Topic models for dynamic translation

合集下载

Geometric Modeling

Geometric ModelingGeometric modeling is a fundamental concept in the field of computer graphics and design. It involves the creation and manipulation of geometric shapes and structures in a virtual environment. This process is essential for creating realistic and visually appealing 3D models that are used in various industries such as animation, architecture, and engineering. One of the key aspects of geometric modeling is the representation of objects in a digital space. This involves defining the shape, size, and position of each element in the model. There are various techniques used for this purpose, including parametric modeling, which allows for the creation of complex shapes by defining mathematical parameters, and polygonal modeling, which uses a mesh of polygons to represent the surface of an object. Another important aspect of geometric modeling is the manipulation of objects within the digital environment. This includes operations such as translation, rotation, and scaling, which allow designers to modify the position and size of objects in a 3D space. These operations are essential for creating dynamic and interactive models that can be used for a variety of applications. In addition to creating and manipulating objects, geometric modeling also involves the representation of materials and textures. This includes defining the surface properties of objects, such as color, reflectivity, and transparency, as well as applying textures to simulate real-world materials like wood, metal, or fabric. These techniques are crucial for creating realistic and visually appealing models that can be used in virtual simulations and animations. One of the challenges of geometric modeling is ensuring accuracy and precision in the representation of objects. This requires a deep understanding of mathematical concepts such as geometry, trigonometry, and calculus, as well as proficiency in software tools such as CAD (Computer-Aided Design) and 3D modeling software. Designers must also have a keen eye for detail and a strong sense of spatial awareness to create models that are both aesthetically pleasing and functionally accurate. Overall, geometric modeling is a complex and multifaceted process that plays a crucial role in the field of computer graphics and design. It requires a combination of technical skills, creative thinking, and attention to detail to create realistic and visually appealing 3D models that can be used in a variety ofindustries. By mastering the principles of geometric modeling, designers can bring their ideas to life in a virtual space and create stunning visual experiences for audiences around the world.。

面向汉维机器翻译的双语关联度优化模型

收稿日期：２０１８０８２２；修回日期：２０１８１０２４基金项目：国家自然科学基金资助项目（Ｕ１７０３１３３）；中科院西部之光项目（２０１７ＸＢＱＮＸＺＡ００５）；中国科学院青年创新促进会的资助项目（２０１７４７２）；新疆维吾尔自治区重大科技专项项目（２０１６Ａ０３００７３）；新疆维吾尔自治区高层次人才引进工程项目（Ｙ８３９０３１２０１）作者简介：潘一荣（１９９２），女，天津人，博士研究生，主要研究方向为自然语言处理、机器翻译；李晓（１９５７），男（通信作者），博导，研究员，硕士，主要研究方向为多语种信息处理、信息系统研究与开发（ｘｉａｏｌｉ＠ｍｓ．ｘｊｂ．ａｃ．ｃｎ）；杨雅婷（１９８５），女，副研究员，硕导，博士，主要研究方向为多语种信息处理技术；董瑞（１９８５），男，助理研究员，博士研究生，主要研究方向为多语种信息处理．面向汉维机器翻译的双语关联度优化模型潘一荣１，２，３，李　晓１，３，杨雅婷１，３，董　瑞１，３（１．中国科学院新疆理化技术研究所，乌鲁木齐８３００１１；２．中国科学院大学，北京１０００４９；３．新疆民族语音语言信息处理实验室，乌鲁木齐８３００１１）摘　要：针对汉语—维吾尔语的统计机器翻译系统中存在的语义无关性问题，提出基于神经网络机器翻译方法的双语关联度优化模型。

该模型利用注意力机制捕获词对齐信息，引入双语短语间的语义相关性和内部词汇匹配度，预测双语短语的生成概率并将其作为双语关联度，以优化统计翻译模型中的短语翻译得分。

在第十一届全国机器翻译研讨会（ＣＷＭＴ２０１５）汉维公开机器翻译数据集上的实验结果表明，与基线系统相比，在使用较小规模的训练数据和词汇表的条件下，所提方法可以同时有效地提高短语级别和句子级别的机器翻译任务性能，分别获得最高２．４９和０．５９的ＢＬＥＵ值提升。

关键词：维吾尔语；神经网络机器翻译；注意力机制；词对齐；生成概率中图分类号：ＴＰ３９１文献标志码：Ａ文章编号：１００１３６９５（２０２０）０３０１９０７２６０５ｄｏｉ：１０．１９７３４／ｊ．ｉｓｓｎ．１００１３６９５．２０１８．０８．０６２５ＢｉｌｉｎｇｕａｌｒｅｌａｔｅｄｎｅｓｓｏｐｔｉｍｉｚａｔｉｏｎｍｏｄｅｌｆｏｒＣｈｉｎｅｓｅＵｙｇｈｕｒｍａｃｈｉｎｅｔｒａｎｓｌａｔｉｏｎＰａｎＹｉｒｏｎｇ１，２，３，ＬｉＸｉａｏ１，３，ＹａｎｇＹａｔｉｎｇ１，３，ＤｏｎｇＲｕｉ１，３（１．ＸｉｎｊｉａｎｇＴｅｃｈｎｉｃａｌＩｎｓｔｉｔｕｔｅｏｆＰｈｙｓｉｃｓ＆Ｃｈｅｍｉｓｔｒｙ，ＣｈｉｎｅｓｅＡｃａｄｅｍｙｏｆＳｃｉｅｎｃｅｓ，Ｕｒｕｍｑｉ８３００１１，Ｃｈｉｎａ；２．ＵｎｉｖｅｒｓｉｔｙｏｆＣｈｉｎｅｓｅＡｃａｄｅｍｙｏｆＳｃｉｅｎｃｅｓ，Ｂｅｉｊｉｎｇ１０００４９，Ｃｈｉｎａ；３．ＸｉｎｊｉａｎｇＬａｂｏｒａｔｏｒｙｏｆＭｉｎｏｒｉｔｙＳｐｅｅｃｈ＆ＬａｎｇｕａｇｅＩｎｆｏｒｍａｔｉｏｎＰｒｏｃｅｓｓｉｎｇ，Ｕｒｕｍｑｉ８３００１１，Ｃｈｉｎａ）Ａｂｓｔｒａｃｔ：ＦｏｃｕｓｅｄｏｎｔｈｅｉｓｓｕｅｏｆｓｅｍａｎｔｉｃｉｎｄｅｐｅｎｄｅｎｃｅｉｎＣｈｉｎｅｓｅＵｙｇｈｕｒｓｔａｔｉｓｔｉｃａｌｍａｃｈｉｎｅｔｒａｎｓｌａｔｉｏｎｓｙｓｔｅｍ，ｔｈｉｓｐａｐｅｒｐｒｏｐｏｓｅｄａｂｉｌｉｎｇｕａｌｒｅｌａｔｅｄｎｅｓｓｏｐｔｉｍｉｚａｔｉｏｎｍｏｄｅｌｂａｓｅｄｏｎｎｅｕｒａｌｍａｃｈｉｎｅｔｒａｎｓｌａｔｉｏｎｍｅｔｈｏｄ．Ｔｈｅｍｏｄｅｌｕｔｉｌｉｚｅｄｔｈｅａｔｔｅｎｔｉｏｎｍｅｃｈａｎｉｓｍｔｏｃａｐｔｕｒｅｗｏｒｄａｌｉｇｎｍｅｎｔｉｎｆｏｒｍａｔｉｏｎａｓｗｅｌｌａｓｉｎｔｒｏｄｕｃｅｄｂｉｌｉｎｇｕａｌｐｈｒａｓｅｓｅｍａｎｔｉｃｒｅｌｅｖａｎｃｅａｎｄｉｎｎｅｒｗｏｒｄｃｏｒｒｅｌａｔｉｏｎｔｏｐｒｅｄｉｃｔｔｈｅｃｏｎｄｉｔｉｏｎａｌｐｒｏｂａｂｉｌｉｔｙｏｆｂｉｌｉｎｇｕａｌｐｈｒａｓｅｐａｉｒ．Ａｎｄｉｔｔｏｏｋｔｈｅｐｒｏｂａｂｉｌｉｔｙａｓｂｉｌｉｎｇｕａｌｒｅｌａｔｅｄｎｅｓｓｔｏｏｐｔｉｍｉｚｅｔｈｅｐｈｒａｓｅｔｒａｎｓｌａｔｉｏｎｓｃｏｒｅｓｉｎｓｔａｔｉｓｔｉｃａｌｔｒａｎｓｌａｔｉｏｎｍｏｄｅｌ．Ｅｘｐｅｒｉｍｅｎｔａｌｒｅｓｕｌｔｓｏｎｔｈｅ１１ｔｈＣｈｉｎａＷｏｒｋｓｈｏｐｏｎＭａｃｈｉｎｅＴｒａｎｓｌａｔｉｏｎ（ＣＷＭＴ２０１５）ＣｈｉｎｅｓｅＵｙｇｈｕｒｐｕｂｌｉｃｍａｃｈｉｎｅｔｒａｎｓｌａｔｉｏｎｄａｔａｓｅｔｓｓｈｏｗｔｈａｔｔｈｅｐｒｏｐｏｓｅｄａｐｐｒｏａｃｈｃａｎａｃｈｉｅｖｅｏｂｖｉｏｕｓｉｍｐｒｏｖｅｍｅｎｔｓｂｏｔｈｉｎｔｈｅｐｈｒａｓｅｌｅｖｅｌａｎｄｔｈｅｓｅｎｔｅｎｃｅｌｅｖｅｌｍａｃｈｉｎｅｔｒａｎｓｌａｔｉｏｎｔａｓｋｓ，ｗｈｉｃｈｏｕｔｐｅｒｆｏｒｍｓｔｈｅｂａｓｅｌｉｎｅｓｙｓｔｅｍｗｉｔｈａｒｅｌａｔｉｖｅｓｍａｌｌｓｃａｌｅｔｒａｉｎｉｎｇｄａｔａａｎｄｖｏｃａｂｕｌａｒｙ．ＴｈｅｈｉｇｈｅｓｔＢＬＥＵｐｏｉｎｔｏｆｔｈｅｐｒｏｐｏｓｅｄａｌｇｏｒｉｔｈｍｇａｉｎｓ２．４９ａｎｄ０．５９ｒｅｓｐｅｃｔｉｖｅｌｙ．Ｋｅｙｗｏｒｄｓ：Ｕｙｇｈｕｒ；ｎｅｕｒａｌｎｅｔｗｏｒｋｍａｃｈｉｎｅｔｒａｎｓｌａｔｉｏｎ；ａｔｔｅｎｔｉｏｎｍｅｃｈａｎｉｓｍ；ｗｏｒｄａｌｉｇｎｍｅｎｔ；ｃｏｎｄｉｔｉｏｎａｌｐｒｏｂａｂｉｌｉｔｙ０　引言在基于短语的统计机器翻译（ｓｔａｔｉｓｔｉｃａｌｍａｃｈｉｎｅｔｒａｎｓｌａｔｉｏｎ，ＳＭＴ）［１］系统中，翻译模型对从平行语料库中抽取的双语短语进行建模，主要包括短语翻译概率、词汇化权重等参数，这些参数作为特征函数并结合对数线性方法，以此训练机器翻译系统，从而获取最优权重分布，在解码时以搜索最有可能的翻译选项，实现双语转换过程。

language and cognition

distinguishing features for each one)
everyday reference; be basic in 3 respects– perception, communication, knowledge organization eg. chair) higher level or more general level than basic one
Language comprehension:
Word recognition Comprehension of sentences Comprehension of text
♥ Word recognition:
Factors the affect the word recognition : frequency & context
[in a word, the successful comprehension of speech sounds is a combination of the innate ability of humans to distinguish minute differences between speech sounds, and the ability to adjust to the acoustic categories of the language they are exposed to.]

nine types schema:
1. Access to words (conceptualization, select a word that corresponds to the chosen concept, morpho-phonological encoding);

TAUS：选择翻译的未来-英文原文

CHOOSE YOUR OWN TRANSLATION FUTURE Created on 14 January 2013Written by Jaap van der Meer中文 | 日本Technology arrived late in the translation services sector. Now it has arrived, it is bound to change everything. In the not‐too‐distant future everyone in the world will be able to speak his or her own language and everyone else will understand. We are entering the Convergence era: translation will be a utility embedded in every app, device, sign board and screen.Businesses will prosper by finding new customers in new markets. Governments and citizens will connect and communicate easily. Consumers will become world‐wise, talking to everyone everywhere as if language barriers never existed.Don’t get me wrong. It will not be perfect, but it will open doors and break down barriers. And it will give a boost to the translation industry, which will be chartered to constantly improve the technology and fill the gaps in global communications.Is this picture too rosy? Not if you believe in the power of translation data, like we do. Translation data is the fuel of machine translation technology. Data powers the engines. The engines may never emulate the human language competence, but they will be good enough to help us converse as we see fit inlanguages we never spoke before or will ever speak. Machine translation, according to Nicholas Ostler, will be the new lingua franca.This is a vision that frightens many insiders in the translation industry. Machine translation was experimented, tried and tested for a long time, but it never passed the test of usefulness. Automation of translation was believed to be a utopia, at least until the vox populi revolution spoke and millions of people started clicking on the automatic translate button in their search pages. But no matter whether the quality is often bad and laughable, people simply like the fact that it is under their control and in real‐time. It’s a sign of the times. Users take charge and drive change.Entering the Convergence eraAs the evolution diagram below shows, the translation industry has undergone a paradigm shift every decade since 1980, but none was as big as the one we are facing now: the Convergence.Source: TAUSThe volume of content is exploding to zettabytes (trillions of gigabytes) of information that can be relevant to billions of new users who click to translate as much as they like. While we make this journey from the 20th century export mentality to the 21st century’s open global society, the mix of language pairs will be shifting from today’s 7 source and 60 target languages to 200 source and 200 target languages in the next ten years. It is utterly clear that a human‐driven translation process alone will not suffice in this new era.In the current phase—the Integration era—enterprises and institutions are busy releasing the translation function from its isolated position. The focus is on integrating translation in enterprise applications such as content management systems. This will help organizations to scale up and translate a lot more than just the usual documents, instructions, brochures and software.But the pressure will keep building to translate more and more content faster or even translate it in real‐time. This opens up tremendous opportunities forinnovators to seize the convergence instrument and offer solutions that did not exist before. (See the Agents of Change: Insiders and Invaders videos.)Two types of convergenceWe highlight two interconnected forms of convergence: pure technology convergence and functional convergence. Technology convergence means combining two or more technologies to create a new compelling product or service offering. Functional convergence means combining functions to create a new solution.The best example of technology and functional convergence in our daily life is the mobile phone. It has now become a camera, a PDA, a navigation tool, and so much more with thousands of new apps being developed to turn this simple handheld device (called a ‘handy’ in German language) into a life‐saving and indispensable kind of extension to our body.In the physical world, the emergence of supermarkets was a form of convergence. The combined offering of coffee and music by Starbucks is also a good example of convergence. In the digital world functional convergence often has a give‐and‐take dimension: the user becomes part of the supply chain. Examples of this are restaurant review websites where users are requested to give ratings and share reviews of the restaurants. The service is free. The owner of the site makes money through advertisements.More innovative examples of functional convergence are location‐based apps (another form of localization). The user—often without knowing it—transmits his or her exact location and receives perfectly matching offers from a shop or restaurant in the neighborhood or an invitation to meet a friend who happens to be walking on the same street.Convergence in the translation industryWe can start imagining what convergence can mean to the translation industry. In fact convergence has already started across technologies and across functions. We have seen the first demonstrations of the integration of speech and machinetranslation technology. Imagine what happens if the technologists get this to work really well. Using tiny little keys on your mobile will no longer be necessary: speech input in one language, and speech output in another language.Of course, the best example of functional convergence in the translation industry is the combination of automatic translation with search. This innovation fired off the vox populi revolution I mentioned above. Millions of end‐users started clicking every day to use real‐time translations. They don’t pay unless you want to call their viewing of advertisements on search pages a form of payment. The owners of search engines decided to extend the service to professional translators.The business model convergence went a step further: for sharing the translation data (translation memories) the industry professionals received customized (improved quality) machine translations. Another recent example of functional convergence in the translation sector is Duolingo: an online gamified language training site. It’s free, but users are helping to translate sentences according to their skill level. This way they return a service while at the same time feeding translation data which helps to improve the platform.In the next ten years we will see numerous new examples of converging functions and technologies. Sometimes this convergence will address just one language pair, domain or market niche. Sometimes it will be applicable on a much wider scale. Together, this convergence is changing the translation industry completely. Translation will quickly become a utility embedded in everything we do. It will be as ubiquitous as electricity and the Internet.More and more, it will be considered a basic necessity for human kind. Language communities not connected yet through this translation utility will make a special effort to become part of it by aggregating the required translation data and sharing it. This is what we call the Viral Effect. It causes the acceleration in the spread of language pairs and domains and the continuous performance improvement of the translation utility.Crowd, Cloud and Big DataOther trends that play along in the Convergence era are Crowd, Cloud and Big Data. The Crowd is part and parcel of functional convergence. Duolingo needs hundreds of thousands of users to make the platform really work well, because it is the voting on the best translations that will improve the overall performance of the system. The Cloud is the natural infrastructure environment to connect with the Crowd and to reach the required scalability and efficiency.Many innovative translation solutions will be characterized as SaaS (Software‐as‐a‐Service), DaaS (Data‐as‐a‐Service), IaaS (Infrastructure‐as‐a‐Service) and PaaS (Platform‐as‐a‐Service), all variances of Cloud‐based solutions. But behind the Crowd and the Cloud is the secret power of Big Data—the biggest trend of all. When IBM’s Watson beat the best Jeopardy players in 2011, it was a milestone event in natural language processing. It proves that the computer can decipher ambiguity, understand jokes and metaphors, as long as it is fed enough data.The importance of Big Data for the translation industry should not be underestimated. Big Data will push the performance of automated translation forward. Big Data will address challenges in many different areas of natural language processing, including machine translation. The computer will be able to run automatic semantic clustering and genre identification processes, meaning that the computer will recognize the industry domain (for instance: medical and radiology) and the type of content (for instance: instruction text or patent application).This is vital for the continuous improvement and customization of machine translation technology. Big Data technologies become crucial since the modern machine translation systems involve more and more parallel data and it is reaching the limit when it is not feasible to process large amounts of data with traditional database management techniques. The computer will also be able to do terminology mining much better if it gets more data.It will identify synonyms, related terms, neologisms, jargon and automatically generate syntactic classification using parallel processing tools. Plain statistical translation models evolve into hybrid models with hierarchical (syntax‐ or alignment‐based) trees allowing the machine translation engines to dolong‐range reordering, creating more fluent and correct translations especially for more distant language pairs.Translation support matching the new content mixIn the Convergence era, the mix of content to be translated is shifting further away from documents and software releases to bits and pieces of text, voice and video published on multiple screens. The end‐user, citizen or patient will be in control—even more than today—and they will drive a continuous stream of translation of official (corporate, public, legislation), social, shared, earned and also private information.Translation memory software fits very well with updates of static documentation pushed by publishers but it will not be very helpful when translating dynamic content pulled by users. Machine translation technology will mature quickly and take over as the primary choice of tools to be used by the translation service sector. New features will be added to MT platforms allowing professional users to add data (customer‐specific or product‐specific translation memories, glossaries and target language texts) that will train and customize the engine almost in real‐time.This self‐service real‐time training of MT engines may be applied to every single job. Personalization of MT is a far jump from the costly and lengthy process of MT development for a generic language pair that we were used to. It will drive the need for translation memory data to be bigger and bigger. For every new job translators will be looking for matching data to fine‐tune the engine. The need for data will be insatiable.So, where does that leave the entrepreneurs in the translation industry—the buyers and providers of translation?Planning for an uncertain futureIn 2010 TAUS organized a series of brainstorming sessions following the scenario‐based planning methodology with translation buyer and provider executives in Copenhagen and Portland (OR) with the aim of planning for anuncertain future to minimize crisis‐driven change and instead pursue opportunity‐driven change. The participants agreed that certain drivers were indisputable (content explosion, the shift to multimedia and mobile media, and the trend to real‐time delivery), but they were uncertain about the answers to three questions:1. Will machine translation take a big role in the translation industry or not?2. Do we have to fear that translation will become a free‐for‐all service?3. Will the closed (competitive) or the open (collaborative) business models prevail?Two of the three questions have been answered in the last couple of years. Yes, MT will play a major role in the translation industry. No, translation will not be free. There is a lot of elasticity in translation pricing, but somehow users always pay for translation. However, the third question is still haunting us. We have not seen clear indicators yet whether closed models or open models will prevail. Both seem to function very well.Open or closed translation futuresThe future of the translation industry could be closed (as it is more or less today) or it could be open and collaborative. In the closed translation future scenario, a few companies will have aggregated all of the world’s translation data that facilitate and support fast and efficient translation of the world’s information in 40,000 or more language pairs. Large and small translation operators—including corporate buyers, governments and institutions—will be dependent on the few data owners to keep their translation engines tuned for every job.In today’s translation world, the translation memories they own or manage for their customers may be sufficient to keep their translation operations running efficiently. But in the Convergence era, it will be harder to predict which content and in what domain or language pair needs translation. New data will always be needed to make new translations possible.In the open translation future scenario, data is shared in collaborative platforms. All translation operators have access to the data on an equal basis and may usethe data to leverage and to develop derivative work, i.e., new machine translation engines. In the open translation future scenario, industry stakeholders agree on common interfaces to connect content, technologies and platforms to ensure a frictionless exchange of translation jobs and data. In the open translation future scenario, industry stakeholders agree on common metrics and benchmarking to measure and compare the performance of automatic translation engines and to track progress.Both scenarios could be true. It is hard to tell today which one has a better chance to win out over the other. In both scenarios we see opportunities for growth. But unless you have a fairly good chance to own all the data that you possibly need in your translation future, your growth opportunities will be much greater in the open translation future scenario.Fork in the roadIn the coming two years—more than ever before—translation buyers and providers will have a decision to make to open or not to open; to collaborate and share or not to collaborate and share. Rather than being taken by surprise, it will be wise to take a conscious decision about your own translation future. We are at a fork in the road. Going one way or the other can make a big difference for the success and growth of your business.Choosing the open translation future scenario means openly sharing your translation memories and convincing your customers and collaborators to do the same. Translation data—other than translation memories—cannot be easily used to reconstruct the original individual source or target language documents. We should look at translation data as data in the same way as the medical industry treats human genome data.Every life sciences company, every university—in fact, everyone in the world—has access to the descriptions of the 1.3 billion chemical base pairs that constitute human DNA. Every company can use human DNA data to develop new medicines and new technologies. This is what stimulates innovation, growth and helps human civilization. Of course, if you choose to share your translation data, you are free not to share confidential data or non‐released product information.Choosing the open translation future scenario means collaborating in translation quality benchmarking and industry metrics. In today’s translation world, every operator has its own way of evaluating translation quality. We have no way to compare and benchmark quality with peers in the industry. To scale up and prepare for growth in the Convergence era, we need to be able to measure the performance of MT engines, as well as track and compare their progress across domains, language pairs and content types.We need to be able to establish best practices—on an industry‐wide scale—when and when not to use MT technology. We need to have industry agreement on acceptable scores, ratings and evaluation techniques. If we don’t have this, it will be harder to meet market expectations and scale up.Force and counterforceFinally, if the translation future still looks frightening to you, relax, because every force has a counterforce. The ubiquitous availability of non‐perfect automated translation will also lead to growth in the need for high quality (non‐automated) translation, transcreation and personalization where old‐fashioned human language skills are unbeatable.The future of translation looks good. It is your choice where you want to be.。

论文写作2(Academic Paper Writing)

relation of equivalence.
ST≈TT or
TT≈ST
This is an approach that has close links with
contrastive linguistics and puts language system
rather than texts on either side of the relation:
We define research broadly as a systematic investigation towards increasing the sum of knowledge. Innovation is vital if a discipline is to grow and prosper.
These models are attempts to construct images of the object of study, images that hopefully make it easier to visualize, understand and analyse. In your own research project, you might take a ready model and simply use its framework and concepts unchanged or you might adapt a given model to your own purposes. Translation Studies has traditionally used three basic types of models.
• Comparative Models
The earliest theoretical model of translation was

Proposal for a Hieronymic Oath

The Translator . V olume 7, Number 2 (2001), 139-154 ISBN 1-900650-50-9Proposal for a Hieronymic OathANDREW CHESTERMAN University of Helsinki, FinlandAbstract. Four current models of translation ethics are described,based on the ideas of representation, service, communication and norms. There are problems with all these models: they are in severalrespects incompatible, and have different ranges of application. An alternative approach is therefore offered based on Alasdair MacIntyre’s ideas about virt ues and the deontic force of excellencein a social practice. This leads to a fifth possible model, an ethicsof professional commitment, comparable with Maria Tymoczko’ssuggestion that translation is a commissive act. At the centre of such a model there might be an official oath, comparable to the Hyppocratic Oath for the medical profession. I end with a pro- posal for a Hieronymic Oath for translators.In the Hopi Indian culture, the Spider Grandmother is the mythical figurewho oversees the spiritual development of humankind. She gave two basicethical rules, not just to Hopis but to all humanity. “She said, ‘Don’t go aroundhurting each other,’ and she said, ‘Try to understand things’” (Heat-Moon1984:187).These two ethical principles – one negative and one positive – may takeus a long way toward an ethics of translation. This paper first offers an analy-sis of the current state of affairs, and then offers an alternative proposal.1. Four current models of translation ethicsI will start by crystallizing the multiplicity of ideas about translation ethicsinto four basic models. There are overlaps between some aspects of thesemodels, but I will mostly disregard them here.Ethics of representation. This model of translation ethics goes way back tothe ideal of the faithful interpreter, and to the translation of sacred texts. Theethical imperative is to represent the source text, or the source author’sintention, accurately, without adding, omitting or changing anything. Acontemporary manifestation of this ethic is to be found in the EU translationservices: political reasons dictate that EU documents in whatever official EU language are legally equivalent to any other language version, perfect repre-sentatives of each other, and that no single version is privileged as a sourcetext, in theory (for a detailed discussion of the EU translation situation, seeKoskinen 2000a).Another line of inheritance of this ethic has to do with the long traditionof representing the Other, the relation with alterity. It comes to the fore par-ticularly during the German Romantic movement and in subsequent theoreticalcontributions stressing the value of allowing the Other to appear in its ownlight, without being domesticated: cf. the general arguments of Schleier-macher, Berman (e.g. 1984) and Venuti (e.g. 1995). These theoretical positionsstress that every translation is an interpretation and inevitably incorporatesdifference. The translator ’ s ethical dilemma is then how to choose and trans-mit a good – or the best – interpretation. To represent is to interpret.If a translation mis represents the Other, the result may be a prejudiced,biased, ideologicallysuspect version, which has unethical consequences forintercultural perceptions and relations. An ethics of representation thushighlights the values of fidelity and truth: the translator must represent thesource text, or source author ’ s intention, or even the source culture, faithfullyand truly, like a good mirror. So-called ‘ abusive fidelity ’ (Lewis 1985) allowsthe Other to appea r as Other, as different, in a foreignizing or minoritizingtranslation. Postmodern approaches problematize the possibility of faithfulrepresentation and stress the ambivalence of the relations between sourceand target texts and cultures (see, for example, Koskinen 2000b), but heretoo the central ethical problem is fundamentally one of representation, of something ‘ standing for ’ something else. In this sense, the representationmodel of ethics is actually a semiotic one: a translation is a sign of the original.Ethics of service. A very different kind of approach is based on the conceptof translation as a commercial service, performed for a client. This is thekind of ethics that underlies (usually implicitly) much of the thinking onfunctional models of translation, especially those of Holz-M äntt äri (1984)and the Skopos theorists. A translator is deemed to act ethically if thetranslation complies with the instructions set by the client and fulfils the aimof the translation as set by the client and accepted or negotiated by thetranslator. A prime quality of good translator-servants is thus loyalty; theyare loyal above all to the client, but also to the target readers and to the originalwriter (cf. Nord 1991:29). They are also efficient (they do not waste time ormoney –especially the client ’ s), and perhaps also as invisible as possible.They provide a commercial service, and clients expect value for money. Anethics of service also underlines the value of time, of meeting deadlines, etc.Ethics of communication. This general model of ethics has received moreattention during the late 20th century, for example in the work of Levinas.Here the emphasis is not on representing the Other but on communicating with others. To recognize the Other as a ‘ subject ’ , with whom one can in-deed communicate, is a primary ethical act, for Levinas, because this steptakes you out of your own ego-confined world (for an accessible but verybrief introduction to Levinas, see Melby 1995:119f.; see also Levinas 1982,1987). In translation theory, the focus is naturally on communicating acrosslinguistic or cultural boundaries.A recent contribution that highlights the ethical aspects of such commu-nication is that of Pym (1997, 2000; for a detailed critique see Koskinen2000b). For Pym, the goal of cross-cultural communication is the mutualbenefit deriving from cooperation, and the ethical goal of translation is tofurther intercultural cooperation between parties who are “ Other ” to eachother. An ethical translator therefore translates in such a way as to optimizethis cooperation. (An ethical translator might also decide, notes Pym, that itwould sometimes be more beneficial not to translate at all, but recommendsome other means of eventually communicating, such as learning the otherlanguage.) An ethical translator ’ s primary loyalty, on this view, is to the trans-lator ’ s profession, situated in an intercultural space, and hence to the wholesystem that makes cross-cultural communication possible, rather than to sourcetext or culture or to target readers or culture. For Pym, the investment intranslation (by client or translator) should not exceed the eventual mutualbenefits accruing – here his position is similar to that taken by an ethics of service.From the point of view of communication, the ethical translator is a me-diator working to achieve cross-cultural understanding. Understanding of what? Of each other, ultimately. But this is achieved via an understanding of texts, messages, signs, intentions, meanings, etc. There is a well-known con-ceptual minefield here, which I will not explore in this context, but with respectto translation we can usefully speak of understanding in the following sense:understanding a translation means arriving at an interpretation that is com-patible with the communicative intention of the author and the translator (andin some cases also the client), to a degree sufficient for a given purpose.Norm-based ethics. This model of translation ethics has arisen either explic-itly or implicitly from descriptive translation studies and norm theory.Following Toury (e.g. 1995), descriptive translation studies investigates thenorms that determine or influence translation production and reception. Thesenorms state what acceptable translation products should look like, and howthey vary from period to period and from culture to culture. The norms thusrepresent expectations, mainly in the target culture, about what translationsare supposed to be like in that culture at that time. The norms are generallyaccepted (in a particular culture) insofar as they appear to serve prevailingvalues, including ethical values such as truth and trust (see Chesterman1997:169f.). Behaving ethically thus means behaving as one is expected tobehave, in accordance with the norms, not surprising the reader or client.Any major breach of these expectations – for instance, a translation that isclearly more literal than the reader might expect, or one that has a specificideological slant, or is abridged or extensively adapted – should, on this view,be signalled overtly by the translator, for example in a preface. One of thecentral values underlying this model is that of trust: if translators behave inpredictable, norm-conforming ways, it is easier to trust them – and the pro-fession as a whole.2. ProblemsThere are several reasons for the current unease in translation studies aboutquestions of ethics. One is the lack of compatibility between available mod-els. Each of the four models outlined above highlights different ethical values:truth (representation), loyalty (service), understanding (communication), trust(norm-based). Are some values higher than others? In the sense that somemay depend on or promote others, the answer to this question is perhaps ‘ yes ’ . People who speak the truth are more likely to be trusted than thosewho do not. People who are loyal are also likely to be trusted by those whomthey are loyal to (but not necessarily by outsiders; indeed, an outsider mighthave good reason not to trust them). Truth and trust may lead to understand-ing. However, none of the models is very clear about what the appropriatelyethical action might be in a situation where values (or loyalties) clash. Onwhat grounds can we simply say: choose this model when it seems appropri-ate, and that model at other times? Would such a solution itself be ethical?The different models do nevertheless have different scopes and limita-tions of application: some models have been applied more to literary or Biblicaltranslation, others to technical or administrative translation. The representa-tion model is vulnerable to arguments about the impossibility of totally truerepresentation, about the relative status of originals and translations, aboutthe illusion of perfect equivalence. How might it be applied to translationtasks that call forradical rewriting or adaptation or improvement to the text?For some scholars, the representation model nevertheless seems to be theonly one, since ethical problems are sometimes discussed almost entirelyfrom this angle (for example, Lane-Mercier 1997).The service model stresses the translator ’ s expertise, but also seems tomake a virtue of translatorial invisibility, weakening the translator ’ s autonomyto some extent. One might even argue that it can promote a mercenary atti-tude and a meek and passive habitus. How might it be applied to a freelancetranslator initiating the translation of an avant-garde Italian poet?The communication model risks expanding the translator ’ s responsibilityto cover aspects of cross-cultural relations that may have more to do withclients and readers than with the translator. Suppose the cross-cultural un-derstanding and cooperation is successful but promotes evidently unethicalends, as for instance in the task of translating instructions for making a cheap nail-bomb? Is this also the translator ’ s responsibility? How do we apply com-munication ethics to an EU situation in which a document may not betranslated in order to be actually read by anyone, but simply in order to existlegitimately, for political and ideological reasons, in another language?The norm-based model seems unduly conservative, underplaying the pos-sibility of change or improvement; but norms do change over time, partly asa result of translatorial action. How might it be applied to a situation where atranslator – perhaps even as requested by the client – seeks to surprise thereaders, to challenge their expectations, and wishes to strengthen this effectby declining to include a warning preface?A further difficulty is the way different models focus on different levelsof ethics. The norm-based model and to some extent the representation modeloperate mainly on the micro-ethical level, concerning the relation betweenthe translator and the text. The other models look more to the macro-ethicallevel, at the relation between the translator and the wider world. How are weto decide where the ethical responsibility of the translator stops – or does itstop at all? Translators are of course responsible for the words they choose towrite, but to what extent are they responsible for the effects these words mayhave? On what readers? Readers may be other than intended ones, after all;and readers of some future generation may react very differently to the samewords. Many scholars over the past few years have pointed to the wider cul-tural and ideological implications of translatorial decisions. How do we definethe limits of the translator ’ s responsibility?And what of the world ’ s responsibility towards translators? This aspecttoo might be considere d to belong to a general ethics of translation andtranslatorial behaviour (see, for instance, UNESCO ’ s Nairobi Recommen-dation). However, I shall not pursue this theme in this article.Our four models also differ with respect to the basic kinds of ethics theyespouse. The service and norm-based models are both examples of contrac-tual ethics. That is, ethical decisions here are based on prior agreements,contracts, expectations, either explicit or internalized; unethical decisions arecriticized because they break a norm or contract. So I act like this becausethis is the norm, because this is the way the provider of a service shouldbehave, this is what the translation instructions say. On the other hand, therepresentation and especially the communication modelsare examples of utili-tarian ethics. That is, ethical decisions are based on their predicted results;unethical decisions can be criticized because of their undesired results. So Iact like this because I want to facilitate communication, or to improve cross-cultural relations: I want to have this kind of effect on you, the reader.A good illustration of the problem of incompatible ethical models is to befound in the different interpretations of the value of clarity. For many trans-lators, clarity is an ethical value that they seek to promote in their work. It isan ethical value for many theorists too, myself included; and even for trans-lation service administrators –see, for example, the current ‘ Fight the Fog’campaign in the EU, which seeks to promote clarity both in original docu-ments and in translations. To be unclear is felt to be a betrayal of loyalty tothe reader, and also to the client, who presumably wishes readers to under-stand a translation. Popper (1945/1962:308) has even argued that clarity is aprecondition for all rational communication, without which society cannotexist. But what is meant by clarity?Different interpretations are offered by our different models of ethics. Inthe representation model, clarity means transparency: the translation shouldbe such that the original is clearly visible, the Other is clearly present, repre-sented as such, undistorted. This would usually lead to some kind of foreignizing translation. On the other hand, if the desired representation is of the author ’ s intention rather than the source text as such, it is this intentionthat should presumably be made transparently visible, and the result mightwell be a domesticating translation. Returning to the mirror metaphor, claritymeans that the mirror must be clean and non-distorting, reflecting whateverit is intended to reflect (intended by whom?).In the service and commmunication models, clarity is interpreted as ac-cessibility. This is a textual quality determining the ease with which readerscan understand a text, its meaning, the message, the author ’ s intention. Aclear translation, in this sense, is one that can be understood without unduetime and effort. Here, the degree of clarity affects the relation between trans-lation and reader, not translation and original.In the norm-based model, clarity is relativized to target culture expecta-tions: the form and degree of the required clarity depends on theseexpectations. Critics of suggestions that clarity is a universal communicationvalue usually interpret cl arity as meaning ‘ directness ’ , i.e. the absence of features such as irony or understatement. These critics then accuse Grice orLeech, for instance, of making Western Protestant discourse values intouniversals. None of our ethical models seem to take this view of clarity, how-ever. Indeed, the norm-based model specifically rejects it. This model says:be as clear as the situation demands, in the way that your readers will expect.In other words, do not be so unclear that your text is unacceptably inaccessi-ble to the people whom you would like to read it. The model stresses that theimplementation of the value of clarity is context-bound and also culture-bound.All in all, these four models are only partial ones; each covers only part of the general ethical field of translation, and each seems therefore inadequateon its own. Maybe we should go back to the beginning and start again. Thefollowing section explores an alternative route to an ethics oftranslation.3. The deontic force of excellenceWe might start with virtues rather than values. This is the position taken bythe philosopher Alasdair MacIntyre (1981), who argues interestingly against founding a general ethics on values, or indeed on rights, because of the irrec-oncilable conflicts between values and the lack of any rational way of prioritizing one right over another. MacIntyre ’ s suggestion is to return to thenotion of a ‘ virtue ’ , which he analyzes in terms of social roles. Social roles,in turn, are embedded in social practices – and this is where his ideas seemhighly relevant to any attempt to construct an ethics for a profession.MacIntyre defines a practice (not, it must be admitted, very clearly) asfollows:By a ‘ practice ’ I am going to mean any coherent and complex formof socially esta blished cooperative human activity through whichgoods internal to that course of activity are realised in the course of trying to achieve those standards of excellence which are appropriateto, and partially definitive of, that form of activity, with the resultthat human powers to achieve excellence, and human conceptions of the ends and goods involved, are systematically extended. (1981:175)Let ’ s unpack that a bit. The key ideas are cooperative human activity and astriving for excellence . Paraphrasing, we could say that a practice is the kindof cooperative activity that involves the desire to get better and better at it.Examples given by MacIntyre are football, chess, architecture, farming, phys-ics, medicine, painting, music, politics. Thus defined, all practices involvehuman relationships.Excelling in a practice brings a sense of satisfaction in its own right (whatMacIntyre calls ‘ internal goods ’ ), quite apart from any external benefits.Entering into a practice means entering into a relationship with its historyand tradition (its narrative, in fact) and its contemporary practitioners. It also means accepting the authority of prevailing standards of excellence (at leastinitially), and striving to achieve them, even to exceed them. Achieving thisexcellence not only enriches the person involved but also the community atlarge. Roughly speaking, a virtue can then be defined as an acquired humanquality that helps a person strive for excellence in a practice (MacIntyre1981:178). Such virtues include trustworthiness, truthfulness, fairness, and the courage to take risks in caring for others. Apart from virtues, practicesalso involve purely technical skills. And they need institutions to support them.How does this point of view help us in searching for a professional ethicsof translation? Some would argue that translation is not a true profession inthe first place, because it does not seem to have a monopoly on a value goal that is not shared by other groups (compare medicine, with the value goal of health; law, for justice; teaching, for human growth; and the police, for secu-rity) (Airaksinen 1991/1993). After all, the values of cross-culturalunderstanding and cooperation are also shared by such people as diplomatsand language-teachers. But translation is clearly a practice. As such, it is cooperative, involves technical skills, is increasingly institutionalized, andseeks its own improvement via quality control systems and the training andaccrediting of recruits. Its institutionalization serves to some extent to pro-tect it and limit its legal accountability.On the other hand, its authority to restrict accreditation is very limited:anyone can set themselves up as a translator – another reason why it is diffi-cult to speak of translation as a true profession. When we speak of an ethicsof translation, do we mean to include amateurs as well as ‘ professionals ’ ?One way of answering this question would be to distinguish between some-one ‘ who is a translator ’ and someone ‘ who does translations (sometimes) ’ .Iwill focus in what follows on the first category, on the practitioner ratherthan the practice.Consider now how this notion of a practitioner naturally leads to the deonticlevel. If I say Louis is a translator, I imply that Louis therefore ought to dowhat a translator ought to do. I further imply that Louis should know what atranslator ought to do, i.e. what a good translator would in fact do; Louis thusmust have some conception of excellence in the practice of translation, somemental image of ‘ a good translator ’ . If I wanted to deny this further implica-tion, I would need to say explicitly that Louis is a bad translator. MacIntyreargues that in this way, with respect to “ functional concepts ” , ‘ ought ’ canindeed be derived from ‘ is ’ (see also C hesterman 1993). Different translatorsmight of course have different mental images of ‘ a good translator ’ , but any-one who calls themselves a translator must have some idea of what translatorsought to do.I have elsewhere shown how a simple deontic logic can be applied totranslation ethics (Chesterman 1997:172f). Roughly speaking, you discoverwhat you ought to do by assessing the outcome of a possible action in termsof the values that are thereby promoted, and comparing this outcome to whatwould probably happen if you did not take the action in question. You thendo whatever leads to the best prospective result: either that particular action,or not. In that analysis, I proposed that the relevant values guiding translatorialdecisions were truth, clarity, understanding and trust; and I showed how eachof these values governed a general translation norm. In the present context, Iwould like to focus not on the values themselves, which act as regulativeideas steering the process of ethical decision-making, but on the qualities of the decision-maker.What virtues must our ethical decision-making translator possess in orderto make the best ethical decisions? Following MacIntyre, I suggest that themost important virtue is simply the desire to make the right decision; that is,the translator must want to be a good translator, must strive for excellence inthe practice of translation. It is interesting to compare this point to the answerMacIntyre (like many others) gives to the perennial question about how tolead a good life: the good life is a life spent seeking the good.Then there are the general virtues I have already mentioned, such as fair-ness (the comparative value assessment of alternative actions must not bedeliberately biased), truthfulness (the assessment must be as honest as possi-ble), and trustworthiness (the translator must be able to defend the decisionstaken, to give evidence of reliability). To these we might also add empathy,i.e. the ability to put oneself in someone else ’ s place –the reader ’ s, the origi-nal author ’ s, the client ’ s – so as to imagine the possible effects of alternativeactions. Then there are the virtues of courage and determination not to giveup until a good solution to a translation problem is found.Other necessary qualities are not virtues as such but play a supportiverole in the striving for excellence. The translator must have adequate knowl-edge of the alternatives available: this meansobvious language skills, includingcontrastive linguistic and cultural knowledge, in order to assess the potentialeffects of different choices, in the widest sense. The translator must also haveadequate technical and research skills in order to discover and evaluate pos-sible alternatives.A first conclusion to be drawn from this approach is that it allows us torestrict the scope of professional ethics to the practice in question. Louis canbe an ethically good translator even if he is not a member of Amnesty Inter-national, Animal Rights, or the Labour Party. This means that the politicalengagement of the translator, which may affect the choice of texts to be trans-lated and also ways of translating them (see Tymoczko 2000), lies outsidethe realm of professional ethics. A translator may be actively engaged insupport of a worthy cause, and may translate in such a way as to support thiscause, but these are factors that are additional to professional ethics proper,not part of them. If Louis is an anti-fascist, and subverts the fascist texts hetranslates, he is allowing his personal ethics to dominate his professionalethics. Fair enough: we must acknowledge that there may sometimes be moreimportant things than professional ethics. It is part of a bus-driver ’ s profes-sional ethics to observe the Highway Code of traffic rules, but in an emergency,to save a life, these might be broken. Professional ethics, thus understood,govern a translator ’ s activities qua translator, not qua political activist orlife-saver.A second conclusion is that translation ethics might be defined in termsof excellence in this practice. I now want to explore a further aspect of thisexcellence, an aspect that characterizes the relation between the practitionerand the values that inspire the practice. This is the notion of commitment.4. An ethics of commitmentI take commitment to be the glue that binds practitioners to the values of thepractice. It is thus alsoa virtue, supporting the striving for excellence, the wanting to be a good translator. A commitment is often stated overtly, as a promise or oath –in the marriage service, for instance. Oaths are quintessen-tially statements of contractual ethics: they constitute contracts, bindingpromises; but they may also have utilitarian aspects, such as reference todesirable or undesirable results. Let us look at two examples of oaths of com-mitment to a practice. The first example is one of the oldest of all –the medicalprofession ’ s Hippocratic Oath. Here it is in full, with paragraph numbers added:The Hippocratic Oath(1)I swear by Apollo the healer, by Aesculapius, by Health and all the pow-ers of healing, and call to witness all the gods and goddesses that I maykeep this Oath and Promise to the best of my ability and judgment.(2)I will pay the same respect to my master in the Science as to my parentsand share my life with him and pay all my debts to him. I will regard hissons as my brothers and teach them the Science, if they desire to learn it,without fee or contract. I will hand on precepts, lectures, and all otherlearning to my sons, to those of my master and to those pupils duly ap-prenticed and sworn, and to none other.(3)I will use my power to help the sick to the best of my ability and judg-ment; I will abstain from harming or wrongdoing any man by it.。

法国欧创_航空制动器

Recent Advances in Aerospace Actuation Systems and Components, June 13-14, 2012, Toulouse, France Contribution to aircraft actuators network analysis thanks to co-simulationBAREILLE MichelAltran Sud-Ouest4, Avenue Didier Daurat31700 Blagnac, FrancePhone: +33 (0)6 31 91 88 46 Email: michel.bareille@SNEESSENS Charles-HenriAltran Sud-Ouest4, Avenue Didier Daurat31700 Blagnac, FrancePhone: +33 (0)6 32 24 19 77Email:charleshenri.sneessens@ROUDIER ThierryKiasTek31, Avenue JF Champollion31000 Toulouse, FrancePhone: +33 (0)6 65 63 17 53Email: thierry.roudier@ABSTRACTNowadays, in the general trend of a more electrical aircraft, designers face an increasing number of constraints to predict the performances of their system in its global environment. Co-simulation is an efficient way to overcome this issue: it allows estimating the behaviour of the equipment by basing itself on models that are implemented in their native software so that there is no major adaptation needed to perform a genuine multi-physics simulation. Two methods are possible to reach this objective: bridge or bus. The object of this paper is therefore to present both techniques with two different examples involving actuators. This will show the interest of co-simulation with practical ways to implement it. The final objective is to provide a solution adapted to each one’s need in order to have a valuable insight on a designer’s product in its real environment.KEYWORDSActuator, co-simulation, hydraulic, mechanical, electrical, electromechanicalI INTRODUCTIONFor several years the digital transient simulation has become essential for designing embedded systems. Over the years, several software solutions are born, each with its strengths and weaknesses. Most of these solutions are generally used to model and simulate a physic or another. For example, even if it cannot be generalized, Saber® [SAB 11] from Synopsys is often used to simulate electrical systems, or b AMESim® [IMA 11] from LMS International is mainly used to simulate hydraulic systems. In conclusion of this observation, we notice that more than one solution is used in a company.Moreover, the increasing complexity of multi-physics systems, and the constant pressure to reduce development costs have led mechanical, software, electronics engineering teams to develop more complete models with a finer level of granularity involving a notion of strong coupling (with adjacent systems implementing different physic). This makes the functional verification of these models impossible without the sub-communicating systems: mechanical, hydraulic, electrical or software. Therefore, the idea consisting in simulating all those models together is quickly becoming a necessity for faster development and requirement match analysis.Figure 1. V-cycleBut in most cases, designers have to wait for a real prototype of the system before they can functionally verify their models. In recent years, several solutions have emerged such as:•HIL (Hardware In the Loop) simulation platforms,•models import into a single simulation environment,•the description of models in a single language,•and co-simulation methods.HIL platforms are expensive and can only be used if electronic equipment (typically microcontrollers) and software code are available. Importing models usually uses code generation methods involving a translation of models which is therefore prone to errors. Modelling in a single language such as Modelica [MOD 12] has the advantage to provide a unique and multi-physic modelling and simulation framework to design the complete system. Nevertheless, there are advantages to use tools dedicated to a physic or another like b AMESim®, or Saber®. One of these advantages has to be seen from a practical point of view. Indeed, designing a system generally involves a large number of engineering teams and suppliers coming from various horizons. Due to the variety of actors involved in the design process, and because each team has already chosen a modelling and simulation solution, which is certainly different from the solution used by another team, it is quite difficult to imagine a unique solution or language to design a complete system.Co-simulating a numerical system consists of simulating its different component subsystems simultaneously, by executing the models of the different subsystems and coordinating the exchange and interpretation of data calculated or required by each one of them. Therefore, by definition, co-simulation is the most flexible method guaranteeing the simulation of the models in their native environment such as Simulink® [SIM 12], Saber®, etc... The solution allows a verification of a complete system without modifying the models of the system. Indeed, a co-simulation environment considers all adjacent models, as they are defined, by coupling the system using synchronization and complex communication methods.II CO-SIMULATION SOLUTIONS2.1 Co-simulation over bridgesBridges can be considered as point to point communication media between two different simulation tools. They are often developed and distributed by the simulation environment provider. For example Synopsys provides a Simulink® bridge to connect with Saber®. LMS International provides an Adams® [ADA 12] bridge (MSC-Software), and a Simulink® bridge to connect with b AMESim®. Because of their high integration with the native simulation environment, bridges generally provide accurate and fast communication links with the connected simulation environments. Nevertheless, several constraints and limitations have to be considered as:•Most of known bridges cannot instantiate more than one instance of a simulator during a co-simulation.• A notion of master/slave natively makes simulators dependent of the simulator driving the co-simulation. In some use cases, such a configurationmay freeze the co-simulation because of a non-synchronization of the solvers.•To build a complex co-simulation platform, several bridges developed by various tool vendors have tobe configured. The diversity of the environmentsconsiderably increases the complexity of using andmaking reusable such a simulation platform.2.2 Bus-based co-simulationA co-simulation bus, as CosiMate® [COS 11], can be considered as a communication controller where all models of a system are simulated in their own environment interacting together using complex synchronization methods. Each couple simulator/model involved in the co-simulation operates as an independent process. In opposition to the bridge-based solution, each couple simulator/model is the master of its own simulation: the co-simulation bus does not force to compute a specific step of computation, but acts as a complex communication arbitrator between the co-simulation attendees. In other words, the co-simulation bus reacts upon requests from the simulation environment. For example, when a simulation environment sends information to the bus as “a computation step is done, waiting for new inputs to compute the next step”, then the co-simulation controller synchronizes with the connected models, exchanges data for the current step, or computes new input values based on variables history (extrapolation method) if data are not available on the bus. Such an implementation provides a lot of flexibility to synchronize the simulators/ models together in an efficient way to guarantee the system simulation convergence and high speed simulations.Figure 2. Co-simulation frameworkIII CO-SIMULATION FEATURES3.1 Synchronization methodsThe co-simulation framework, CosiMate®, provides two synchronization methods: event-driven and synchronized modes, which perfectly fit with the heterogeneity of various systems involved in a simulation platform. Indeed, these methods take account of the behavioural differences of languages and the simulators used.With the event-driven co-simulation mode, the co-simulation bus does not process any notion of time. Data are transmitted as soon as they are available on the co-simulation bus. Sucha method allows the connection between event-driven (HDL simulator) and sequential (C code) simulation environments. The synchronized or timed mode was developed to interconnect solvers (Simulink®, Saber®…) together. The method implements a synchronization of thecommunications based on multiple clock rates. Each clock rate Tn represents the desired frequency to synchronize a given couple simulator/model with the co-simulation bus. The co-simulation router, bus controller, verifies all clock rates, and exchanges data on the bus. Then the router accordingly synchronizes each pair of connected simulators.Figure 3. Synchronization exampleThis simple method of synchronization, based on multiple fixed communication frequencies, mainly answers the needs of the various solver types, continuous or discrete [NIC 06], fixed or variable steps, involved in a co-simulation. However a wrong choice of a communication frequency may imply sometimes a non-convergence of the system simulation, or may result in long simulations. To avoid this kind of behaviour and also to get the best compromise between simulation speed and accuracy, we have introduced with CosiMate® the data prediction algorithm (DPA). The concept uses polynomial or more advanced extrapolation methods in order to calculate inputs of a model between two successive synchronization. i.e. when new values are not available on the bus. The combination of both methods: synchronization and extrapolation, provides more flexibility to the end user in the choice of the bus frequencies without deteriorating the results of the co-simulation.IV ELECTROMECHANICAL ACTUATORS SYSTEMThis first application deals with a complete architecture of multiple actuators as a flight control system. Our study concerns an electromechanical actuator (EMA). The system is described in fig 9 and is a common solution for aerospace application.Figure 4. Electromechanical actuator system Regarding the different physics of the systems, for this particular study, we focus on electrical, mechanical and control parts of the problem. The system model is composed of two parts•Electrical part with Saber® software•Mechanical and control parts with Matlab®/ Simulink®.The aim of this study is to highlight if there is any interest in making co-simulation for EMA’s application in a flight control environment and to define the methodology to use. Priority is given to dynamic phenomena modelling linked to the electrical and mechanical part. The objectives are to define a more accurate model for each physic, to study the behaviour and to perform a tuning of the parameters of the system with time saving and a better model accuracy. Those models will be interesting to study with a complete architecture of several actuators; time saving will become more important in this case.As a Saber® module already exists in order to connect Saber® and Simulink®, we begin by analyzing the bridge approach.5.1 Co-simulation interfaceThe definition of the interface is important and will influence the results quality and computation time. Systems behaviours are taken into account by switching to physics models but also by the properties of the bridge solution, e.a time step, communication and synchronization methods. In this case, we used a simplified functional model with a simplified inverter (since the inverter frequency requires a higher frequency exchange). For the behavioural inverter simulation, the choice of modelling the inverter in Saber® can be done to obtain better accuracy.Figure 5 Co-simulation Interface5.2 Parameters influenceDepending on the objective of modelling the system and the use of results we need different accuracies. As we have a master/slave couple, in our case Saber® leads the simulation. The exchange step is the most important parameter to define since it greatly influences the simulation results. During a step Saber® will send to Matlab the same constant valueeven if Saber® data changes.Figure 6 Exchange step influenceThe computation method can also be decisive since some methods will not converge to the results. This method will also influence the computation time and depending on the model’s complexity the method must be adapted. We are trying to reduce computation time while maintaining results accuracy by this way.5.3 Preliminary resultsTo evaluate the bridge method, the first case studied is the functional model with one actuator. The results are quite satisfying. With 1e-3 step size we reduce the computation time by at least 20% with very little differences on co-simulation results (in red).Figure 7 Difference on co-simulation results (in red)The second case concerns a multiple actuator network. The results are similar and the computation time is reduced by at least 15% compared to full Saber® model.Expectations with the bus methodology might bring better results if considering the next example.5.4 Next stepsTo make a full study of co-simulation for this actuator application, we plan to realize a simulation with a bus approach. Then, we’ll define some guidelines to optimize the model for co-simulation purpose and network application. A behavioural model will also be tested and strong case studies will be deployed. Finally, we plan to use those techniques for operational application.V HYDRAULIC / MECHANICAL ACTUATORS SYSTEMThis application test bench concerns an integration study and performances assessment of a complex hydraulic / mechanical system. The objective here is not to develop a model in order to validate a design but to make an existing heterogeneous simulation platform work in order to verify the performances of an actuator’s system.It is composed of elementary models modelled in different packages:• Local hydraulics and Mechanical part onb AMESim® / Adams®,• Hydraulic general power network in Saber®.4.1 Bridge approachThe first approach to make all these models simulate together consisted in using Simulink as a “bridge” (communication interface) to exchange the data between concurrent software. The data exchange flow was the following:Figure 8. “Bridge” co-simulation platformThe exchanged variables are:• Forces from Adams® to b AMESim®, • Actuators velocity and position from bAMESim® to Adams®,• Flows from b AMESim® to Saber®, • Pressures from Saber® to b AMESim®.This first approach proved to be sufficient, although difficult to set up, for a first simulations loop. However, it was not robust with respect to input data modifications: at each new issue of the hydraulic or mechanical system, one had to re-set the simulation parameters (internal to each software, plus exchange rate) without any assurance that the simulation would reach the end of the cycle planned.4.2 Bus-based approachThat is why a more robust approach was investigated, based on the CosiMate® tool. The resulting simplified co-simulation platform is depicted hereunder:Figure 9“Bus based” co-simulation platformIn addition to this simulation architecture, some additional interfaces were also added to help the simulation converge: •Low pass filters on the b AMESim® side to smooth the waveforms received,•Predictors on the Saber® side to reproduce signal continuity between two exchange steps.The results show a real robustness improvement of the platform over complete simulation cycles and also with respect to models modification. The graphs below display the results obtained for pressure and flow compared the ones obtained with a standard “ping pong” technique (exchange of simulation results between Saber® and b AMESim® / Adams® and run of several loops to reach result stability):•In orange: ping pong results,•In black: co-simulation results.Figure 10. Pressure in forward and return lines withrespect to timeFigure 11. Flow in forward and return lines withrespect to timeThe curves obtained are quite comparable except for a small delay at the end of the cycle for the co-simulation. This can be explained by the fact that the simulation is not exactly identical in the ping pong technique and in the co-simulation. In particular, a valve limiting the flow in the forward line implemented in Saber® has been moved into b AMESim® for the ping pong. If the models are not strictly identical, the difference of behaviour could explain the graphs obtained.The point is that the time difference at the end of the simulation is 1% of the global cycle duration, which is acceptable considering that the two techniques are different. However, the most significant improvement with respect to the ping pong is the simulation time. In a platform composedof a computer which features are:•CPU: Intel core 2 duo T7500 @ 2.2Ghz•RAM: 3Gothe complete cycle duration is around 18h (black curves) whereas it took 5 iterations of 12h (60h) to obtain the ping pong ones (orange curves). The simulation time was therefore reduced by 70%.CONCLUSIONWith regard to the increasing complexity of actuator systems, co-simulation represents an efficient way to investigate difficult multi-physics issues. This paper presents two different techniques to implement it: by bridge or by bus. Co-simulation by bridge requires compatibility between packages provided by a given simulator or by dedicated modules within the tools. A bus avoids this issue by managing several packages together and by ensuring the correct data transmission. It applies therefore to complex simulation environments in which tools are not meant to work together. With simulation objectives strongly define, a solution can be provided thanks to co-simulation and this method can help designers dealing with multi-physic constraints.The perspectives brought by this technique are therefore numerous, not only from a technical point of view but also considering the design team relationship with its suppliers and customers.Indeed, today, the work sharing between companies concerning the development of large systems implies the delegation of bigger and bigger work packages to different suppliers. The most convenient way to verify the coherencyof the global design consists in letting each partner use its own optimized tool and to join them all eventually in a common co-simulation platform.This approach leads naturally to develop shared platforms and resources to allow fast and reliable simulations. This could be implemented in a unique workshop with the different tools installed in the same computer. The other solution is to use different computers, each running its own package in a distributed simulation approach. This configuration could be adapted to various industrial sectors and environments and is the subject of our next research axis. REFERENCESAdams (2012), /Products/CAE-Tools/Adams.aspx, Online documentation, MSC-Software, Santa Ana, United States, 2012CosiMate (2011.04), , Online documentation, KiasTek, Toulouse, France, 2012 Modelica (2012), , Online documentation, Modelica Association, 2012Nicolescu G., Bouchhima F. and Gheorghe L. (2006),CODIS – A Framework for Continuous/Discrete SystemsCo-Simulation, Proceedings of the 2nd IFAC conference on Analysis and Design Hybrid Systems, Alghero, Italy, June 7–9, 2006b AMESim (Rev11), /LMS-Imagine-Lab-AMESim, Online documentation, LMS International, Leuven, Belgium, 2012Saber (2011.09), /Systems/Saber, Online documentation, Synopsys, Mountain View, CA, United States, 2012Simulink (R2012a), /products /simulink/, Online documentation, The Mathworks, Natick, MA, United States, 2012NOTATIONSVariables and parametersTn Communication time step (s)T’n Computation time step (s)ACKNOWLEDGEMENTSThe authors want to thank particularly David Sillonnet ( Altran Consultant in electrical department) and Clément Mur (Trainee in Altran actuator team) who participate to this subject through simulation and modelling activities.。

notesontranslationcriticism

NOTES ON TRANSLATION CRITICISMModel, Newmark Textbooksources: House ‘Quality’, HouseA criticism of a translation is different from a review of a translation.Review = comment on new translations, description and evaluation as to whether they are worth reading and buyingCriticism = a broader activity, analysis in detail, evaluating old and new translations ,assuming that readers know the translationTranslation criticism should take into account all the factors and elements in the process of translation (translation as a communicative act: intention, function, text tupe,register, strategies, principles, rules, constraints, audience)It comprises activities which are part of the process of translation (analysis andinterpretation of the ST), but it is different from the forms of criticism involved in thisprocessTranslation criticism should not be a mere identification of errors, an intuitive or highlysubjective appraisal judging translations as ‘good’, ‘bad’. ‘faithful’ withou these adjectives.Similarly, reviews should-describe the quality of a translation with more than a single adjective and- refrain from trashing the translator’s work on the basis of isolated errorsCriticisim of translation quality should be grounded on thorough analysis and descriptionSome critics prefer to eschew value judgements, prefer not to proclaim one translationbetter than another (Hatim and Mason 1990b: 1)More concern with understanding how translated texts work (rather than with traditionalcponcepts of quality) and seek to define the translator’s method (Vilikovsky) and purpose (Newmark: 1998: 75)There is discussion about whether evaluation should take into account the Source Text : Toury notes that comparisons between translations and originals often lead to anenumeration of errors and a reverence for the original (1978: 26)Most critics carry out a comparative criticism of original and translation? Newmark (Textbook): five-part modelanalysis of source textcomparison of it and the translationcomments about the translation’s potential role as a translation? Hatim and Mason (1990b) outline a set of comparative parameters; their principalinterest lies in the ‘cultural semiotics of language’．Using the notions of genre, discourse and text, they focus not on individualwords but on a ‘thread of discourse which is sustained through acommunicative transaction’ (10)? de Beaugrande (1978) evaluative criteria should address the ‘presuppostions and expectations about texts’ shared by readers and writers in each languageNon-comparative modelsLefevere (1981b) focus on the product of translation in the context of the targetculture rather than on the translation process (see polysystem theory)Toury (1978, 1980c) his work with translational norms suggests evaluative centred onthe target system aloneCriticism should take into account the presence of ideology in translation.Critics may also have their own hidden ideology conditioning their criticismA reviewer’s motiviation may be political, or of other nature.For instance, in his study pf Matthew Arnold’s lecture ‘On Translating Homer”，Venuti (1995: 118-45) has shown not only that Arnold’s attack on FrancisNewman’s translation of the Iliad served to marginalize Newman’s work, butalso the extent to which a polemics about acceptable translation strategies canbu simultaneously one about cultural politics.Hamlet][Cladera on Moratín’s translation ofCriticism of translations can be found intranslator’s prefaces and annotations (many new translations try to improve orrectify previous translations; prefaces and annotation contain evaluativecomments)complimentary poems and essays about the work of other translators (often inmetaphorical language -> they must be read in the context of prevailingrhetorical conventions),scholarly writing about translation theory, andappraisals embedded in fictional commentary (a simile from Don Quixote that likens works in translation to the wrong side of a Flemish tapestry provides Cervanteswith the opportunity to pass judgment on his contemporaries -Moner 1990: 519-22)Criteria for the evaluation of translationThey depend on one’s view of or approach to translation, on one’s theory of translation. Communicative approach:They focus on determining the ‘dynamic equivalence’ (Nida 1964) between source and translation‘dynamic equivalence’= the manner in which receptors of the translated text respond to it must beequivalent to the manner in which the receptors of the source text respond to the source textNida postulated three criteria for an optimal translation-general efficiency of the communicative process-comprehension of intent-equivalence of responseFor House (‘Quality’), these criteria prove to be as vague and non-verifiable asthose by the intuitive-anecdotal approach.A functional-pragmatic modelA model that attempts to avoid anecdotalism, reductionism, programmatic statements andintuitively implausible one-sided considerations of the ST and TT aloneRevisited): a model based on pragmatic theories of language.House (‘Quality’,Analysis of linguistic-situational particularities of the source and target textsA comparison of th two textsAn assessment of their relative matchThe basic requirement for equivalence is that the translationshould have a function which is equivalent to that of the original, (function = consistsof an ideational and an interpersonal functional component, in Halliday’s se should also employ equivalent pragmatic means for achieving that functionInitial analysis of the original according to a set of situational dimensions, for whichlinguistic correlates are establishedThe resulting textual profile of the original characterizes its functionThe function is the norm against the which the translation is measuredAnalogous analysis on the translationFrom this analysis derives the textual profile and function of the translationComparison of both the original’s and the translation’s textual profiles and functions.The degree to which the textual profile and function of the translation match theprofile and function of the original is the degree to which the translation isadequate in qualityIn this comparison, some mismatches will occur. Two kinds of mismatchesDimensional mismatches: pragmatic errors that have to do with language users andlanguage useNon-dimensional mismatches: in the denotative meanings of original and translationelements and breaches of the target language system at various levelsThe final qualitative judgment consists of a listing of both types of errors and of astatement of the relative match of the two functional componentsNewmark (Textbook chapter 17 “Translation Criticism”）Criteria:-accuracy, economy, both according to the translator’s standards and to the critic’ｓstandards,-without reference to the SL: smoothness, naturalness, easy flow, readability, absence ofinterferencePlan:1. analysis of ST stressing its intention and functional aspects2. analysis of-the translator’s interpretation of the SL text’s purpose,-his translation method and-the translation’s likely readeship3. selective but representative detailed comparison of ST and TT4. evaluation of the translation (a) in the translator’s terms, (b) in the critic’s5. where appropriate, an assessment of the likely place of the translation in the TL cultureor discipline1. analysis of ST :- author’s purpose, his attitude toward the topic- characterisation of the readership- category and type of text- quality of the language to determine the translator’s degree of licence informative text -> clichés [metaphor that have perhaps temporarily outlivedtheir usefulness, that are used as a substitute for clear thought, oftenemotively, but wihout corresponding to the facts of the matter, set trends, ajewel in the crown, ] are reduced to neutral languageauthoritative texts -> clichés are retained- state the topic or themesunless they are referred to (don’t discuss author’s life, works, general background,in the text)[underline particular problems posed by ST: title, structure, level of language,metaphors, cultural words, proper names, insitutional names, neologism,‘untranslatable words, technical terms, ambiguity, meta-language, puns, soundeffects,2. analysis of2.1 the translator’s purpose: – you should understand (not criticise) why he hasused procedures for a specific aimIs he deliberatly antiquating the language? moderating the figurative language?livening up simple sentences with colloquial and idiomatic phrases?Is he trying to counter the under-translating tendency of all translations bydeliberately over-translating?To what exent has the TT been deculturalised, or transferred to the TL culture?It is too easy to pounce on a translation’s howlers, listing them one after theother:-false friends, stretched synonyms, stiff or old-fashioned structure,anachronistic colloquialisms, literal translations of stock metaphorsIf you do so, you have to provide reasons why3. comparison[A translation critic determines the general properties -first of ST, and then of TT-and uses the underlined words (see Last reading p. 17 ) as a basis for a detailedcomparison of the two textsunderlined words: neologisms, metaphors, cultural words, and institutionalterms peculiar to the SL, proper names, technical terms and‘untranslatable words’ (the ones with no ready one-to-tone equivalent)]how the translator has solved the particular problems posed by STgroup problems under general heads:title, structure, shifts, metaphors, cultural words, translationese, propernames, neologism, ‘untranslatable’ words, ambiguity, level of language,meta-language, puns, sound-effectdiscuss problems and do not prescribe a correct or better translation4. evaluation of the translation- assess the referential and pragmatic accuracy of the TT by the translator’ｓstandardsIs the TT successful in its own terms)- assess the referential and pragmatic accuracy of the TT by your standardsassess the quality and extent of the semantic deficit in the TT. Was itinevitable, was it because of the translator’s deficiencies - assess the TT as a piece of writing, independently of the ST:in personal or authoritative text, has the translator captured the ideolect ofthe original?5. assessment of the importance of the translation in the TL culture or disciplinewas it in fact worth translating?what kind of influence will it have on the language, literature, the ideas in its newmelieu?Quality in translationA good translation fulfils its intentioninformative texts -> it conveys the facts acceptablyvocative text -> it has its purposed effectexpressive text ->judged ‘adequate’ if explains what the text is about (cf. many Penguin Plain Prose translations)judged ‘good’ if it is ‘distinguised, if the translator was excepcionally sensitiveas “form is almost as important as content, there is often a tension between theexpressive and the aesthetic functions of language and therefore a merely‘adequate’ translation may be useful to explain what the text is about (cf.many Penguin Plain Prose translations), but a good translation has to be‘distinguised’ and the translator exceptionally sensitive; for me, theexemplar is Andreas Mayor’s translation of Proust’sLe Temps retrouvé－” (p. 192)‘Time RegainedExamples of translation criticism in Part II, Text 10 to Text 13,。

“翻译学导论”期末考试题型及课程复习提纲

“翻译学导论”期末考试题型及课程复习提纲考试题型：I. Match the following translation theorists with their representative works. [10 points, 1 point each]II. Interpret the following terms. [20 points, 5 points each]III. Interpret the following concepts. [40 points, 20 points each]IV. Analyze and comment on the following translated version(s) by applying any translation theory you are familiar with. [30 points]复习提纲一、作者和代表作的搭配1. Roman Jakobson On Linguistic Aspects of Translation (1959)2. E.A. Nida Toward a Science of Translating (1964)3. E.A. Nida & Taber The Theory and Practice of Translation (1969)4. James Holmes The Name and Nature of Translation Studies (1972)5. George Steiner After Babel (1975)6. Susan Bassnett Translation Studies (1980)7. Peter Newmark Approaches to Translation (1981)8. Wolfram Wilss The Science of Translation (1982)9. Theo Hermans The Manipulation of Literature (1985)10. Peter Newmark A Textbook of Translation (1988)11. Mary Snell-Hornby Translation Studies: An Integrated Approach (1988）12. Christiane Nord Text Analysis in Translation (1989)13. Susan Bassnett & AndréLefevere Translation, History and Culture (1990)14. Andre Lefevere Translation, Rewriting, and the Manipulation of Literary Fame (1992)15. Edwin Gentzler Contemporary Translation Theories (1993)16. Gideon Toury Descriptive Translation Studies (1995)17. Lawrence Venuti The Translator’s Invisibility (1995/2008)18. Christiane Nord Translating as a Purposeful Activity (1997)19. Luise Von Flotow Translation and Gender (1997)20. Maria Tymoczko Translation in a Postcolonial Context (1999)二、解释以下术语1. word-for-word/literal translation and sense-for-sense/free translation2. metaphrase, paraphrase, imitation3. Formal equivalence and dynamic/functional equivlence4. Semantic Translation and Communicative Translation5. documentary translation and instrumental translation6. translational action7. skopo theory8. polysystem theory9. norms10. equivalence11. foreignization and domestication12. translation/translating13. Prescriptive Translation Studies and Descriptive Translation Studies三、理解以下观点1. Translating consists in reproducing in the receptor language the closest natural equivalent of thesource-language message, first in terms of meaning and secondly in terms of style. (Nida & Taber, The Theory and Practice of Translation, 1969: 12)2. 翻译（translating）即解释，具体地说，翻译的过程就是，在跨文化的历史语境中，具有历史性的译者使自己的视域与源语文本视域互相发生融合而形成新视域，并用浸润着目的语文化的语言符号将新视域重新固定下来形成新文本的过程。

4 学术论文摘要的模板和翻译

Abstract Sample Study
---Stylistic Features of Webchat English
• With the development of computers and the internet, webchat has become more and more popular. This particular communicative form generates some unique linguistic features which can hardly be found in other variants of English. Yet, to analyze its linguistic features from the point of view of stylistics is still a new realm. Adopting the theories of modern stylistics, this paper analyzes the English used in webchat from four levels: graphology, lexicon, syntax/grammar and context. (74 words)
摘要的理解
• 摘要是以提供文献内容梗概为目的，不加评论和补充解释，简明、确切地记述文献重要内容的短文。
• 摘要的基本要素：研究目的、方法、结果和结论。
Abstract or Summary?
• Abstract: placed before the text part (usual way)
2. Describing methodology: In this move the author gives a good indication of the experimental design, including information on the data, procedures or methods used and, if necessary, the scope of the research reported.

1、下载文档前请自行甄别文档内容的完整性，平台不提供额外的编辑、内容补充、找答案等附加服务。
2、"仅部分预览"的文档,不可在线预览部分如存在完整性等问题,可反馈申请退款(可完整预览的文档不适用该条件!)。
3、如文档侵犯您的权益，请联系客服反馈,我们会尽快为您处理(人工客服工作时间：9:00-18:30)。

Proceedingsofthe50thAnnualMeetingoftheAssociationforComputationalLinguistics,pages115–119,Jeju,RepublicofKorea,8-14July2012.c󰀂2012AssociationforComputationalLinguistics

TopicModelsforDynamicTranslationModelAdaptationVladimirEidelmanComputerScienceandUMIACSUniversityofMarylandCollegePark,MDvlad@umiacs.umd.eduJordanBoyd-GraberiSchoolandUMIACSUniversityofMarylandCollegePark,MDjbg@umiacs.umd.eduPhilipResnikLinguisticsandUMIACSUniversityofMarylandCollegePark,MDresnik@umd.edu

AbstractWeproposeanapproachthatbiasesmachinetranslationsystemstowardrelevanttransla-tionsbasedontopic-speciﬁccontexts,wheretopicsareinducedinanunsupervisedwayusingtopicmodels;thiscanbethoughtofasinducingsubcorporaforadaptationwith-outanyhumanannotation.Weusethesetopicdistributionstocomputetopic-dependentlex-icalweightingprobabilitiesanddirectlyin-corporatethemintoourtranslationmodelasfeatures.Conditioninglexicalprobabilitiesonthetopicbiasestranslationstowardtopic-relevantoutput,resultinginsigniﬁcantim-provementsofupto1BLEUand3TERonChinesetoEnglishtranslationoverastrongbaseline.

1IntroductionTheperformanceofastatisticalmachinetranslation(SMT)systemonatranslationtaskdependslargelyonthesuitabilityoftheavailableparalleltrainingdata.Domains(e.g.,newswirevs.blogs)mayvarywidelyintheirlexicalchoicesandstylisticprefer-ences,andwhatmaybepreferableinageneralset-ting,orinonedomain,isnotnecessarilypreferableinanotherdomain.Indeed,sometimesthedomaincanchangethemeaningofaphraseentirely.Inafoodrelatedcontext,theChinesesentence“粉丝很多”(“fˇens¯ihˇendu¯o”)wouldmean“Theyhavealotofvermicelli”;however,inaninformalIn-ternetconversation,thissentencewouldmean“Theyhavealotoffans”.Withoutthebroadercontext,itisimpossibletodeterminethecorrecttranslationinotherwiseidenticalsentences.

Thisproblemhasledtoasubstantialamountofrecentworkintryingtobias,oradapt,thetransla-tionmodel(TM)towardparticulardomainsofinter-est(Axelrodetal.,2011;Fosteretal.,2010;Snoveretal.,2008).1TheintuitionbehindTMadapta-tionistoincreasethelikelihoodofselectingrele-vantphrasesfortranslation.Matsoukasetal.(2009)introducedassigningapairofbinaryfeaturestoeachtrainingsentence,indicatingsentences’genreandcollectionasawaytocapturedomains.Theythenlearnamappingfromthesefeaturestosen-tenceweights,usethesentenceweightstobiasthemodelprobabilityestimatesandsubsequentlylearnthemodelweights.Assentenceweightswerefoundtobemostbeneﬁcialforlexicalweighting,Chiangetal.(2011)extendsthesamenotionofcondition-ingonprovenance(i.e.,theoriginofthetext)byre-movingtheseparatemappingstep,directlyoptimiz-ingtheweightofthegenreandcollectionfeaturesbycomputingaseparatewordtranslationtableforeachfeature,estimatedfromonlythosesentencesthatcomprisethatgenreorcollection.Thecommonthreadthroughoutpriorworkistheconceptofadomain.Adomainistypicallyahardconstraintthatisexternallyimposedandhandla-beled,suchasgenreorcorpuscollection.Forex-ample,asentenceeithercomesfromnewswire,orweblog,butnotboth.However,thisposessev-eralproblems.First,sinceasentencecontributesitscountsonlytothetranslationtableforthesourceitcamefrom,manywordpairswillbeunobservedforagiventable.Thissparsityrequiressmoothing.Sec-ond,wemaynotknowthe(sub)corporaourtraining

1Languagemodeladaptationisalsoprevalentbutisnotthe

focusofthiswork.

115datacomefrom;andevenifwedo,“subcorpus”maynotbethemostusefulnotionofdomainforbettertranslations.Wetakeaﬁner-grained,ﬂexible,unsupervisedap-proachforlexicalweightingbydomain.Weinduceunsuperviseddomainsfromlargecorpora,andweincorporatesoft,probabilisticdomainmembershipintoatranslationmodel.Unsupervisedmodelingofthetrainingdataproducesnaturallyoccurringsub-corpora,generalizingbeyondcorpusandgenre.De-pendingonthemodelusedtoselectsubcorpora,wecanbiasourtranslationtowardanyarbitrarydistinc-tion.Thisreducestheproblemtoidentifyingwhatautomaticallydeﬁnedsubsetsofthetrainingcorpusmaybebeneﬁcialfortranslation.Inthiswork,weconsidertheunderlyinglatenttopicsofthedocuments(Bleietal.,2003).TopicmodelinghasreceivedsomeuseinSMT,forin-stanceBilingualLSAadaptation(Tametal.,2007),andtheBiTAMmodel(ZhaoandXing,2006),whichusesabilingualtopicmodelforlearningalignment.Inourcase,bybuildingatopicdistri-butionforthesourcesideofthetrainingdata,weabstractthenotionofdomaintoincludeautomati-callyderivedsubcorporawithprobabilisticmember-ship.Thistopicmodelinfersthetopicdistributionofatestsetandbiasessentencetranslationstoap-propriatetopics.Weaccomplishthisbyintroduc-ingtopicdependentlexicalprobabilitiesdirectlyasfeaturesinthetranslationmodel,andinterpolatingthemlog-linearlywithourotherfeatures,thusallow-ingustodiscriminativelyoptimizetheirweightsonanarbitraryobjectivefunction.Incorporatingthesefeaturesintoourhierarchicalphrase-basedtransla-tionsystemsigniﬁcantlyimprovedtranslationper-formance,byupto1BLEUand3TERoverastrongChinesetoEnglishbaseline.2ModelDescriptionLexicalWeightingLexicalweightingfeatureses-timatethequalityofaphrasepairbycombiningthelexicaltranslationprobabilitiesofthewordsinthephrase2(Koehnetal.,2003).Lexicalcondi-tionalprobabilitiesp(e|f)areobtainedwithmaxi-mumlikelihoodestimatesfromrelativefrequencies2Forhierarchicalsystems,thesecorrespondtotranslationrules.c(f,e)/󰀂ec(f,e).Phrasepairprobabilitiesp(e|f)arecomputedfromtheseasdescribedinKoehnetal.(2003).Chiangetal.(2011)showedthatisitbeneﬁ-cialtoconditionthelexicalweightingfeaturesonprovenancebyassigningeachsentencepairasetoffeatures,fs(e|f),oneforeachdomains,whichcomputeanewwordtranslationtableps(e|f)esti-matedfromonlythosesentenceswhichbelongtos:cs(f,e)/󰀂ecs(f,e),wherecs(·)isthenumberofoccurrencesofthewordpairins.