What did you mean by that?! The Customer Voice In Your Language

Size: px

Start display at page:

Download "What did you mean by that?! The Customer Voice In Your Language"

Reynard Wilkerson
5 years ago
Views:

1 What did you mean by that?! The Customer Voice In Your Language Daphne Guericke VP, Content Analytics

2 Agenda Ø Importance of understanding the use of social media globally Ø How far can machine transla8on take sen8ment analysis Ø What needs to go into technology that captures sen8ment in different languages

3 What is the percentage of North American internet users in the world? q Less than 10% q 10-15% q 15-20% q More than 20% Pop Quiz

4 Internet Users: DistribuCon by Regions

5 World Internet Users World Regions PopulaCon (2012 est.) Internet Users June 2012 PenetraCon (% PopulaCon) Users % North America 348,280, ,748, % 11.40% Oceania / Australia 35,903,569 24,270, % 1.00% Europe 820,918, ,820, % 21.50% La8n America / Caribbean 593,688, ,692, % 10.60% Middle East 223,608,203 89,890, % 3.70% Asia 3,922,066,987 1,078,568, % 44.80% Africa 1,073,380, ,447, % 7.00% WORLD TOTAL 7,017,846,922 2,407,121, % %

6 Facebook Growth % outside of NA

7 hup:// media- around- the- world by- insites- consul8ng

9 hup:// media- around- the- world by- insites- consul8ng

10 55% of social network users are connected to brands. 8 out of 10 consumers want to help brands. hup:// media- around- the- world by- insites- consul8ng

11 The pressure is on to analyze text written in languages that the analyst cannot read. Meta Brown

13 ABH Experiment Ø SenCment analysis on machine translated text Data German (997) and Japanese (588) customer reviews (Amazon) (BOSE headphones) Equal amounts of sentences from 1-2 star and 4-5 star reviews Approach German and Japanese na8ve speakers rate the sentences English na8ve speakers rate the machine- translated sentences Ra8ng Scale: Posi8ve, Nega8ve, Pos/Neg, Neutral, Unratable

14 Japanese: Machine Translated Sentences About half of the sentences rated nega8ve by Japanese annotators received a corresponding nega8ve ra8ng by English annotators.

15 Japanese: Machine Translated Sentences Only 60% of the sentences rated posi8ve by Japanese annotators got a corresponding posi8ve ra8ng by English annotators.

16 ABH Experiment: Conclusion Ø Machine- translated text is not accurate enough to reliably perform sen8ment analysis on. Ø Cultural differences are not captured in machine- translated text. Ø Recommenda)on: Data should be analyzed in the na8ve language.

17 Language Pack Development Ø Requirements for linguiscc analysis Lexicon and special word lists Part of Speech Morphological rules Syntax parsing (dependency, n- grams) Named En88es Co- reference, anaphora resolu8on

18 Language Pack Development (cont.) Ø Requirements for sencment resources Sen8ment lexicon Sen8ment modifiers and modifier rules Sen8ment excep8on rules Language- specific considera8ons Customiza8on and tuning

19 SenCment Lexicon Ø List of words with sencment value Requirement: Language- specific lists to represent the usage and sen8ment values in the target language. If basis is English seed list: Find equivalents in the target language Modify original value of sen8ment words based on usage in target language Expand list to include sen8ment words specific to target language Resources available on internet

20 Modifier Lists and Modifier Rules Ø Modifiers = Words and phrases that intensify, decrease or reverse sen8ment value (e.g. very, absolutely, not, a good deal ) Ø Modifier rules = Descrip8on of effect of modifier Example: o Word trouble = nega8ve value of - 2 o a lot of trouble = nega)ve value of - 3

21 SenCment ExcepCon Rules Ø Context- based rules that adjust original sencment value of word Examples: used to be + adjec)ve - - > reverse clean = posi)ve This restaurant used to be clean. = nega)ve would + prefer - - > nega8ve free wi- fi = posi8ve I would prefer they had free wi- fi.

22 Dutch Example Ø No discnccon between adjeccves and adverbs Example: English: "terrible" (adjec8ve, nega8ve sen8ment) "terribly (adverb, intensifying modifier) Dutch: "verschrikkelijk" o "De service is verschrikkelijk. (The service is awful.) o "Ik vind jou verschrikkelijk aardig. (I find you awfully nice, I like you a lot.)

23 Italian Example Ø SenCment word Verb: impazzire" = nega8ve (to go crazy) Ø ExcepCon rule posi8ve+da+impazzire => posi8ve "La vista è bella da impazzire" (The view is worth going crazy over. It is to die for.)

24 Italian Example Ø Modifier rule troppo+posi8ve => nega8ve Ø ExcepCon rule troppo+bello,buono,carino,bravo => posi8ve "La stanza era troppo carina" (Our room was really cute.)

25 Chinese Example Ø English rule: "too+adjeccve => negacve Chinese too 太 = "very" and "too" 他太出色了 = He is totally outstanding. Machine transla8on= "He was too good." Conclusion: Special Chinese modifier rules are needed

26 Other Language- Specific ConsideraCons Specific construc8ons You will like this not! Dutch: Star8ng a sentence with "grappig" (funny) - - > you will say something nega8ve Spelling variants We had to wait a looooong )me. Punctua8on and emo8cons (^_^) vs. :- ) (^_~) vs. ;- ) (o_o) vs. :- 0 And lots others

27 Summary: Global companies have global sencment Ø Understand Social Media around the world Ø Know where your global customers are talking Ø Be aware of lost in (machine) transla8on Ø Look or solu8on that are na8ve in language and culture Ø Consult analysts who understand the culture

28 Dank u! Vragen? Teşekkür ederiz! Sorular? תודה! שאלות? Cảm ơn bạn! Câu hỏi? Tack! Frågor? श"#$य! सव ल? Спасибо! Вопросы? ありがとうございました! 質問は? Gracias! Preguntas? Vielen Dank!!! Fragen??? 谢谢! 有问题吗? Dank je wel! Vragen? Dankie! Vrae? Grazie! Domande? 당신을감사하십시오! 질문이있으십니까? Ďakujem vám! Otázky?!" ధన%&'(! ప*శ,(? Köszönöm! Kérdései vannak? Dziękujemy! Pytania? Je vous remercie! Ques8ons?

29 Dank u! Vragen? Teşekkür ederiz! Sorular? תודה! שאלות? Cảm ơn bạn! Câu hỏi? Tack! Frågor? श"#$य! सव ल? Спасибо! Вопросы? ありがとうございました! 質問は? Gracias! Preguntas? Vielen Dank!!! Fragen??? 谢谢! 有问题吗? Dank je wel! Vragen? Dankie! Vrae? Grazie! Domande? 당신을감사하십시오! 질문이있으십니까? Ďakujem vám! Otázky?!" ధన%&'(! ప*శ,(? Köszönöm! Kérdései vannak? Dziękujemy! Pytania? Je vous remercie! Ques8ons?

20 TH OECD-JAPAN SEMINAR

OECD DIRECTORATE FOR EDUCATION AND SKILLS 20 TH OECD-JAPAN SEMINAR Roger Smyth Tokyo Centre, JAPAN 28 July, 2018 This presentation: The economic and demographic context Life-long learning: barriers to