WIT3: Web Inventory of Transcribed and Translated Talks

WIT3: Web Inventory of Transcribed and Translated Talks Mauro Cettolo Christian Girardi FBK – Fondazione Bruno Kessler Trento, Italy {surname}@fbk.eu Marcello Federico Abstract We describe here a Web inventory named WIT3 that offers access to a collection of transcribed and translated talks. The core of WIT3 is the TED Talks corpus, that basically redistributes the original content published by the TED Conference web- site (http://www.ted.com). Since 2007, the TED Conference, based in California, has been posting all video recordings of its talks together with subtitles in English and their translations in more than 80 lan- guages. Aside from its cultural and so- cial relevance, this content, which is pub- lished under the Creative Commons BY- NC-ND license, also represents a precious language resource for the machine transla- tion research community, thanks to its size, variety of topics, and covered languages. This effort repurposes the original content in a way which is more convenient for ma- chine translation researchers. 1 Introduction Data play a key role in machine learning as they are the main source of information to infer parameter values of the employed mathematical model. In statistical machine translation (SMT), learn- ing is performed on parallel texts, i.e. documents, sentences or even fragments of sentences with their translation(s). Large amounts of in-domain paral- lel data are usually required to properly train trans- lation and reordering models. Unfortunately, parallel data are a scarce re- source, which are freely available only for some language pairs and for few, very specific domains. c⃝ 2012 European Association for Machine Translation. For example, MultiUN (Eisele and Chen, 2010) provides large parallel texts (300 million words) but for only 6 languages; Europarl (Koehn, 2005) consists of the translation into most European lan- guages of the proceedings of the European Par- liament (at most 50 million words); JRC-Acquis1 comprises the total body of European Union law applicable to the Member States, written in 22 European languages (35 million words); other smaller parallel corpora in specific domains are included in OPUS (Tiedemann, 2009) for various languages. On the other hand, it is unfeasible for research laboratories to cover all possible needs in terms of parallel texts by resorting to professional transla- tors, given their high cost. The data available at the TED website2 is there- fore particularly valuable for the MT community. TED is a nonprofit organization that invites “the world’s most fascinating thinkers and doers [...] to give the talk of their lives”. The site makes avail- able under the Creative Commons BY-NC-ND li- cense the video recordings of the best TED talks, all subtitled in English and translated in many other languages by volunteers worldwide. The set of subtitles represents a precious multilingual par- allel corpus since its size continuously increases (more than 900 TED talks had been collected at the end of 2011), subtitles are available in a signif- icant number of languages (82 now, to be extended to 90 in the near future) and topics covered span the whole of human knowledge, making such …

Content truncated to 500 words