A Taxonomy of Weeds: A Field Guide for Corpus Curators to Winnowing the Parallel Text Harvest Katherine M. Young Jeremy Gwinnup Lane O. Schwartz N-Space Analysis, LLC katherine.young.ctr.1@us.af.mil Air Force Research Laboratory jeremy.gwinnup.1@us.af.mil University of Illinois at Urbana-Champaign lanes@illinois.edu 31 October 2016 Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Modern MT requires corpora Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Weeds among the wheat “Weed is possibly a be�er metaphor than dirt for many of these issues with SMT training data: a wild plant growing where it is not wanted and in compe��on with cul�vated plants” – Simard (2014) Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Weeds among the wheat ἐν δὲ τῷ καθεύδειν τοὺς ἀνθρώπους ἦλθεν αὐτοῦ ὁ ἐχθρὸς καὶ ἐπέσπειρεν ζιζάνια ἀνὰ μέσον τοῦ σίτου καὶ ἀπῆλθεν. Parable of the weeds —Greek New Testament Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Some types of weeds zizania ex homine zizania ex machina • Mixed alphabet spellings • Wrong language text • Mixed morphology • Historical encoding errors • Translitera�ons of NE and • Bidirec�onal reversal Borrowings • Sentence alignment errors • Underachieving transla�on • Sentence-internal fragments • Overachieving transla�on • Harvested MT • Transla�on direc�onality Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Zizania ex machina Weeds of mechanical origin Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Wrong language text IWSLT 2014 Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Wrong language text IWSLT 2014 Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Wrong language text Common Crawl Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Wrong language text Common Crawl Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Wrong language text • U�������� I (і І) • Y�(ї Ї) • G�� W��� U����� (ґ Ґ) • I� (є Є). Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Wrong language text Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Wrong language text +0.34 BLEU on WMT’15. Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Historical encoding errors Ñïðàâêà ïî ãîðîäàì Ðîññèè è ìèðà Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Historical encoding errors Ñïðàâêà ïî ãîðîäàì Ðîññèè è ìèðà Справка по городам России и мира Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Historical encoding errors In ru-en common crawl, some Russian text is actually encoded as Windows-1251 but is incorrectly interpreted as UTF-8 Correct by shi�ing 350 hex code points into Unicode Cyrillic range Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Historical encoding errors йquipe chвteau Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Historical encoding errors йquipe chвteau équipe château Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Historical encoding errors In fr-en common crawl, some French text has the reverse problem Correct by shi�ing − 350 hex code points Unicode La�n range Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Bidirec�onal reversal Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Bidirec�onal reversal Extrac�on of RTL text from PDF • PDF is a display format • Extracted text is reversed, character-by-character No�ceable because the teh-marbuta character [ ة ] should not begin a word Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Sentence-internal fragments Last year I showed these two slides so that demonstrate that the arc�c ice cap, which for most of the last three million years has been the size of the lower 48 states, has shrunk by 40 percent. L’année dernière, je vous ai présenté ces deux diaposi�ves qui mon- traient que la calo�e glacière arc�que, qui pendant ces 3 derniers millions d’année avait la taille des Etats-Unis sans l’Alaska, qui pen- dant ces 3 derniers millions d’année avait la taille des Etats-Unis sans l’Alaska, avait diminué de 40%. Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Sentence-internal fragments +1.53 BLEU on IWSLT’14 Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Sentence-internal fragments It’s the difference between divergent thinkingand convergent thinking. You have to separate the two so that you can diverge your thoughtsand come up with this great collec�on of ideas, and then once you have this great collec�onof ideas, you focus on the convergent thinking. Chunking errors in the QED Corpus. Resolved using greedy search for two valid words. Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Harvested MT Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Zizania ex homine Weeds of human origin Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Under-achieving transla�on Errors in human transla�ons: • Translitera�on instead of transla�on Awards and Reviews → Награды и Ревью /nagradɨ i rev’ju/ • Codeswitching English: In the top ten, India comes in the last Urdu: “cer�fica�on”, “top ten”, and “number” Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Over-achieving transla�on Explicita�on in human transla�ons: • Explana�ons added • Acronyms expanded • Visual aids described in target side of TED talks • Disfluencies removed from transla�on of speech Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Mixed alphabets oнa cейчас Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Mixed alphabets Word La�n (L) or Cyrillic (C) Meaning oнa LCL she cейчас LCCCCC now MP3-плеер LLL-CCCCC MP3-player MP3плеер LLLCCCCC MP3player амазон.com CCCCCC.LLL amazon.com iпациент LCCCCCCC iPa�ent Russian word oнa is encoded such that the La�n characters o and a are used instead of the more appropriate (but visually indis�nguishable) Cyrillic equivalents. Mo�va�on Weeds of mechanical origin Weeds of human origin Conclusion AMTA 2016 Katherine M. Young, Jeremy Gwinnup, Lane O. Schwartz
Recommend
More recommend
