will not eliminate them, simply because the volume of translation to be performed is so huge, and constantly growing, and because of the limitations of current and 1 In fact, one can get perfect translations from one kind of system, but at the cost of radically restricting what an author can say, so one should perhaps think of such systems as (multilingual) text creation aids, rather than MT systems. The basic idea is similar to that of a phrase book, which provides the user with a collection of canned phrase to use. This is fine, provided the canned text contains what the user wants to say. Fortunately, there are some situations where this is the case. 2 Of course, the sorts of errors one finds in draft translations produced by a human translator will be rather different from those that one finds in translations produced by machine. 4

15 forseeable MT systems. While not an immediate prospect, it could, of course, turn out that MT enslaves human translators, by controlling the translation process, and forcing them to work on the problems it throws up, at its speed. There are no doubt examples of this happening to other professions. However, there are not many such examples, and it is not likely to happen with MT. What is more likely is that the process of producing draft translations, along with the often tedious business of looking up unknown words in dictionaries, and ensuring terminological consistency, will become automated, leaving human translators free to spend time on increasing clarity and improving style, and to translate more important and interesting documents editorials rather than weather reports, for example. This idea borne out in practice: the job satisfaction of the human translators in the Canadian Meteorological Centerimproved when METEO was installed, and their job became one of checking and trying to find ways to improve the system output, rather than translating the weather bulletins by hand (the concrete effect of this was a greatly reduced turnover in translation staff at the Center). The Japanese have developed a system that you can talk to on the phone. It translates what you say into Japanese, and translates the other speaker s replies into English. The claim that the Japanese have a speech to speech translation system, of the kind described above, is pure science fiction. It is true that speech-to-speech translation is a topic of current research, and there are laboratory prototypes that can deal with a very restricted range of questions. But this research is mainly aimed at investigating how the various technologies involved in speech and language processing can be integrated, and is limited to very restricted domains (hotel bookings, for example), and messages (offering little more than a phrase book in these domains). It will be several years before even this sort of system will be in any sort of real use. This is partly because of the limitations of speech systems, which are currently fine for recognizing isolated words, uttered by a single speaker, for which the system has been specially trained, in quiet conditions, but which do not go far beyond this. However, it is also because of the limitations of the MT system. Against these misconceptions, we should place the genuine facts about MT. MT is useful. The METEO system has been in daily use since As of 1990, it was regularly translating around words daily. In the 1980s, The diesel engine manufacturers Perkins Engines was saving around 4000 and up to 15 weeks on each manual translated. 5

16 While MT systems sometimes produce howlers, there are many situations where the ability of MT systems to produce reliable, if less than perfect, translations at high speed is valuable. In some circumstances, MT systems can produce good quality output: less than 4% of ME- TEO output requires any correction by human translators at all (and most of these are due to transmission errors in the original texts). Even where the quality is lower, it is often easier and cheaper to revise draft quality MT output than to translate entirely by hand. MT does not threaten translators jobs. The need for translation is vast and unlikely to diminish, and the limitations of current MT systems are too great. However, MT systems can take over some of the boring, repetitive translation jobs and allow human translation to concentrate on more interesting tasks, where their specialist skills are really needed. Speech-to-Speech MT is still a research topic. In general, there are many open research problems to be solved before MT systems will be come close to the abilities of human translators. Not only are there are many open research problems in MT, but building an MT system is an arduous and time consuming job, involving the construction of grammars and very large monolingual and bilingual dictionaries. There is no magic solution to this. In practice, before an MT system becomes really useful, a user will typically have to invest a considerable amount of effort in customizing it. The correct conclusion is that MT, although imperfect, is not only a possibility, but an actuality. But it is important to see the product in a proper perspective, to be aware of its strong points and shortcomings. While there have been many variants, most MT systems, and certainly those that have found practical application, have parts that can be named for the chapters in a linguistic textbook. They have lexical, morphological, syntactic, and possibly semantic components, one for each of the two languages, for treating basic words, complex words, sentences and meanings. Each feeds into the next until the last one in the chain produces a very abstract representation of the sentence. There is also a transfer component, the only one that is specialized for a particular pair of languages, which converts the most abstract source representation that can be achieved into a corresponding abstract target representation. The target sentence is produced from this essentially by reversing the analysis process. Some systems make use of a so-called interlingua or intermediate language, in which case the transfer stage is divided into two steps, one translating a source sentence into the 6

17 interlingua and the other translating the result of this into an abstract representation in the target language. Machine Translation started out with the hope and expectation that most of the work of translation could be handled by a system which contained all the information we find in a standard paper bilingual dictionary. Source language words would be replaced with their target language translational equivalents, as determined by the built-in dictionary, and where necessary the order of the words in the input sentences would be rearranged by special rules into something more characteristic of the target language. In effect, correct translations suitable for immediate use would be manufactured in two simple steps. This corresponds to the view that translation is nothing more than word substitution (determined by the dictionary) and reordering (determined by reordering rules). Reason and experience show that good MT cannot be produced by such delightfully simple means. As all translators know, word for word translation doesn t produce a satisfying target language text, not even when some local reordering rules (e.g. for the position of the adjective with regard to the noun which it modifies) have been included in the system. Translating a text requires not only a good knowledge of the vocabulary of both source and target language, but also of their grammar the system of rules which specifies which sentences are well-formed in a particular language and which are not. Additionally it requires some element of real world knowledge knowledge of the nature of things out in the world and how they work together and technical knowledge of the text s subject area. Researchers certainly believe that much can be done to satisfy these requirements, but producing systems which actually do so is far from easy. Most effort in the past 10 years or so has gone into increasing the subtlety, breadth and depth of the linguistic or grammatical knowledge available to systems. In growing into some sort of maturity, the MT world has also come to realize that the text in translation out assumption the assumption that MT is solely a matter of switching on the machine and watching a faultless translation come flying out was rather too naive. A translation process starts with providing the MT system with usable input. It is quite common that texts which are submitted for translation need to be adapted (for example, typographically, or in terms of format) before the system can deal with them. And when a text can actually be submitted to an MT system, and the system produces a translation, the output is almost invariably deemed to be grammatically and translationally imperfect. Despite the increased complexity of MT systems they will never within the forseeable future be able to handle all types of text reliably and accurately. This normally means that the translation will have to be corrected (post-edited) and usually the person best equipped to do this is a translator. 7

18 MT will only be profitable in environments that can exploit the strong points to the full. As a consequence, we see that the main impact of MT in the immediate future will be in large corporate environments where substantial amounts of translation are performed. The implication of this is that MT is not (yet) for the individual self-employed translator working from home, or the untrained lay-person who has the occasional letter to write. This is not a matter of cost: MT systems sell at anywhere between a few hundred pounds and over It is a matter of effective use. The aim of MT is to achieve faster, and thus cheaper, translation. The lay-person or self-employed translator would probably have to spend so much time on dictionary updating and/or postediting that MT would not be worthwhile. There is also the problem of getting input texts in machine readable form, otherwise the effort of typing will outweigh any gains of automation. The real gains come from integrating the MT system into the whole document processing environment, and they are greatest when several users can share, for example, the effort of updating dictionaries, efficiencies of avoiding unnecessary retranslation, and the benefits of terminological consistency. 1.3 The Need for Translation Technology Advances in Information Technology (IT) have combined with modern communication requirements to foster translation automation. The history of the relationship between technology and translation goes back to the beginnings of the Cold War, as in the 1950s competition between the United States and the Soviet Union was so intensive at every level that thousands of documents were translated from Russian to English and vice versa. However, such high demand revealed the inefficiency of the translation process, above all in specialized areas of knowledge, increasing interest in the idea of a translation machine. Although the Cold War has now ended, and despite the importance of globalization, which tends to break down cultural, economic and linguistic barriers, translation has not become obsolete, because of the desire on the part of nations to retain their independence and cultural identity, especially as expressed through their own language. This phenomenon can clearly be seen within the European Union, where translation remains a crucial activity. The Internet with its universal access to information and instant communication between users has created a physical and geographical freedom for translators that was inconceivable in the past. IT has produced a screen culture that tends to replace the print culture, with printed documents being dispensed with and information being accessed and relayed directly through computers ( , databases and other stored information). These computer documents are instantly available 8

19 and can be opened and processed with far greater flexibility than printed matter, with the result that the status of information itself has changed, becoming either temporary or permanent according to need. Over the last two decades we have witnessed the enormous growth of information technology with the accompanying advantages of speed, visual impact, ease of use, convenience, and costeffectiveness. At the same time, with the development of the global market, industry and commerce function more than ever on an international scale, with increasing freedom and flexibility in terms of exchange of products and services. The nature and function of translation is inevitably affected by these changes. There is the need for countries to cooperate in many spheres, such as ecological (Greenpeace), economic (free trade agreements) humanitarian (Doctors without Borders) and educational (exchange programs), etc. Despite the importance of English, there is the commonly-held belief that people have the right to use their own language, yet the diversity of languages should not be an obstacle to mutual understanding. Solutions to linguistic problems must be found in order to allow information to circulate freely and to facilitate bilateral and multilateral relationships. Thus different aspects of modern life have led to the need for more efficient methods of translation. At the present time the demand for translations is not satisfied because there are not enough human translators, or because individuals and organizations do not recognize translation as a complex activity requiring a high level of skill, and are therefore not prepared to pay what it is worth. In other words, translation is sometimes avoided because it is considered to be too expensive. In part, human translation is expensive because the productivity of a human being is essentially limited. Statistics vary, but in general to produce a good translation of a difficult text a translator cannot process more than 4-6 pages or 2,000 words per day. The economic necessity of finding a cheaper solution to international exchange has resulted in continuing technological progress in terms of translation tools designed to respond to the translator s need for immediately-available information and non-sequential access to extensive databases. The social or political importance of MT arises from the socio-political importance of translation in communities where more than one language is generally spoken. Here the only viable alternative to rather widespread use of translation is the adoption of a single common lingua franca, which (despite what one might first think) is not a particularly attractive alternative, because it involves the dominance of the chosen language, to the disadvantage of speakers of the other languages, and raises the prospect of the other languages becoming second-class, and ultimately disappearing. Since the loss of a language often involves the disappearance of a distinctive culture, and a way of thinking, this is a loss that should matter to everyone. So translation is necessary for communication for ordinary human interaction, and for gathering the information one needs to 9

20 play a full part in society. Being allowed to express yourself in your own language, and to receive information that directly affects you in the same medium, seems to be an important, if often violated, right. And it is one that depends on the availability of translation. The problem is that the demand for translation in the modern world far outstrips any possible supply. Part of the problem is that there are too few human translators, and that there is a limit on how far their productivity can be increased without automation. In short, it seems as though automation of translation is a social and political necessity for modern societies which do not wish to impose a common language on their members. The commercial importance of MT is a result of related factors. First, translation itself is commercially important: faced with a choice between a product with an instruction manual in English, and one whose manual is written in Japanese, most English speakers will buy the former and in the case of a repair manual for a piece of manufacturing machinery or the manual for a safety critical system, this is not just a matter of taste. Secondly, translation is expensive. Translation is a highly skilled job, requiring much more than mere knowledge of a number of languages, and in some countries at least, translators salaries are comparable to other highly trained professionals. Moreover, delays in translation are costly. Estimates vary, but producing high quality translations of difficult material, a professional translator may average no more than about 4-6 pages of translation (perhaps 2000 words) per day, and it is quite easy for delays in translating product documentation to erode the market lead time of a new product. It has been estimated that some 40-45% of the running costs of European Community institutions are language costs, of which translation and interpreting are the main element. This would give a cost of something like 300 million per annum. This figure relates to translations actually done, and is a tiny fraction of the cost that would be involved in doing all the translations that could, or should be done. 3 Scientifically, MT is interesting, because it is an obvious application and testing ground for many ideas in Computer Science, Artificial Intelligence, and Linguistics, and some of the most important developments in these fields have begun in MT. To illustrate this: the origins of Prolog, the first widely available logic programming language, which formed a key part of the Japanese Fifth Generation programme of research in the late 1980s, can be found in the Q-Systems language, originally developed for MT. Philosophically, MT is interesting, because it represents an attempt to automate an activity that can require the full range of human knowledge that is, for any piece of human knowledge, it 3 These estimates of CEC translation costs are from Patterson (1982) [55]. 10

21 is possible to think of a context where the knowledge is required. For example, getting the correct translation of negatively charged electrons and protons into French depends on knowing that protons are positively charged, so the interpretation cannot be something like negatively charged electrons and negatively charged protons. In this sense, the extent to which one can automate translation is an indication of the extent to which one can automate thinking. Despite this, very few people, even those who are involved in producing or commissioning translations, have much idea of what is involved in MT today, either at the practical level of what it means to have and use an MT system, or at the level of what is technically feasible, and what is science fiction. In the whole of the UK there are perhaps five companies who use MT for making commercial translations on a day-to-day basis. In continental Europe, where the need for commercial translation is for historical reasons greater, the number is larger, but it still represents an extremely small proportion of the overall translation effort that is actually undertaken. In Japan, where there is an enormous need for translation of Japanese into English, MT is just beginning to become established on a commercial scale, and some familiarity with MT is becoming a standard part of the training of a professional translator. 1.4 History of Machine Translation Machine translation has recently celebrated its 50th birthday. This is a short life span for a science, but in that period remarkable progress has been made, mirroring the advances in the contributing disciplines of computer science and linguistics. It is possible to trace ideas about mechanizing translation processes back to the seventeenth century, but realistic possibilities came only in the 20th century. In the mid 1930s, a French- Armenian Georges Artsrouni and a Russian Petr Troyanskii applied for patents for translating machines. Of the two, Troyanskii s was the more significant, proposing not only a method for an automatic bilingual dictionary, but also a scheme for coding interlingual grammatical roles (based on Esperanto) and an outline of how analysis and synthesis might work. However, Troyanskii s ideas were not known about until the end of the 1950s. Before then, the computer had been born. Soon after the first appearance of electronic calculators research began on using computers as aids for translating natural languages. The beginning may be dated to a letter in March 1947 from Warren Weaver of the Rockefeller Foundation to cyberneticist Norbert Wiener. Two years later, Weaver wrote a memorandum (July 1949), putting forward various proposals, based on the wartime successes in code breaking, the developments by Claude Shannon in information the- 11

22 ory and speculations about universal principles underlying natural languages. Within a few years research on MT had begun at many US universities, and in 1954 the first public demonstration of the feasibility of machine translation was given 4. Although using a very restricted vocabulary and grammar it was sufficiently impressive to stimulate massive funding of MT in the United States and to inspire the establishment of MT projects throughout the world. The earliest systems consisted primarily of large bilingual dictionaries where entries for words of the source language gave one or more equivalents in the target language, and some rules for producing the correct word order in the output. It was soon recognized that specific dictionarydriven rules for syntactic ordering were too complex and increasingly ad hoc, and the need for more systematic methods of syntactic analysis became evident. A number of projects were inspired by contemporary developments in linguistics, particularly in models of formal grammar, and they seemed to offer the prospect of greatly improved translation. Optimism remained at a high level for the first decade of research, with many predictions of imminent breakthroughs. However, disillusion grew as researchers encountered semantic barriers for which they saw no straightforward solutions. There were some operational systems the Mark II system 5 installed at the USAF Foreign Technology Division, and the Georgetown University system at the US Atomic Energy Authority and at Euratom in Italy but the quality of output was disappointing. By 1964, the US government sponsors had become increasingly concerned at the lack of progress; they set up the Automatic Language Processing Advisory Committee (AL- PAC), which concluded in a famous 1966 report that MT was slower, less accurate and twice as expensive as human translation and that there is no immediate or predictable prospect of useful machine translation. It saw no need for further investment in MT research; and instead it recommended the development of machine aids for translators, such as automatic dictionaries, and the continued support of basic research in computational linguistics. Although widely condemned as biased and short-sighted, the ALPAC report brought a virtual end to MT research in the United States for over a decade and it had great impact elsewhere in the Soviet Union and in Europe. However, research did continue in Canada, in France and in Germany. Within a few years the Systran system was installed for use by the USAF (1970), and shortly afterwards by the Commission of the European Communities for translating its rapidly growing volumes of documentation (1976). In the same year, another successful operational system appeared in Canada, the Meteo system for translating weather reports, developed at Montreal University. 4 A collaboration by IBM and Georgetown University 5 Which developed by IBM and Washington University 12

23 In the 1960s in the US and the Soviet Union MT activity had concentrated on Russian- English and English-Russian translation of scientific and technical documents for a relatively small number of potential users, who would accept the crude unrevised output for the sake of rapid access to information. From the mid-1970s onwards the demand for MT came from quite different sources with different needs and different languages. The administrative and commercial demands of multilingual communities and multinational trade stimulated the demand for translation in Europe, Canada and Japan beyond the capacity of the traditional translation services. The demand was now for cost-effective machine-aided translation systems that could deal with commercial and technical documentation in the principal languages of international commerce. The 1980s witnessed the emergence of a wide variety of MT system types, and from a widening number of countries. First there were a number of mainframe systems, whose use continues to the present day. Apart from Systran, now operating in many pairs of languages, there was Logos (German-English and English-French); the internally developed systems at the Pan American Health Organization (Spanish-English and English-Spanish); the Metal system (German-English); and major systems for English-Japanese and Japanese-English translation from Japanese computer companies. The wide availability of microcomputers and of text-processing software created a market for cheaper MT systems, exploited in North America and Europe by companies such as ALPS, Weidner, Linguistic Products, and Globalink, and by many Japanese companies, e.g. Sharp, NEC, Oki, Mitsubishi, Sanyo. Other microcomputer-based systems appeared from China, Taiwan, Korea, Eastern Europe, the Soviet Union, etc. Throughout the 1980s research on more advanced methods and techniques continued. For most of the decade, the dominant strategy was that of indirect translation via intermediary representations, sometimes interlingual in nature, involving semantic as well as morphological and syntactic analysis and sometimes non-linguistic knowledge bases. The most notable projects of the period were the GETA-Ariane (Grenoble), SUSY (Saarbrucken), Mu (Kyoto), DLT (Utrecht), Rosetta (Eindhoven), the knowledge-based project at Carnegie-Mellon University (Pittsburgh), and two international multilingual projects: Eurotra, supported by the European Communities, and the Japanese CICC project with participants in China, Indonesia and Thailand. In early 1990s, the end of the decade was a major turning point. Firstly, a group from IBM published the results of experiments on a system (Candide) based purely on statistical methods. Secondly, certain Japanese groups began to use methods based on corpora of translation examples, i.e. using the approach now called example-based translation. In both approaches the distinctive 13

24 feature was that no syntactic or semantic rules are used in the analysis of texts or in the selection of lexical equivalents; both approaches differed from earlier rule-based methods in the exploitation of large text corpora. A third innovation was the start of research on speech translation, involving the integration of speech recognition, speech synthesis, and translation modules the latter mixing rule-based and corpus-based approaches. The major projects are at ATR (Nara, Japan), the collaborative JANUS project (ATR, Carnegie-Mellon University and the University of Karlsruhe), and in Germany the government-funded Verbmobil project. However, traditional rule-based projects have continued, e.g. the Catalyst project at Carnegie-Mellon University, the project at the University of Maryland, and the ARPA-funded research (Pangloss) at three US universities. Another feature of the early 1990s was the changing focus of MT activity from pure research to practical applications, to the development of translator workstations for professional translators, to work on controlled language and domain-restricted systems, and to the application of translation components in multilingual information systems. These trends have continued into the later 1990s. In particular, the use of MT and translation aids (translator workstations) by large corporations has grown rapidly a particularly impressive increase is seen in the area of software localisation (i.e. the adaptation and translation of equipment and documentation for new markets). There has been a huge growth in sales of MT software for personal computers (primarily for use by non-translators) and even more significantly, the growing availability of MT from on-line networked services (e.g. AltaVista, and many others). The demand has been met not just by new systems but also by downsized and improved versions of previous mainframe systems. While in these applications, the need may be for reasonably good quality translation (particularly if the results are intended for publication), there has been even more rapid growth of automatic translation for direct Internet applications (electronic mail, Web pages, etc.), where the need is for fast real-time response with less importance attached to quality. With these developments, MT software is becoming a mass-market product, as familiar as word processing and desktop publishing. 1.5 Natural Language Processing Natural Language Processing (NLP) is both a modern computational technology and a method of investigating and evaluating claims about human language itself. Some prefer the term Computational Linguistics in order to capture this latter function, but NLP is a term that links back 14