Machine Translation –A Rosetta stone for the 21th century? Maria Hedblom Kognitionsvetenskap 2 Ht 2010 Abstract Machine Translation(MT) is a sub-genre in Artificial Intelligence that deals with automatic translations between different languages. Historically it has been research for roughly the last century but has been idealised since the 1700th century when Descartes presented the idea of a universal language. There are several different problems facing the translation process, linguistic problems such as differences in grammar, word construction and ambiguous words. But also problems concerning the context in a text, how metaphors and anecdotes are to be translated. To attempt to solve these problems there are different types of MT's. All more or less derived from the more traditional approaches such as Direct MT and Transfer system MT. While Direct MT is a direct translation word for word with a minimum of grammar rules, Transfer system is composed by a large rule book and a simple dictionary. Simplified one can say that Direct MT turned into the Corpus based MT and that Transfer MT today is more referred to as Rule based MT. Knowledge based MT is one of the most common forms of the Rule based MT and example based MT is one of the most common Corpus based MT. Another very common translation system is the Statistical MT where probability rules such as Baye's rule of the Expectation Maximization Algorithm are used. The final version of MT that I have chosen to research is the Hybrid MT. A combination of several or all the MT's above. Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 Table of content Abstract......................................................................................................... 1. Introduction............................................................................................2 1.1 Purpose..................................................................................................................................2 1.2 Word definition and comments.............................................................................................2 1.3 Why Machine Translation?...................................................................................................3 1.4 Introducing Machine Translation..........................................................................................4 2. Problems with Machine Translation.....................................................5 2.1 Metaphors and Anecdotes.....................................................................................................6 2.2 Ambiguity and Fertility.........................................................................................................7 3. Traditional Machine translation............................................................8 3.1 Direct MT.............................................................................................................................8 3.2 Interlingua system ................................................................................................................9 3.3 Transfer system MT............................................................................................................10 4. Modern Machine Translation...............................................................11 4.1 Rule based MT....................................................................................................................11 4.2 Corpus based MT................................................................................................................12 4.3 Statistic based MT..............................................................................................................13 4.4. Hybrid MT ........................................................................................................................14 5. Discussion............................................................................................15 6. References............................................................................................17 1 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 1. Introduction 1.1 Purpose Part of the examination in the course Artificiell Intelligens II(Artificial Intelligence II) with course code 729G11, was to research a subject within Artificial Intelligence and to write an essay of the knowledge learnt. The purpose of this rapport is therefore to increase, first and foremost, my own knowledge in this particular subject but I hope that others who read this may find it helpful to understand the basics in Machine Translation. The rapport will focus on trying to explain some of the problems of Machine Translation, it is different forms, and how their artificial intelligence makes them unique amongst the others. The reason why I have chosen to study Machine Translation is because I am interested in how different language affect us humans and how Machine Translation can help us understand some of these differences both linguistically and culturally. 1.2 Word definition and comments Abbreviations: MT Machine Translation KBMT Knowledge based Machine Translation RBMT Rule based Machine Translation SMT Statistical based Machine Translation HMT Hybrid Machine Translation AI Artificial Intelligence EM Expectation Maximization 2 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 How to read it: Language[Word/sentence] This indicates that the translation of the words or sentence is in the defined language. Word/sentence Important term Other: René Descartes French philosopher, scientist and military living between 15961650. Known among other things for his involvement in rationalism and the theory that knowledge is achieved from thinking.1 Rosetta Stone A basalt stone, dating from 196 BC, with an inscription found in the Egyptian city Rosetta in 1799. The inscription is written in three languages: Greek, Egyptian demotic hieroglyphs and Egyptian hieroglyphs. Mostly known for its influence on deshiffering the Egyptian hieroglyphs.2 Star Trek A science fiction tv-show, movie series and world wide phenomena. 1.3 Why Machine Translation? Thanks to modern technology such as the Internet, cell-phones and a highly developed infrastructure we can easily have conversations with people all over the world, do business between nations that speak different languages and travel just for the fun of it. The Internet, online literature and efficiency in reaching as many as possible with a certain message, are only few of several reasons why Machine Translation(MT) is so important. It can help us overcome the barriers that different languages create and increase the communication between nations and therefore also the welfare of the world. Unfortunately this is not quite as easy as one would like, to translate a text from one language to another, a reason to why this research is still very current and ongoing. Machine Translation has proven to be very useful within several areas and I will mention 1 The Philosophy Net, http://www.thephilosophynet.com/descartes.htm, 2010-09-26 2 Nationalencyklopedin, http://www.ne.se/rosettestenen. 2010-09-24 3 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 a few of these to help you understand how useful software such as this can be. The first, perhaps unnecessary, area to mention is the possibility to make a Literary translation from a text. This has been found to be very difficult, and is still not possible for all texts, but it is rather easy to produce a Rough translation of a text using a MT. This is very useful when you, for example, look at a foreign web page and wish to get a gist of the contents.3 Another area is Restricted-source translation. This means that a text concerning a certain subject is translated. An example that has proved to be rather successful is the METEO, a translating system made by the TAUM group at the University of Montreal that is been in use since the 1977's4, to translated weather rapports from English to French.5 The last area, that I will mention, that MT have been successful in is Preedited translation: a human writes a text based on the rules the MT in question uses, in other words a restricted language, to make it easier to translate this text to different languages afterwards. These restricted languages are often called Caterpillar English because the first one to try this translation model was Caterpillar Corp. when they tried to make the translation of manuals more efficient.6 1.4 Introducing Machine Translation In 1799 French officers found a stone, dating from 196 BC, in the Egyptian city of Rosetta. A stone with an inscription written in three different languages. The content of the stone inscription was irrelevant but it did contribute with the final piece of the puzzle to be able to interpret the Egyptian Hieroglyphs. Ever since then the Rosetta Stone is a symbol for translation. Today we might not find stones to help us understand other languages but instead we have invented Machine Translation, a system that allows us to do as little as possible and still understand other languages. But what is Machine Translation? “The job of a translator is to render in one language the meaning expressed by a passage of text in another language.”7 For MT this is preferable done with as little of, if any at all, human interaction as possible. 3 4 5 6 7 Russel,S, Norvig, P; Artificial Intelligence: A Modern Approach, New Jersey 2003, p.851 Chandious, http://www.chandioux.com/profile.asp 2010-09-25 Russel, p.851 Russel, p.851 Brown, P.F, Cocke, J, Della Pietra A, Della Pietra, J.V, Jelinek, F, Lafferty, J.D, Mercer, R.L, Roossin, P.S; A Statistical Approach to Machine Translation, Computational Lingustics Volume 16 (1990) 2:79-85 4 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 Even though Machine Translation has been a practical process only since the 1950's, the idea of machine translation has been alive for more than 300 years. In the 17th century René Descartes introduced the idea of a universal language. A language that in different tongues shared both the same content and symbols. Today there isn't such a wide spread language apart from perhaps the mathematical symbols. But we can find universal languages in smaller “universes”. For example the Chinese writing system, where there are many different dialects but each and every one of them uses the same writing system. English could also be considered as a form of a universal language. How did this then turn into actual Machine Translation? It wasn't until the early 20th century when the technology had advanced enough for MT to start evolving. A few pioneers, among them George Artsrouni and Petr Smirnov-Troyanskii, patented in the early 1930´s their ideas. Artsrouni had made a storage device on paper tape which could find the equivalent word in other languages. But it was Smirnov-Troyanskii's idea that had more significance to modern MT. He proposed a three staged translation step. The first, including a linguist from the source language, was to make a logical analysis turning the sentence into base forms and syntactic functions. The second step was to let a machine translate it to the other language from these base forms, and lastly another linguistic was to make the final touch to the translation.8 It may seem simple enough but it is very similar to today's MT, even though step one and three also is done by machines rather by humans. During the 1950´s and the development of computers, MT had prime years and the expectations were very high. Surprisingly little of the early MT was based on theoretical linguistics9. Instead the approaches focused on word-for-word translation and statistical methods. After a couple of decades it was found that the expectations may have been a bit overrated and the business was rather drawn back for some time. The first real successful MT after the first wave was in the 1977's when the METEO by the TAUM group at the university of Montreal.10 Today MT researchers know the limitations and instead of trying to annihilate them fully, focus is put on trying to produce good enough translations despite the problems. 8 Hutchins, W.J; Machine Translation: A Brief History, p.3 9 Nirenburg, S; Readings in Machine Translation, 2003, p.4 10 Hutchins, p.7 5 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 2. Problems with Machine Translation As I have already mentioned there are several problems that prevents Machine Translation(MT) from being able to translate a text perfectly. Sergei Nirenburg gives a list of six elements that the MT must manage before a perfect translation is possible. The first one is the Field of discourse, the MT must be able to recognise the general subject of the text. The second is that it must be able to Recognise coherent word groups, such as idioms and compound nouns. The third is to know the Syntactic function for each word. The forth is to understand The selectional relations between words in open classes, that is, nouns, verbs, adjectives, and adverbs. The fifth is Antecedents, meaning the ability to understand how words are built and to place translated words in the right order based on previous statements in the text. These first five are more focused on a linguistic approach, with some exception, and can more or less be solved using linguistics rules and corpuses in different forms. Even though there is knowledge of where we should start to look for a solution to these problems, it is not as MT creators have all the answers, it is still very difficult. But perhaps the most challenging part is the last one on Nirenburgs list, for an MT to actually Understand the context of a text. For it to not only understand the words and the gist but to actually follow the reasoning, the underlying contents and meanings. It is likely that this is the last piece of the puzzle of MT, if it ever is solved.11 2.1 Metaphors and Anecdotes As Nirenburg said understanding the context is one of the most difficult parts of MT. Metaphors and anecdotes are two of these difficult things a MT has to translate and are therefore suitable to explore further. Machine Translation programs use different kinds of Artificial Intelligence to translate a text, which is explained later. But take a person trying to learn a language, it is a good enough simile. To help him he has got a dictionary and a grammar book. He will most likely translate English[It's raining cats and dogs] literately, which most likely won't make any sense to him, the expression does not indicate that it in fact is raining mammals. The direct translation would be Swedish[Det regnar katter och hundar] whereas a more appropriate translation would be 11 Nirenburg, p.7 6 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 Swedish[Det spöregnar] which basically means that it rains a lot. To fully understand the expression he will need someone, or something, to inform him of that this is a phrase that is not meant to be taken literately but as an expression for something else. This can be done in a large lexicon where all metaphors are listed, which probably would work pretty well even though it is both memory and time insufficient. But what about anecdotes? Anecdotes are stories that is trying to say something else than the actual contents. These are usually build on the local culture and are therefore very difficult to explain. Religious text for example is full of anecdotes. The Holy Bible's Jesus speaks more or less only in anecdotes as He delivers His message12. An MT would probably be able to give a fair enough translation, but it is likely that the real message would be misinterpreted. It is interesting to question whether a machine translator can lead to loss of, apart from the obvious context, also of culture. For example religious texts or even poetry wouldn't perhaps be the first thing you'd want a machine to translate, would it? 2.2 Ambiguity and Fertility Ambiguity means that a word or a text can have two or more meanings, of course making MT more difficult. Take for example the word English[Light], light as in either not dark or not heavy. Ambiguity comes in two forms, it is either Lexical or Structural. The example from before is a typical Lexical ambiguity, meaning that one word have one or several different meanings. If we assume a MT has knowledge over English and Swedish grammar and knows that Coca Cola is a name, a translation might look like this: English[Coca Cola light] → Swedish[Ljus Coca Cola] or Swedish[Lätt Coca Cola] This is perhaps not exactly what we're trying to express. Another example is English[The queen can't bear children ] → Swedish[Drottningen kan inte få barn] or Swedish[Drottningen står inte ut med barn] In an attempt to prevent this lexical ambiguity to ruin MT, Fertility was invented. It is a form mostly used in Statistical based MT and means that a word with fertility n is copied n times to 12 The Holy Bible, Matthew 1:1 - John 21:25 7 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 make sure that all possible translations is considered.13 Structural ambiguity is when a sentence can be interpreted wrong. For example English[The policeman killed the man with a gun] a sentence that means that the policeman used a gun to kill a man, or that the policeman killed a guy that had a gun. Neither interpretation is wrong, nor is one more correct than the other one. The only way to know which interpretation is more correct than the other is to analyse the context of the text, something a MT does not manage.14 3. Traditional Machine translation The traditional Machine Translation(MT) is what today's MT is based on, even though it evolves more and more when new ideas get into the picture. But to understand where MT is today we need to have knowledge over the original versions of MT. 3.1 Direct MT This form of MT is the most basic one. It translates the individual words in a sentence from one language to another using a two-way dictionary. To its help it uses very simple grammar rules, only the most basic and general ones. Irregular verb morphologies are often incorrect since they are bent as general verbs and often the word order lack in perfection due to the word-for-word translation system.15 Even though the MT makes many faults in the translation, the translation is usually comprehensible to a human reader. However since the translation choose the words from a statistical point of view, the most common translation first, and not depending on the content of the text, it is likely that some of the source texts meaning can be lost in the translation, especially in more advanced texts.16 The picture below is an attempt to show how Direct MT works in translation from different languages, A-D. 13 14 15 16 Russel, p.855 Bach, K; Ambiguity, http://userwww.sfsu.edu/~kbach/ambguity.html, 2010-09-26 Watters, P.A, Patel, M; Semantic processing preformance of Internet machine translation system, p.153 Watters, p.155 8 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 Fi gure 1: Explains the principles of Direct MT. One of the oldest and most widely used Direct MT's is SYSTRAN, a translation software that has been active for over 40 years and is integrated in several online MT17, for example Yahoo's “Babelfish”18. To make a simple example over how it works the following sentence has been translated using a Direct MT: English[The dog is in his house] → Swedish[Hunden är i hans hus] Those of us who speak Swedish can see that even though the words are translated correctly the translation of “his → hans” is inappropriate. The pronoun should be bent into “sitt” instead of “hans”. 3.2 Interlingua system This is actually a sub version of Direct MT rather than a separate MT version in itself. Even though the basics in how the translation works is fundamentally the same, it is still unique in its way to reach a translation. This form of Direct MT simply converts the words into a universal language that is created for the MT simply to translate it once more to the language we originally were interested in. Sometimes this universal language is Esperanto or English. This has a lot of benefits concerning the desire to translate a text into many different languages. Since we no longer need a two-way dictionary an Interlingua MT can easily add new languages to be translated without 17 SYSTRAN, http://www.systran.co.uk/systran, 2010-10-03 18 Babelfish, http://babelfish.yahoo.com/, 2010-10-03 9 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 much effort. However its time efficiency is lower than that of an ordinary Direct MT.19 Compare figure 2 that shows how an interlingua system works and that of an ordinary Direct MT in figure 1. Figure 2: picture showing how different languages are translated through a Interlingua language. An example made with English as an interlingual language could look like this: Spanish[El perro está in su casa] → Interlingua[The dog is in his house] → Swedish[Hunden är i hans hus] Important to notice in interlingua translations is that since there are two translations, one from the source language and one from the interlingua language, a lot more information can be lost or interpreted wrongly on the way to the target language. But even though it may lose information and misinterpret it is a lot more practical when several languages are to be interpreted since it only needs to translate it once from the source language. 19 Watters, p.153 10 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 Figure 3: explains how a language A can be translated to several languages using an interlangua. 3.3 Transfer system MT Transfer system MT is slightly more clever than the direct translation MT because it is based on a database of translation rules rather than a dictionary. Whenever a sentence matches one of the rules, or examples, it is translated directly using a dictionary such as in direct translation. It goes from the source language to a morphological and syntactic analysis to produce a sort of interlingua on the base forms of the source language, from this it translates it to the base forms of the target language and from there a better translation is made to create the final step in the translation.20 See figure 4. Figure 4: describing a transfer system. This is a MT that can occur at lexical, syntactic or semantic lever, whereas the direct MT only works on a lexical level. This of course makes the MT superior in the quality of the translation but there are still a lot of problems if you for example want to translate ”White House” from English to Spanish it would follow rules such as this: 20 Sánchez-Martínez,F, Ney, N; Using Alignment Templates to Infer Shallow-Transfer Machine Translation Rules, Berlin 2006, p.755 11 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 English [adjective noun] → Spanish[noun adjective] English[White House] → Spanish[Casa Blanca] However correct this literal translation may be, it is perhaps not necessarily what we meant. When we want to translate the government building of the United States we end up with a classical movie. 4. Modern Machine Translation 4.1 Rule based MT This has its origin in Transfer system MT where it in likeness uses a database of rules, usually based on morphological and bilingual dictionaries, to translate a text.21 One problem is that since it never stores new information it is incapable of handling new texts.22 One of the more common rule-based MTs is the Knowledge based MT(KBMT). It works from the point of view that you need to understand the text to be able to translate it. To return to the White House-example from before, it would be easy to put it into context and from knowledge based MT know that White House is not to be translated because it is a name. Even though KBMT have been found to be quite accurate in its translations it needs a lot of computer memory. This has made it rather difficult to apply to larger texts such a book or even a newspaper.23 4.2 Corpus based MT Corpus based MT gets all information from large corpuses were grammatical rules have been cast aside to use pairs of the different languages to translate words and sentences. Example based MT is one of the bigger versions of a Corpus based MT. It uses Translation templates which is a bilingual pair of sentences or phrases where words are coupled and replaced by variables .24 The goal is to have a large, and good, enough corpus to be able to directly translate word after word in a sentence based on the translation templates. The idea is 21 Sánchez-Martínez, p.759 22 Carl, M, Pease, C, Iomdin, L.L, Streiter, O; Towards a Dynamic Linkage of Example-based and Rule-based Machine Translation, Machine Translation (2000) 15:223-257, p.225 23 Knight, K, Luk, S.K; Building a Large-Scale Knowledge Base for Machine Translation, 1994 https://www.aaai.org/Papers/AAAI/1994/AAAI94-118.pdf, 2010-09-24 24 Sanchez-Martínez, p.757 12 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 that since these pairs always are correct, if we find one of the pairs in the source text, we don't need grammar rules to know the corresponding phrase or word. Figure 5: Describes the idea behind Translation templates of a phrase in two different languages. However, most Example based MT's usually uses some grammar to get the corpus to a minimum, this however almost turns the Example based MT into a Transfer system MT, with the important exception that in Example based MT the rules are very specific, almost one for every case, whereas in Transfer systems MT the rules are more general. 4.3 Statistic based MT Statistic based MT(SMT) is, together with Hybrid MT, one of the most frequently used MT's today. It is a Corpus based MT that usually translate using the Expectation Maximization Algorithm or Baye's rule, two similar ways to determine the probability of a translation being accurate.25 The general idea is that the translation will be from the most likely translated word. To know which word is most common statistical data that is gathered from several bilingual corpuses, today the Internet is the biggest source. In Statistic based MT we are interested in the probability that “White House” is in fact to be interpreted as “Casa Blanca” rather than the actual “White House” and based on the probability the MT chose the translation. A SMT build on probability rules such as Baye's rule works like this: 25 Sofianopoulos, S, Tambouratzis, G; Multi-objective optimisation of real valued parameters of a hybrid MT system using Genetic Algorithms, Pattern Recognition Letters 31 (2010), p.1672 13 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 Baye's rule says that: P(S | E) = P(E | S)*P(S) / P(E) This however equals the following equation because the P(E) is constant. P(S | E) = P(E | S)*P(S) Where: P(E | S) : is how probable that the English sentence is a translation to the Swedish one. P(S) : is the probability that it is this Swedish sentence P(E) : is the probability that it is this English sentence This means that the probability that the English sentence is a correct translation of the Swedish sentence is equal to the probability that the Swedish sentence is a correct translation of the English sentence multiplied with the probability that the Swedish sentence is correct.26 Expectation Maximization(EM) Algorithm is another mathematical approach to determinate the likelihood, rather than the probability, of missed data or in this case, missed translations, and from this the likelihood of the translation to be the best. Then it repeats this steps until only one translation is left. Hopefully it is the most accurate one.27 4.4. Hybrid MT There are a number of different types of Hybrid MT(HMT):s since the essence of this version of Machine Translation focus upon a mix of the other ones. There is still a lot of research being done within this type of MT and due to this it is difficult to give a complete explanation of HMT. But to help you get the general idea of the concept an attempt to explain how one of the more common hybrid MT's works. Namely the Generation-Heavy MT(GHMT). GHMT comes from a history of primarily transfer and interlingua MT but is in itself neither. Due to the fact that it has more knowledge of words and lexical translations it is more of a knowledge based MT than a transfer system. However since this lexical knowledge is represented in a transfer-like interlingua tree it is assumable to categorize GHMT as a partly transfer system.28 26 Russel, S, Norvig, P; Artificial Intelligence: A Modern Approach, New Jersey 2003, p.853 27 Borman, S; The expectation Maximization Algorithm: A short tutorial, 2004 28 Habash, N, Dorr, B, Monz, C; Symbolic to statistical hybridization: Extending generation heavy machine 14 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 The data result from the lexical tree is then analysed and ranked from a statistical point of view to get the most likely translation in the top of the tree. The next step is a rather complicated one based of several step in which I will not venture in further in the believe that that will most likely complicate matters more than it will be beneficial to the understanding of Hybrid MT. What happens in this next step is, very much simplified, the final translation from the interlingua to the desired language including organizing the words after appropriate grammar and sentence structure as done by a traditional Knowledge based MT.29 See figure 6 for an overview of how GHMT works. Figure 6: Simplified explanation of how the GHMT works. 5. Discussion That Machine Translation is something important has already been established since it is useful in more areas than one and it is likely that the MT systems of today will get more powerful and be able to produce even better translations than today. The different forms of MT all have benefits and disadvantages and it is likely that the best MT that ever will be created is a Hybrid MT since it can take the best aspects of all the other MTs. However I do believe that we will never be able to create a MT that is able to translate every possible text, from all different languages perfectly. Even if we were able to create a MT that could handle the grammar correctly, with all its exceptions, metaphors, ambiguity and so on, I simply believe that a language is more than linguistics and words. A language is a culture and is therefore difficult enough for a human to translate, since a culture is something that has to be experienced. translation, 2009, p.28 29 Habash, p.32 15 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 This said, I do enjoy the world of MT very much and use it very frequently, if not daily, and perhaps it is not so important to take the cultural part of MT so seriously. As long as we are aware of the limitations of the technology, and realise that MT might misinterpret or interpret culture that we as humans of another culture cannot understand, we still can benefit from the possibilities that a language is available to us even if it is trough a veil. Even though I have disbeliefs of whether we ever will create a perfect MT I do believe that MT will be evolved into not only written language but also spoken language. This is quite a fresh research area that is still under construction but I see it as a possible future that one day we won't need an interpreter to be able to directly communicate with someone in another language. Perhaps all we will need is a small gadget that translates the other persons speech directly into our ears. In this way, to experience the language in a more appropriate and direct association to it is culture it is likely that the mistakes that the translation gadget makes won't be of the same magnitude since a human is there to correct it. This is a wonderful thought, to be able to directly talk to people of other languages but I believe that this will be found to be even more difficult than the MTs for written language. When we speak we have a lot of different ways of speaking; slang, dialects, different ways of word order, pauses. If you believe that you speak as they do in the movies, you either only watches dogmatic documentaries or you're unaware of how you speak. However I do believe that this will be a possibility and not only something you'd see on Star Trek. But perhaps it is likelier that the world has united under a universal language before this is perfected, maybe under a language such as Mandarin or perhaps English. 16 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Maria Hedblom, marhe503 Kognitionsvetenskap 2 6. References Articles: Sean Borman, The expectation Maximization Algorithm: A short tutorial, 2004 http://www.seanborman.com/publications/EM_algorithm.pdf, 2010-10-02 Peter F. Brown, John Cocke, A. Della Pietra, Vincent J. Della Pietra, Fredrick Jelinek, John D. Lafferty, Robert L.Mercer & Paul S. Roossin; A Statistical Approach to Machine Translation, Computational Lingustics Volume 16 (1990) 2:79-85 http://acl.ldc.upenn.edu/J/J90/J90-2002.pdf, 2010-09-24 Michael Carl, Catherine Pease, Leonid L. Iomdin & Oliver Streiter; Towards a Dynamic Linkage of Example-based and Rule-based Machine Translation, Machine Translation (2000) Nr 15:223-257. Nizar Habash, Bonnie Dorr & Christof Monz; Symbolic-to-statistical hybridization: Extending generation-heavy machine translation, Mach Translat (2009) Nr 23:23–63 W. John Hutchins; Machine Translation: A Brief History, Oxford 1995 http://aymara.org/biblio/mtranslation.pdf, 2010-10-02 Kevin Knight & Steve K. Luk; Building a Large-Scale Knowledge Base for Machine Translation, 1994, https://www.aaai.org/Papers/AAAI/1994/AAAI94-118.pdf, 2010-09-24 Felipe Sanchez-Martínez & Hermann Ney; Using Alignment Templates to Infer ShallowTransfer Machine Translation Rules, Berlin 2006 http://www.springerlink.com.lt.ltag.bibl.liu.se/content/6044810011064370/, 2010-10-03 Sokratis Sofianopoulos & George Tambouratzis; Multi-objective optimisation of real valued parameters of a hybrid MT system using Genetic Algorithms, Pattern Recognition Letters (2010) Nr 31:1672–1682 Paul A. Watters & Malti Patel; Semantic processing preformance of Internet machine translation system, Internet Research: Electronic Networkning Applications and Policy Volume 9(1999) Nr 2:153-160 Books: Sergei Nierenburg, Harold L. Somers & Yorick A. Wilks: Reading in Machine Translation, 2003 http://cognet.mit.edu.lt.ltag.bibl.liu.se/library/books/mitpress/0262140748/cache/I.pdf Suart Russel & Peter Norvig; Artificial Intelligence: A Modern Approach, New Jersey 2003 17 Machine Translation - A Rosetta stone for the 21th century? 2010-10-03 Web pages: Kent Bach; Ambiguity, http://userwww.sfsu.edu/~kbach/ambguity.html, 2010-09-26 Chandious, http://www.chandioux.com/profile.asp, 2010-09-25 Cultivate Interactive, Marieke Napier, 2002 http://www.cultivate-int.org/issue2/mt/ Nationalencyklopedin, http://www.ne.se/rosettestenen. 2010-09-24 SYSTRAN, http://www.systran.co.uk/systran, 2010-10-03 18 Maria Hedblom, marhe503 Kognitionsvetenskap 2