Generating Story Reviews Using Phrases

Hiroshi Ota and Kazuhide Yamamoto. Generating Story Reviews Using Phrases Expressing Emotion. Proceedings of the Annual meetings of the Pacific Asia Conference on Language, Information and Computation (PACLIC 22), pp.302-310 (2008.11)

Transcript

1. 1 Generating Story Review Using Phrases Expressing Emotion 　　　　　Nagaoka University of Technology 　　　Hiroshi OTA, Kazuhide YAMAMOTO
2. 2 What we have done Construct the story review generator input: miss the medal output: The scene where he miss the medal is agonizing. Task: Generating text, Subjective expression
3. 3 Review Flow [human] Oh…sad ① Read Feel Something Write the Review ② ③ The scene where … is sad.
4. 4 Review Flow [system] ??? ① Input ② Judge the emotion Generate the Review ③ ???
5. 5 Preparing the Lexicon • 2 types of the expression about emotion – directly : happy,sad,scare, ... – Indirectly: we can eat the special dinner • The story suggest the emotion – Indirectly • System needs the Indirectly Lexicon – Emotion Emerged Expression (E3)
6. 6 E3 Lexicons joy sad fear のどを　潤す Wet one's whistle 川に　落ちる Falls in the drink 事件が　おきる happen accident 旅行に　行く Go trip 職場を　去る Go into retirement 泥沼に　はまる Fall into slough 結婚式を　あげる Celebrate a marriage 腰を　やる Suffer from backache 指を　切る Cut one's finger
7. 7 Construction of E3 Lexicons Blog 悲 Sad・E3 ①Extract Emotion Blogs ②Add to Lexicons 怖 喜 Sad Joy Fear 100 10 Deviation score for each phrases [Fujimura et al. 04]
8. 8 Evaluation:E3 joy sad fear 327,702 37,439 13,238 Right:23%, Wrong: 7% 70% lexicons are judged the No-emotion
9. 9 Review Flow [system] Sad ① Input ②Judge the emotion Generating Review ③ ??? E3 Lexicons
10. 10 Review sentence  Includes Object and Attitude  Emotion expression model [Nakayama et al.05] The scene where he holds back his tears is agonizing. E3 Emotion Expression Object／Scene Attitude／emotion
11. 11 Generating Review  Use existing review as the template Existing review: The scene where he holds back his tears is too agonizing Miss the medal Generate Book(B)’s review Replace a scene of Book(A) a scene of Book(B)
12. 12 Review Generator Template Review system ②Input: ⑤output:　　　　　 ①Extract Review Sentence as Template ④Select Template “miss the medal” The scene where he miss the medal is agonizing E3 ③Judge the emotion
13. 13  The object is clear  Include the emotion expression  Specific Noun + Particle + Emotion Expression Review Template The scene of the dialogue on the bus is pleasant. バスでの掛け合いの場面が面白い
14. 14 How to select the template? • What is the good choice • Input emotion and Template emotion should be the same? • Coherence of the emotion in one sentence and the naturalness of sentence She receive a present that is sad.
15. 15 Validation: Coherence and Naturalness Prepare the 2 pattern of sentences. Coherence and Incoherence about emotion  Generate 3 reviews against an E3.  An E3 has only one emotion.  3 reviews are made by 3 templates, each templates has different emotion expression.
16. 16 Judge Natural / Unnatural Coherence sentence： 30/30 Incoherence sentence： 11/60 (Input from E3 lexicon) The scene that *** is joyous. The scene that *** is sad.　　　 The scene that *** is fear. Input: falls in the drink (sad) (coherent) (incoherent) (incoherent) (incoherent) Natural sentence / Evaluated Number
17. 17 Discussion  Incoherent but natural sentence  Object is only two segment  Imagine some situation   Adding the Information of object, particular emotion may occur  Scene that falls in the drink  Who？When？Where？
18. 18 Conclusion  Generate the sentence which include the subjective expression  Construct E3 Lexicon from blog  Propose the method generating Review sentence  Extract the review Flame in automatically  Emotion coherence and naturalness
20. 20  Generate text which include Subjective expression  Opinion・Evaluate・Emotion  Review
21. 21 Construct flow of E3 Blog Sad SVM e Sad e other Sad・E3 ①Extract emotion blog ④Construct Lexicon ③Emotion Classification ②Construct classifier fear Joy
22. 22 Extract Emotion blog  Sad blog requirements  1) Title include “sad”  2) “sad” is majority emotion expression in the text Table1.extract result Requirement (1) (2) (1)&(2) Accuracy （sad/extract） 0.5 (10/20) 0.6 (12/20) 0.85 (17/20)
23. 23 Division Model  SVM  feature： content word  Training data  correct example： “Sad” blog  Wrong example： “Joy” blog, “fear” blog Table2.result of 10-fold cross validation　　　　 Classify Model joy sad fear Acc. 70.9 71.1 71.1
24. 24 Evaluation about E3 joy sad fear correct 19 8 7 wrong 0 2 3 no-emotion 20 29 21 Can't evaluate 7 6 9 原因を→調べる Research the cause 圧力を→思う Think pressure
25. 25 Scoring [Fujimura04]  piece: syntactic piece  Pe(piece)：probability occurring at e  scoree(piece)：scoring about emotion-e
26. 26 Unit of E3 Lexicons  Syntactic Piece  A modifier and a modificand [Aoki et al. 07]  Structure information
27. 27 Review System hold back his tear oh、agnoizing.. ①Input Judge Emotion Generate review ② ③ E3 lexicons The scene where he holds his tears is agonizing.