STOA Workshop State of the art Machine Translation - Current challenges and future opportunities 3 December Report

Transcription

1 STOA Workshop State of the art Machine Translation - Current challenges and future opportunities 3 December 2013 Report Jan van der Meer MT as the New Lingua Franca In this age of constant development of communication technologies there are endless zettabytes of information waiting to be translated into over 2,000 languages in this age of hyper-globalisation. We are moving into all new types of content, dynamic content with social media that we need to worry about Linguistic diversity is the new reality Language is one of the main barriers to accessing the global economy. Despite the growth of the sector is the translation industry ready to deal with this growing necessity, and offer the required services? Definitely not, we only have maybe 200,000 in the world and 10,000-15,000 translation agencies, and they are just scratching the surface. There is so much more work to be done. The Translation Technology Landscape Report, carried out by TAUS Segregating the past and future evolution of translation over the five decades Translation through documents, Localisation through software, Translation memory technology Globalisation through simship, Started dealing with the web Integration through integration in enterprise systems, Translation is becoming a strategic function and it s now that we start looking at MT as another resource, as another support Convergence through translation being embedded in every app, on every screen, Translation becoming a utility. Embedded in everything we do. Translation technology converges with search, speech and knowledge management. Economic, technological and social trends are leading a paradigm shift where translation is moving from being a luxury to being a utility. Evolving data, technology, interoperability and metrics are making MT the New Lingua Franca. Resulting in trends such as cloud, crowd, data, mobile, social and convergence pushing the translation sector ahead through new challenges and opportunities. 1

2 Georg Rehm Europe s Language in the Digital Age: Multilingual Technologies for Overcoming Language Barriers and Preventing Digital Language Extinction. One of the most poignant challenges in the modern world of languages is to ensure that a person s mother tongue does not become a disadvantage, through the use of advanced technologies. The necessary research has made considerable progress but not at a fast enough pace to be able to deal with the growing issue within the next years. A major collaborative push from stakeholders, researchers, industries, funding, and policy makers is required in order to meet the growing demand of a multilingual society. The Language White Paper Series compiled by META NET reports on the state of our languages in the digital age and the level of support through language technology, covering 31 languages. It looks at the societal and technological problems, challenges and the economic opportunities that are involved in language translation. The white paper showed that there are inconsistencies across Europe in its languages and technologies. With the possible danger of 21 European languages suffering digital extinction. META NETs Strategic research Agenda for Multilingual Europe 2020 calls for a translingual cloud, social intelligence and e-participation, socially-aware interactive assistants, a European service platform for language technologies and core technologies for language analysis and production. These proposals offer Europe the chance to situate itself at the top of this growing industry or a core area of economic growth. Hans Uszkoreit Attacking Quality Limitations: New Approaches to translation Technologies. Even today s technology has many more applications, especially by better combination with human translation, but greater disruptive changes are still ahead of us. MT has not yet even scratched the surface of business. The presentation covered seven main points that are required for progress in the area of MT. 1. Analytical investigation of quality barriers, Concentration on nearly good translations. Better metrics needed for the quality measurements of HT and MT. 2. Inclusion of professional translators, Their involvement, data and educated opinions are vital to create high quality translation technology. 3. Improvement of automatic quality estimation, 4. Specialised technologies for different purposes, There is too much influence on in-bound translation, the demand is for outbound translation, and this is where the sector will expand and profit. 5. Creative combination of different approaches, 2

3 Hybrid MT is vital for success 6. Exploitation of the best monolingual processing technologies, 7. Putting a strong focus on semantics, There needs to be major pushes in research and to establish large web-based web services. The European Union has the ability to support R&D through funding, service procurement and policy development. Heidi Depraetere Current and Future Applications of Machine Translation With any new technology you need to first evaluate it and see what it can bring to you before you consider using it, and MT is no different. MT is expensive and out of reach for the majority of SMEs The promise of semantic machine translation (SMT) is the main selling point for MT MT will play a bigger role in the future as volume explosion and new content types brings translation beyond human scale, meaning intelligent content needs to be multilingual. Alex Weibel Speech Machine Translation: Current Status and Long Term prospects. If we ignore the spoken version of language we cannot realise this vision of the future. If we could do translation of speech there would be numerous wonderful applications. Speech adds a few challenges to MT. First of all it adds errors. There are speech recognition front ends that feed into a MT system that add errors. There are significant sources of errors depending on your environment, e.g. the microphone issues and speaking style. Speech machine translation has been used in humanitarian deployment and has the possibility of translating on-going monologue such as speeches and lectures, and can be used on a mass scale through modern hand held communication devices e.g. JIBBIGO, a mobile application. An EU funded project called EU Bridge will hopefully develop such technologies. Automated Speech Recognition (ASR) has tremendous potential but again, speech variation and the use of unrecognisable words affect it, so high accuracy is tantamount to its success. Cross-lingual spoken communication between the people of the world is possible. I believe that it is possible to come to a world where MT becomes the glue between us, and that includes speech that we can communicate in our own spoken language, and for this obviously there is a long list of things that still need to be done, in both the scientific and organisations aspects. 3

4 Jan Hajič The Future of MT - A Paradigm Change: From Word-based Statistical Machine Translation Towards High-quality, Hybrid Systems. Current technologies are phrased-based statistical MT systems Hybrid systems in the future will consist of standard statistical MT with structural and semantic features added, creating multi architecture systems Areas of research to pursue for a qualitative leap include, linguistics, semantics, textual analysis, discourse, spoken language, and knowledge representation, machine learning in complex spaces, evaluation, neurolinguistics, cognitive science and psycholinguistics. 4

5 Machine Translation Workshop - Some points raised in closing discussion Copyright law- Audience We need something like an exemption of the copyright law. Copyright law basically has some regulations on how to protect copyright and there are exemptions for academic use. We really need something like an exemption for the use of language resources for the purposes of training, models for speech recognition and MT, some fair use concept, and something that allows people to avail of publically available resources e.g. data on the internet. Hans Uszkareit: When it comes to policies in the area of copyright protection or especially of international property, such as data and language data, things are in a flux. Right now there are several directorate generals put together, that are committed to prepare a better draft. People are extremely worried about using data for scientific purposes. Plus the European Commission tries to be very protective when it comes to the intellectual property industry. The suggestion would be, not to put focus in a very specific direction. The right legal instrument for making language data and other data resources could involve the use of semantic data, which we just heard is more freely available for research and technology development. Jaap van der Meer: One of the biggest issues is that the differences in how copyright is protected in North America compared to in Europe is creating a sort of inequality. It gives North American companies a great advantage over European companies when it comes to training MT engines. This is a critical issue, and needs to be addressed by European policy makers in order to modify and clarify copyright law. Statistical paradigms- Jan Hajic: Even if the core technologies in the MT changes I still believe that it will be statistically based. For example if we can get language analysis of the semantic properties for all of these various languages, this data would be statistical. If someone creates graphs (trees) for a couple thousand sentences, that data would also be a statistical component of the whole translation system. Data security- Hans Uszkareit: MT can only become better with the use of more language data. Learning algorithms can grow and improve on any data that is out there, without compromising the privacy or security of an individual or corporation. So unless we understand this we are stuck with these localised small data sources, which are impeding the growth of the MT. So I think this issue needs to be overcome on a political level. 5