90 likes | 355 Views
arTenTen A new, vast corpus for Arabic. Yonatan Belinkov , Nizar Habash , AdamKilgarriff , Noam Ordan , Ryan Roth, Vit Suchomel MIT/Columbia/Lexical Computing Ltd./ Univ Saarlandes/Masaryk Univ Cz. We all want corpora to be. Bigger Better More text types Richer metadata Cleaner
E N D
arTenTenA new, vast corpus for Arabic YonatanBelinkov, NizarHabash, AdamKilgarriff, NoamOrdan, Ryan Roth, VitSuchomel MIT/Columbia/Lexical Computing Ltd./ UnivSaarlandes/MasarykUnivCz
We all want corpora to be • Bigger • Better • More text types • Richer metadata • Cleaner • Better linguistic processing
Arabic • Since 2003: Arabic Gigaword • Good on most fronts except variety • Newswire only • Leeds • 2005 Arabic web corpus (oldish) • Others • Mostly • small • or not available • or newswire
arTenTen • TenTen family • See paper in main conference • Web crawled • Spiderling • Pomikalek and Suchomel, WAC 2012 • Cleaning and deduplication • justText, Onion (Pomikalek)
Size • 5.8 b space-separated tokens Fully processed: • 200M words • Tokenise, lemmatise, POS-tag by MADA, Columbia U • Sketch grammar: new work (Belinkov) Varieties/dialects • We don’t know yet
Availability • In Sketch Engine • demo
Encoding • ‘Vertical’ format • Sketch Engine input format • One word per line, tab-separated columns • Twenty-nine • Structural markup: XML
For each word word (as written, in Arabic) transdiac lemmalemma_arnon_voc_lemmanon_voc_lemma_ar stem tagbw pref3 pref3tag pref2 pref2tagpref1 pref1tag pref0 pref0tag person aspectvox modus gender number state case encliticgloss source