BLEU_is_Not_Suitable_for_the_Evaluation_of_Text_Simplification.pdf

Transcript

1. BLEU is Not Suitable for the Evaluation of Text Simplification

   Elior Sulem, Omri Abend, Ari Rappoport Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing(EMNLP2018) Nagaoka University of Technology Takumi Maruyama Literature review: 1

2. Abstract Ø Text simplification(TS) )!0 )! %  ' Ø

   BLEU TS )! *# Ø /- • $&)! "(, • $& .
   + )!0 ' 2

3. Introduction Ø 2 2)
   (! • Wiki-8REF (Xu et al. 2016)

   − .8'5 -
    − +,. 43 0  • Hsplit − /1*$
    − +, 0 Ø 6& MT-based simplification system, sentence splitting system2) (2)7, ) Ø '5 2)7 2) #"% 3

4. Gold-Standard Splitting Corpus

5. Gold-Standard Splitting Corpus Ø Wiki-8REF (Xu et al. 2016)*!$$%) Ø

   1$4annotators(Native:2, Native-like proficiency: 2)#, Ø 2+-guidlines 1. $$'/
   &"(%  2. % .
   %  5

6. Gold-Standard Splitting Corpus Ø 4 
    ( 359) • HSplit1:

   annotator=Native, guideline=1 • HSplit2: annotator=Native-like, guideline=1 • HSplit3: annotator=Native, guideline=2 • HSplit4: annotator=Native-like, guideline=2 6

7. Experiments

8. Experimental Setup (Metrics) Ø BLEU • BP: brevity penalty •!"

   : weights (usually uniform) •#" : modified n-gram precisions Ø iBLEU • $, &, ': input text, output text, reference text • (: parameter 8

9. Experimental Setup (Metrics) Ø Flesch-Kincaid Grade Level (FK) Ø SARI

   Ø Levenshtein distance (!"#$ ) 9

10. Experimental Setup (Human evaluation) Ø 703
     Ø   •

    Grammaticality (G) • Meaning preservation (M) • Simplicity (S) • Structural Simplicity (StS) Ø 5  10

11. Experimental Setup (Systems) Ø Standard Reference Setting • Systems/Corpora without

    splits − Wiki-8REF(  
    ) − Six MT-based simplification systems (NTS, SBMT-SARI etc.) • All systems/Corpora − Wiki-8REF + HSplit − Six MT-based simplification systems (NTS, SBMT-SARI etc.) Ø HSplit as Reference Setting − HSplit − Six sentence splitting systems (DSS, SEMoses etc.) 11

12. Results with Standard Reference Setting 12 Sentence-level spearman correlation (and

    p-values)

13. Results with Standard Reference Setting 13 BLEUS, StS  
    

    Sentence-level spearman correlation (and p-values)

14. Results with Standard Reference Setting 14    
    GM

      Sentence-level spearman correlation (and p-values)

15. Results with Standard Reference Setting 15  BLEU  

     BLEU-8ref
    LDspearman correlation: 0.86 BLEU-8ref
    LDspearman correlation: 0.82 BLEU-8ref
    LDspearman correlation: 0.52 BLEU-8ref
    LDspearman correlation: 0.55 Sentence-level spearman correlation (and p-values)

16. Results with Standard Reference Setting 16 Sentence-level spearman correlation (and

    p-values) SARI    
    

17. Experimental Setup (Systems) Ø Standard Reference Setting • Systems/Corpora without

    splits − Wiki-8REF(  
    ) − Six MT-based simplification systems (NTS, SBMT-SARI etc.) • All systems/Corpora − Wiki-8REF + HSplit − Six MT-based simplification systems (NTS, SBMT-SARI etc.) Ø HSplit as Reference Setting − HSplit − Six sentence splitting systems (DSS, SEMoses etc.) 17

18. Results with HSplit as Reference Setting 18 Sentence-level spearman correlation

    (and p-values) BLEUS, StS  
    

19.  Ø BLEU Text simplification   • BLEUSimplicity 

    
      • BLEUGrammaticality, Meaning preservation  Levenshitein distanceGrammaticality, Meaning preservation  Ø “BLEU should not be used for the evaluation of text simplification in general and sentence splitting” 19