IPSJ SIG Technical Report Vol.2018-NL-237 No /9/26 Wikipedia 1,a) 2 3 1,4 Wikipedia 7 6,701 Wikipedia Wikipedia Hanawa Kazuaki 1,a) Sasaki Akira

Size: px

Start display at page:

Download "IPSJ SIG Technical Report Vol.2018-NL-237 No /9/26 Wikipedia 1,a) 2 3 1,4 Wikipedia 7 6,701 Wikipedia Wikipedia Hanawa Kazuaki 1,a) Sasaki Akira"

りさこつちた
2 years ago
Views:

1 Wikipedia 1,a) 2 3 1,4 Wikipedia 7 6,701 Wikipedia Wikipedia Hanawa Kazuaki 1,a) Sasaki Akira 2 Okazaki Naoaki 3 Inui Kentaro 1,4 1. [1]. [2], [3], [4] [5], [6], [7] [8], [9] [10] [11], [12], [13] (TPP) TPP TPP (TPP ) TPP AIP a) hanawa@ecei.tohoku.ac.jp Sasaki [11] Boltuzic and Snajder [12] microstructures Bar-Haim [13] ( ) () 1): ( i. Wikipedia ii. (); ( 1 ) Wikipedia *1 7 6,701 (3.1 ). ( 2 ) Wikipedia *1 OpenResources 1

2 ターゲット文プレミアムフライデーに対する賛否そんなすぐに働き方を変えることはできない. プレミアムフライデーが促進する否定的消費活動をもっと増やしたほうがいい. プレミアムフライデーが促進する肯定的反対賛成外部知識を利用した賛否分類 (4 章 ) 大阪都構想集団的自衛権プレミアムフライデープレミアムフライデーは 2017 年 ( 平成 29 年 ) に日本国政府が提唱した個人消費喚起キャンペーンである 15 時に仕事を終えることを奨励する働き方改革と連携し知識源としての Wikipedia 記事関係抽出 (3 章 ) 促進する促進される抑制する抑制される ( 原発, 抑制, 環境 ) ( 大阪都構想, 促進, 大阪都 ) ( プレミアムフライデー, 促進される, 日本国政府 ) ( プレミアムフライデー, 促進, 個人消費 ) ( プレミアムフライデー, 促進, 15 時に仕事が終わる ) ( プレミアムフライデー, 促進, 働き方改革 ) トピックに関する促進 / 抑制関係知識 1 Wikipedia *1 (3 ) ( 3 ) (4 ) ( 4 ) F 3 (5 ) Twitter [14] [15] SNS Twitter Wikipedia Wikipedia Wikipedia Wikipedia TF*IDF TF DF Wikipedia TPP ,000 1,795 1 TPP 7. 2, ,

3 % () 56.3 (: ) / (Wikipedia ) 26.3 (: TPP) / (Wikipedia ) 13.9 (: TPP) 2.5 (topic: ) 2 / 491 (10%) / [16], [17], [18] *2 A B A B A B A B % ( ) 2 X X = 40.2% (= 26.3% %) /2,3 TPP TPP TPP Wikipedia / 26.3% Wikipedia / 3. Wikipedia / 3.1 Wikipedia / / Wikipedia / Hanawa [18] Wikipedia / *2 // [13] Boltuzic and Snajder [12] / 8 Boltuzic and Snajder 8 /A B A B 4 : [title] B Sup: [title] B By: A [title] SupBy: A [title] A B [title] Hanawa [18] 1,494 Wikipedia / 7 Wikipedia /, Sup, By, SupBy Yahoo! * Wikipedia /, Sup, By, SupBy 4 Wikipedia IOB2 (B-, I-, B-Sup, I-Sup ) BiLSTM-CRF[19] 300 Wikipedia *4 Hanawa [18] *3 *4 3

4 TPP Sup By SupBy IOB2 α α O Wikipedia t rwikipedia m (KB)D (t, r, m) D. (1) KB D Wikipedia 4. 3 / z N s = w 1, w 2,, w N 3 y R 3 y s 4.1 (KB ) s BiLSTM h w t x t R dw ELMo [20] ELMo t R de d w, d e ELMo w t x t = x t ELMo t 2 BiLSTM x 1,, x N h 1,, h N h 1,, h N h t, c t = LSTM(x t, h t 1, c t 1 ), (2) h t, c t = LSTM(x t, h t+1, c t+1 ). (3) h t, c t R d h (t = 1,, N) LSTM( LSTM) h t, c t R d h (t = 1,, N) LSTM( LSTM) d h h t, c t, h t, c t [ h N ; h 1 ] h t h t h h R 2d h [ h t ; h t ] (t = 1,, N) h 3 y = softmax(w h + b). (4) W R 3 2d h b R KB KB D z s ( 2) s D m m r 5 D t z D z = {(m, r) (t, r, m) D t = z} (5) p t w t D z z p t, Sup, By, SupBy, None None w t D z. TPP z = TPP, s = (,,,, ) D z = {(, )} p 1, p 2, p 3 None p 4, p 5 w t z x t w t x t x t = x t emb(p t ). (6) emb(p) : p R dr p d r x t 2 3 x t 4

5 None Non e None None None None Sup Sup By SupBy Sup Sup Sup By SupBy 2 / 4.3 KB KB D s KB 3 t LSTM w t m i KB i 5 z D z m i r i D z i (m i, r i ) (i {1,..., D z })D z m r key-value (m) 1 BiLSTM (, ) (, ) BiLSTM v i R d h v i R d h D z m i w t h t h t v i v i LSTM a t,0, a t,1,, a t, Dz a t,0 w t D z a t,i (i {1,..., D z }) w t D z i a t,i h t v i exp(sim( h t, v i )) a t,i = Dz i =0 exp(sim( h t, v i )) sim (7) sim( h t, h t v i (if 0 <i) v i )=, (8) κ (if i =0) i <0 i =0 κ() w t D z κ a t,0 r i i {1,..., D z } i (m i, r i ) D z r i (i =0)None w t D z q t = a t,i emb(r i ). (9) i=0 q t 2 LSTM [ h t ; h t ] [ h t ; h t ; q t ; q t ] (/) Wikipedia d w = d e = d h = 300, 5

6 None Sup Sup By SupBy None Sup Sup By SupBy Sup Sup By SupBy 3 / d r = 100, κ = 10, α =0.85. Wikipedia *6 ELMo Wikipedia 3 BiLM (LSTM ) Adam. Mohammad [1] F F avg =(F favor + F against )/2 10 ( ) F F avg F ( ) By z TPP D z = {(TPP, ), (TPP, By)} : ()0.476 () Wikipedia Wikipedia Wikipedia KB / (0.507) F , / Precision(p) Recall(r) (α =0.85) Precision Recall 6

7 TPP Majority baseline (F ) (a) トピック : TPP S1 = ( TPP, Sup, 食料自給力 ) 順向逆向アテンションスコア None None S1 S1 None None None None None None が最のもの Sup アテンション By スコアの合計アテンションスコアが最のもの SupBy None S1 None None None None None None None None None Sup アテンション By スコアの合計 SupBy None (b) トピック : 原発 P1 = ( 原発,, 高レベル放射性廃棄物処理費用 ) 順向逆向例文アテンションスコアが最のもの None None None None P1 P1 None アテンションスコアの合計アテンションスコアが最のものアテンションスコアの合計 Sup By SupBy None None None None None None None None Sup By SupBy None [1], [2] () [21], [22], [23] 2 (SNS [24] [25]) SNS SNS ( [10] [26] [27], [28], and [11] ) Boltuzic and Snajder [12] microstructures / 8 / Bar-Haim [13] / Web Wikipedia A vs B A and B 7

8 TPP p r p r p r p r p r p r p r Sup By SupBy / Precision (p) Recall (r) (TPP NAFTA ) (TPP ) , % / Wikipedia / / ( F ) Wikipedia SNS 2 end-to-end JST CREST(: JPMJCR1301) [1] Mohammad, S., Kiritchenko, S., Sobhani, P., Zhu, X. and Cherry, C.: Semeval-2016 task 6: Detecting stance in tweets, ceedings of the 10th International Workshop on Semantic Evaluation (SemEval), pp (2016). [2] Thomas, M., Pang, B. and Lee, L.: Get out the vote: Determining support or opposition from Congressional floor-debate transcripts, ceedings of the 2006 Conference on Empirical Methods in Natural Language cessing (EMNLP), pp (2006). [3] Murakami, A. and Raymond, R.: Support or oppose?: classifying positions in online debates from reply activities and opinion expressions, ceedings of the 23rd International Conference on Computational Linguistics (COLING), pp (2010). [4] Somasundaran, S. and Wiebe, J.: Recognizing stances in ideological on-line debates, ceedings of the NAACL HLT 2010 Workshop on Computational Approaches to Analysis and Generation of Emotion in Text, pp (2010). [5] Abu-Jbara, A., Dasigi, P., Diab, M. and Radev, D.: Subgroup Detection in Ideological Discussions, ceedings of the 50th Annual Meeting of the Association for Computational Linguistics (ACL), pp (2012). [6] Qiu, M., Yang, L. and Jiang, J.: Modeling interaction features for debate side clustering, ceedings of the 22nd ACM international conference on Conference on information and knowledge management (CIKM), pp (2013). [7] Hasan, K. S. and Ng, V.: Extra-Linguistic Constraints on Stance Recognition in Ideological Debates, ceedings of the 51st Annual Meeting of the Association for Computational Linguistics (ACL), pp (2013). [8] Kim, S.-M. and Hovy, E.: Crystal: Analyzing predictive opinions on the web, ceedings of the 2007 Joint Conference on Empirical Methods in Natural Language cessing and Computational Natural Language Learning (EMNLP-CoNLL), pp (2007). [9] Bermingham, A. and Smeaton, A.: On using Twitter to monitor political sentiment and predict election results, ceedings of the Workshop on Sentiment Analysis where AI meets Psychology (SAAIP), pp (2011). [10] Ferreira, W. and Vlachos, A.: Emergent: a novel data-set for stance classification, ceedings of the 2016 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pp (2016). [11] Sasaki, A., Mizuno, J., Okazaki, N. and Inui, K.: Stance Classification by Recognizing Related Events about Targets, ceedings of the 2017 IEEE/WIC/ACM International Conference on Web Intelligence (WI), pp (2016). [12] Boltuzic, F. and Šnajder, J.: Toward Stance Classification Based on Claim Microstructures, ceedings of the 8th Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis (WASSA), pp (2017). [13] Bar-Haim, R., Bhattacharya, I., Dinuzzo, F., Saha, A. and Slonim, N.: Stance Classification of Context- Dependent Claims, ceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics (EACL), pp (2017). [14] Gray, J., Chambers, L. and Bounegru, L.: The Data 8

9 Journalism Handbook, O Reilly Media (2012). [15] Graefe, A.: Guide to Automated Journalism, Technical report, The Tow Center for Digital Journalism (2016). [16] Hashimoto, C., Torisawa, K., De Saeger, S., Oh, J.-H. and Kazama, J.: Excitatory or inhibitory: a new semantic orientation extracts contradiction and causality from the web, ceedings of the 2012 Conference on Empirical Methods on Natural Language cessing and Computational Natural Language Learning (EMNLP- CoNLL), pp (2012). [17] Fluck, J., Madan, S., Ellendorff, T. R., Mevissen, T., Clematide, S., van der Lek, A. and Rinaldi, F.: Track 4 Overview: Extraction of Causal Network Information in Biological Expression Language (BEL), ceedings of the Fifth BioCreative Challenge Evaluation Workshop, pp (2015). [18] Hanawa, K., Sasaki, A., Okazaki, N. and Inui, K.: A Crowdsourcing Approach for Annotating Causal Relation Instances in Wikipedia, ceedings of the 31st Pacific Asia Conference on Language, Information and Computation (PACLIC) (2017). [19] Huang, Z., Xu, W. and Yu, K.: Bidirectional LSTM- CRF models for sequence tagging, arxiv preprint arxiv: (2015). [20] Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K. and Zettlemoyer, L.: Deep Contextualized Word Representations, ceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (NAACL-HLT), pp (2018). [21] Anand, P., Walker, M., Abbott, R., Tree, J. E. F., Bowmani, R. and Minor, M.: Cats Rule and Dogs Drool!: Classifying Stance in Online Debate, ceedings of the 2nd Workshop on Computational Approaches to Subjectivity and Sentiment Analysis (WASSA), pp. 1 9 (2011). [22] Zarrella, G. and Marsh, A.: MITRE at SemEval-2016 Task 6: Transfer Learning for Stance Detection, ceedings of the 10th International Workshop on Semantic Evaluation (SemEval), pp (2016). [23] Du, J., Xu, R., He, Y. and Gui, L.: Stance classification with target-specific neural attention networks, ceedings of the 26th International Joint Conferences on Artificial Intelligence (IJCAI), pp (2017). [24] Ebrahimi, J., Dou, D. and Lowd, D.: Weakly supervised tweet stance classification by relational bootstrapping, ceedings of the 2016 Conference on Empirical Methods in Natural Language cessing (EMNLP), pp (2016). [25] Wei, W., Zhang, X., Liu, X., Chen, W. and Wang, T.: pkudblab at semeval-2016 task 6: A specific convolutional neural network system for effective stance detection, ceedings of the 10th International Workshop on Semantic Evaluation (SemEval), pp (2016). [26] Jindal, N. and Liu, B.: Identifying comparative sentences in text documents, ceedings of the 29th annual international ACM SIGIR conference on Research and development in information retrieval (SIGIR), pp (2006). [27] Somasundaran, S. and Wiebe, J.: Recognizing stances in online debates, ceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language cessing of the AFNLP (ACL-IJCNLP), pp (2009). [28] Cabrio, E. and Villata, S.: A Natural Language Bipolar Argumentation Approach to Support Users in Online Debate Interactions, Argument and Computation, Vol. 4, No. 3, pp (2013). 9

_314I01BM浅谷2.indd

_314I01BM浅谷2.indd 587 ネットワークの表現学習 1 1 1 1 Deep Learning [1] Google [2] Deep Learning [3] [4] 2014 Deepwalk [5] 1 2 [6] [7] [8] 1 2 1 word2vec[9] word2vec 1 http://www.ai-gakkai.or.jp/my-bookmark_vol31-no4 588 31 4 2016