Task description
Page Contents
- The objective of the task is to translate tweets between languages of the Iberian Peninsula: Catalan (ca), Basque (eu), Spanish (es), Galician (gl), and Portuguese (pt). Translations for the following pairs will be evaluated: eu-es, ca-es, gl-es and pt-es.
- We will provide a development corpus (eu-es, ca-es), as well as test corpora for all language pairs.
- Corpora for eu-es and ca-es language pairs have been harvested from bilingual Twitter accounts that simultaneously post tweets in two languages. These tweets have been manually aligned and double-checked. Owing to the collection methodology, most of the tweets are formally written.
- Corpora for the other two pairs (gl-es and pt-es) have been generated by manually translating the above corpora.
- We will provide the Twitter account IDs both for the development and test corpora.
- The participants of the task will make use of the development corpora (aligned corpora) to develop and tweak their systems. Additionally, the participants can also make use of complementary resources, including Twitter’s free API.prime time replica watches review
- Finally, the participants will have to produce machine translations of the tweets provided in the test corpus, which will only be provided in one language.
- We will also ask participants to assist with the manual evaluation of the translations submitted by others, provided that one of the team members is proficient in at least one of the languages.fake watches
Evaluation
- The task consists in translating each tweet to the target language.
- The translated tweets cannot exceed the maximum length of 140 characters. If the length is exceeded, tweets will be truncated to 140 characters.
- Hashtags will also be considered a translatable part of the tweets, while URLs and user mentions will be annotated as: URLURL, IDIDID.
- For each language pair, the evaluation will be performed in both directions (so that there is no need to distribute the translations).
- The performance of the systems will be assessed with well-known, automatic evaluation measures: Bleu, NIST, TER and METEOR, compared against a single reference.
- Thanks to the manual evaluation work collaboratively performed by participants, we will also provide a qualitative analysis of the systems.replica watch omega