Sentence_simplification_with_deep_reinforcement_learning.pdf

Transcript

1. Sentence Simplification
   with Deep Reinforcement
   Learning
   Proceedings of the 2017 Conference on Empirical Methods in
   Natural Language Processing, 2017 pp. 584–594
   Zhang Xingxing, Lapata Mirella
   Nagaoka University of Technology Takumi Maruyama
   

   View Slide

2. Abstract
   Ø Sentence simplification aims to make sentences
   easier to read and understand
   Ø This paper proposes encoder-decoder model
   coupled with a deep reinforcement learning
   frame work for text simplification
   Ø The proposed model outperforms competitive
   simplification systems on experiments.
   2
   

   View Slide

3. Reinforcement Learning for Sentence Simplification
   Ø This paper proposes following two models:
   • Deep Reinforcement learning sentence simplification
   model (DRESS)
   • DRESS + Lexical Simplification model (DRESS-LS)
   3
   

   View Slide

4. DRESS
   Ø The overview of deep reinforcement learning
   simplification model
   4
   

   View Slide

5. DRESS
   Ø The overview of deep reinforcement learning
   simplification model
   Encoder-Decoder model
   5
   

   View Slide

6. DRESS
   Ø The overview of deep reinforcement learning
   simplification model
   Reinforcement Learning
   6
   

   View Slide

7. DRESS
   Ø Reward
   ! ", $, %
   $ = '(!( ", $, %
   $
   + '*!* ", %
   $
   + '+!+ %
   $
   ,-, ,., ,/ ∈ [0, 1]
   where
   5-: The simplicity reward
   5.: The relevance reward
   5/: The fluency reward
   7
   

   View Slide

8. DRESS
   Ø Reward
   • Simplicity:
   • Relevance:
   • Fluency:
   !" = $%&'( ), +
   ,, ,
   + . − $ %&'( ), ,, +
   ,
   !0 = 123 4)
   , 4+
   ,
   =
   4)
   5 4+
   ,
   4)
   4+
   ,
   !6 = 789
   .
   +
   ,
   :
   ;<.
   +
   ,
   =2>?@A
   B
   C;
   |B
   CE:;G.
   4)
   and 4+
   ,
   are sentence vectors
   8
   

   View Slide

9. DRESS-LS
   Ø Lexical simplification is a task that replaces
   complex words with simpler alternatives
   Ø This paper uses pre-trained encoder-decoder
   model for lexical simplification
   Ø ! "#
   "$:#&$
   , ( = 1 − , -./
   "#
   "$:#&$
   , ( + ,-/1
   "#
   (, 2#
   Where , ∈ [0,1]
   9
   

   View Slide

10. Experimental Setup
    Ø Three simplification datasets
    • WikiSmall (Zhu et al. 2010)
    • WikiLarge (Kauchak 2013, Woodsend and Lapata 2011, Zhu et al. 2010)
    • Newsela (Xu et al. 2015)
    Dataset Train Dev. Test
    WikiSmall 89,042 205 100
    WikiLarge 296,402 2,000 359
    Newsela 94,208 1,129 1,076
    10
    

    View Slide

11. Experimental Setup
    Ø Comparison systems
    • PBMT-R:
    • Hybrid:
    A hybrid semantic-based model that combine
    simplification model and monolingual machine
    translation model
    • SBMT-SARI:
    A syntax-based translation model trained with PPDB and
    tuned with SARI
    A monolingual phrase base machine translation
    with a reranking post-processing step
    11
    

    View Slide

12. Results
    (Five point Likert scale) 12
    

    View Slide

13. Results
    (Five point Likert scale) 13
    

    View Slide

14. Results
    (Five point Likert scale) 14
    

    View Slide

15. Results
    15
    

    View Slide

16. Results
    16
    

    View Slide