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 3.1 1 2 3 4 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 http://www.cl.ecei.tohoku.ac.jp/index.php? OpenResources 1
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 ) 2. 2.1 Twitter [14] [15] SNS Twitter Wikipedia Wikipedia Wikipedia Wikipedia 2.2 2015 4 2017 6 260 TF*IDF TF DF Wikipedia TPP 53 802 230 153 744 218 86 592 308 47 783 202 239 259 380 168 352 262 160 468 195 906 4,000 1,795 1 TPP 7. 2,000 7 14,000 2.3 1 5 4 1 2.4 2
% () 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 1 56.3% ( ) 2 X X = 40.2% (= 26.3% + 13.9%) /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! *3 1 10 10 2 3 3.2 Wikipedia /, Sup, By, SupBy 4 Wikipedia IOB2 (B-, I-, B-Sup, I-Sup ) BiLSTM-CRF[19] 300 Wikipedia *4 Hanawa [18] *3 http://crowdsourcing.yahoo.co.jp/ *4 https://github.com/overlast/word-vector-web-api 3
TPP 333 6 179 257 169 121 39 257 17 122 190 165 120 42 Sup 67 2 163 74 115 46 25 By 131 7 77 108 64 45 21 SupBy 145 3 86 96 51 30 6 3 3.1 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 3 4.2 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
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 ] 5. 5.1 (/) 7 5 1 1 Wikipedia d w = d e = d h = 300, 5
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 ( ) 5.2 4 F F avg F 0.478 ( ) By z TPP D z = {(TPP, ), (TPP, By)} : 0.475 ()0.476 () 2. 3.2 Wikipedia Wikipedia Wikipedia KB / (0.507) F 0.491 3.2, 3.2 5 / Precision(p) Recall(r) (α =0.85) Precision Recall 6
TPP Majority baseline 0.375 0.400 0.431 0.425 0.228 0.310 0.363 0.374 0.490 0.258 0.489 0.430 0.332 0.533 0.470 0.466 0.498 0.267 0.490 0.439 0.348 0.543 0.488 0.478 0.488 0.257 0.502 0.431 0.331 0.528 0.475 0.470 0.499 0.260 0.502 0.448 0.352 0.549 0.494 0.475 0.484 0.255 0.488 0.447 0.359 0.544 0.477 0.475 0.489 0.258 0.514 0.448 0.360 0.561 0.482 0.486 0.501 0.264 0.525 0.444 0.368 0.548 0.477 0.483 0.516 0.262 0.529 0.452 0.369 0.566 0.489 0.490 0.491 0.250 0.488 0.425 0.347 0.539 0.475 0.468 0.489 0.251 0.495 0.436 0.352 0.546 0.479 0.476 0.491 0.252 0.508 0.430 0.352 0.564 0.481 0.473 0.509 0.263 0.527 0.442 0.360 0.570 0.499 0.491 0.500 0.249 0.515 0.448 0.361 0.566 0.486 0.490 0.523 0.258 0.539 0.465 0.362 0.582 0.500 0.507 4 (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 0.02 0.01 0.00 0.00 0.01 0.02 0.04 0.02 0.03 0.01 Sup 0.00 0.29 0.82 0.76 0.43 0.14 0.01 0.03 0.00 0.01 アテンション By 0.00 0.01 0.00 0.01 0.00 0.01 0.00 0.01 0.02 0.07 スコアの合計 アテンションスコアが最 のもの SupBy 0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.01 0.00 0.00 None 0.98 0.69 0.18 0.23 0.56 0.83 0.95 0.93 0.95 0.91 S1 None None None None None None None None None 0.00 0.02 0.01 0.01 0.02 0.00 0.01 0.02 0.00 0.00 Sup 0.99 0.18 0.00 0.00 0.01 0.06 0.01 0.17 0.28 0.12 アテンション By 0.00 0.00 0.03 0.15 0.09 0.00 0.00 0.00 0.00 0.00 スコアの合計 SupBy 0.00 0.00 0.00 0.01 0.01 0.00 0.01 0.00 0.00 0.01 None 0.01 0.90 0.96 0.83 0.87 0.94 0.97 0.81 0.72 0.87 (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 0.00 0.00 0.03 0.18 0.95 0.82 0.21 Sup 0.01 0.00 0.09 0.00 0.00 0.00 0.00 By 0.00 0.01 0.01 0.00 0.00 0.01 0.02 SupBy 0.01 0.06 0.02 0.01 0.00 0.00 0.00 None 0.98 0.93 0.85 0.81 0.05 0.17 0.73 None None None None None None None 0.11 0.15 0.07 0.01 0.00 0.00 0.00 Sup 0.00 0.00 0.00 0.02 0.01 0.01 0.01 By 0.01 0.02 0.00 0.00 0.00 0.01 0.00 SupBy 0.02 0.02 0.06 0.04 0.07 0.10 0.00 None 0.86 0.81 0.87 0.93 0.82 0.88 0.99 4 6. [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
TPP p r p r p r p r p r p r p r 0.41 0.12 0.48 0.37 0.26 0.27 0.29 0.30 0.38 0.15 0.48 0.18 0.43 0.10 Sup 0.67 0.09 1.00 1.00 0.58 0.06 0.34 0.18 1.00 0.09 1.00 0.15 0.59 0.29 By 0.18 0.17 0.38 0.29 0.14 0.18 0.23 0.10 0.19 0.13 0.16 0.13 0.50 0.06 SupBy 0.32 0.08 0.0 0.0 0.29 0.15 0.19 0.05 0.25 0.18 0.30 0.08 0.85 0.10 5 / Precision (p) Recall (r) (TPP NAFTA ) (TPP ) 7. 7 6,701 40.2% / Wikipedia / / ( F 1.4 2.9 ) 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. 31 41 (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. 327 335 (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. 869 875 (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. 116 124 (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. 399 409 (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. 873 878 (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. 816 821 (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. 1056 1064 (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. 2 10 (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. 1163 1168 (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. 582 587 (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. 74 80 (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. 251 261 (2017). [14] Gray, J., Chambers, L. and Bounegru, L.: The Data 8
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. 619 630 (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. 333 346 (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:1508.01991 (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. 2227 2237 (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. 458 463 (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. 3988 3994 (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. 1012 1017 (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. 384 388 (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. 244 251 (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. 226 234 (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. 209 230 (2013). 9