Comments
Description
Transcript
ACL 2016 参加報告
ACL 2016 参加報告 ~ 会議概要と研究傾向分析 ~ 笹野遼平(東工大) Contents • ACL 2016 について • 投稿/査読等に関する詳細 • 採択論文の傾向 2016/09/09 ACL2016参加報告(笹野) 2 ACLとは • The annual meeting of the Association for Computational Linguistics (ACL) – 計算言語学(CL)・自然言語処理(NLP)分野のトップ会議 (個人的印象: [ ACL ≥ NAACL = EMNLP ] > COLING = EACL > IJCNLP > …) • 1963年より毎年開催で今年は54回目 – 1968年までの名称はThe Association for Machine Translation and Computational Linguistics (AMTCL) – 開催時期は例年6月~8月 • 開催地は、欧州→北米→アジア/オセアニアで巡回 – Sofia’13 → Baltimore’14 → Beijing’15 → – Berlin’16 → Vancouver’17 → Melbourne’18 → … 2016/09/09 ACL2016参加報告(笹野) 3 • 開催期間 ACL 2016の概要 – August 7-12, 2016 (本会議: August 8-10) • 開催場所 – 口頭発表: Humboldt University, Berlin – ポスター: Maritim proArte Hotel, Berlin • 規模 (=過去最大) – 参加者数: 1500人以上 – 発表論文: 354 (long:231 + short:97 + TACL:25) 2016/09/09 ACL2016参加報告(笹野) 4 Berlin, Germany 東京 2016/09/09 ACL2016参加報告(笹野) 5 Humboldt University @ Berlin Berlin Tegel Airport 2016/09/09 Humboldt University ACL2016参加報告(笹野) 6 Humboldt University 2016/09/09 ACL2016参加報告(笹野) 7 本会議1日目のスケジュール 8:45-9:00 9:00-10:10 10:10-10:40 Opening Session (Audimax + Kinosaal) Invited talk I: Amber Boydstun (Audimax + Kinosaal) Coffee Break Session 1A: Session 1B: Session 1C: Session 1D: Session 1F: Session 1E: Semantic Information Machine Word Noncomposi 10:40-12:00 Parsing I Parsing I Extraction Translation I Meaning I tionality 2094 Audimax Kinosaal 2002 3038 2091 12:00-13:40 Lunch Break Session 2A: Session 2B: Session 2C: Session 2E: Session 2F: Session 2D: Word Events and Sentiment Information Phonoly and 13:40-15:00 Parsing II 7並列! Vectors I Schemas Analysis Retrieval Morphology 3038 Audimax Kinosaal 2002 2094 2091 15:00-15:30 Coffee Break Session 3A: Session 3B: Session 3F: Session 3C: Session 3D: Session 3E: Session 3G: Question Sentence Entities and 15:30-17:10 Parsing III Dialog Generation Topics Answering I Vectors Coreference 2002 3038 2094 2097 Audimax Kinosaal 2091 18:00-21:00 Poster Session I, Demo Session A, Student Research Workshop, and Dinner (Maritim Hotel) 3時間! 2016/09/09 ACL2016参加報告(笹野) 8 本会議2日目のスケジュール 9:00-10:10 10:10-10:40 Invited talk II: Mark Steedman (Audimax + Kinosaal) Coffee Break Session 4A: Session 4B: Session 4C: Session 4D: Session 4F: Relations and Session 4E: Semantic Word Tasks and Document 10:40-12:00 Knowledge Parsing IV Parsing II Meaning II Datasets Analysis Bases 2094 Kinosaal 2002 3038 2091 Audimax 12:00-13:40 Lunch Break Session 5B: Session 5A: Session 5C: Session 5D: Semantics Session 5E: Deep Machine Text and Potpourri I 13:40-15:00 Learning Translation II Classification Generation (short) (short) (short) (short) (short) 2094 Audimax 2002 3038 Kinosaal 15:00-15:30 Coffee Break Session 6A: Session 6B: Session 6C: Session 6E: Session 6F: Session 6G: Session 6D: Machine Word Vectors Machine Language Summarizati Learner 15:30-17:10 Discourse Learning II Translation III and Vision on Language 3038 Audimax Kinosaal 2002 2094 2091 2097 17:30-19:00 Poster Session II, Demo Session B, and Dinner (Maritim Hotel) 19:00-23:00 Boat Cruise 2016/09/09 1.5時間 ACL2016参加報告(笹野) 3時間! 9 本会議3日目のスケジュール 9:00-09:40 9:40-10:10 10:10-11:50 11:50-13:30 ACL President's Talk (Audimax + Kinosaal) Session 7A: Outstanding Papers I Audimax 15:30-17:00 17:00-19:00 2016/09/09 Outstanding Paper Session 5本×2=10本(他に1日目に1本) Lunch Break ACL Business Meeting (open to all) Audimax 13:30-15:00 15:00-15:30 Session 7B: Outstanding Papers II Kinosaal Coffee Break Session 8A: Question Answering II (short) Audimax Coffee Break Session 8B: Word Vectors III (short) Kinosaal Session 8C: Topics and Discourse (short) 2002 Session 8D: Syntax and Morphology (short) 3038 Session 8E: Potpourri II (short) 2094 Awards and Closing Session (Audimax + Kinosaal) ACL2016参加報告(笹野) 10 Contents • ACL 2016 について • 投稿/査読等に関する詳細 • 採択論文の傾向 2016/09/09 ACL2016参加報告(笹野) 11 Innovations (Katrin Erk and Noah Smith) https://www.aclweb.org/adminwiki/index.php?title=2016Q3_Reports:_Program_Chairs 1. Best Paper以外にOutstanding Paperを選出 2. Longの発表を20分から15分に短縮(+質疑5分) 3. 投稿時に任意の長さのAppendixが追加可能に 4. Area Chairによるメタレビューを導入 5. Short Paperの締め切りがLong Paperより先 6. 投稿時のArea(分野)にOtherを新設 2016/09/09 ACL2016参加報告(笹野) 12 論文投稿関連のスケジュール 日付 February 29, 2016 イベント Deadline for short paper submission (Mar 2, 2016 Notification of NAACL 2016) March 18, 2016 Deadline for long paper submission April 15, 2016 Notification of acceptance for short paper May 24, 2016 Notification of acceptance for long paper August 7–11, 2016 Conference 2016/09/09 ACL2016参加報告(笹野) 13 採択率と変動(=Longのみ) • Short Paper: 19.9% (投稿数: 463, 採択数: 92) • Long Paper : 28.0% (投稿数: 825, 採択数: 231) 900 800 30% 過去最多! 25% 700 20% 600 500 15% 400 300 200 10% 過去最多! 5% 100 0 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 投稿数 2016/09/09 かなり高め 0% 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 採択率 採択数 ACL2016参加報告(笹野) 14 AREA AREAごとの分析(Short+Long) Semantics SemanticsのLongの採択率: 37.7% (43/114) ⇔ short:19.7% Information Extraction, Question Answering, and Text Mining Sentiment Analysis and Opinion Mining Document Analysis Machine Translation 180 56 31.1% 17.1% 105 12 11.4% 3.7% 投稿数 採択数 採択率 割合 170 102 94 Tagging, Chunking, Syntax, and Parsing 81 Machine Learning 68 Social Media 69 Resources and Evaluation 65 Discourse and Pragmatics 60 Other 61 Summarization その他7AREA total 2016/09/09 48 Cognitive Modeling and Psycholinguisticsの Longの採択率: 41.2% (7/17) ⇔ short:12.5% ACL2016参加報告(笹野) 185 1288 36 21 24 27 14 21 23 15 19 7 53 328 21.2% 20.6% 25.5% 11.0% 6.4% 7.3% 33.3% 8.2% 30.9% 6.4% 20.3% 4.3% 35.4% 7.0% 31.7% 5.8% 24.6% 14.6% 28.7% 4.6% 2.1% 16.1% 25.5% 100.0% 15 Contents • ACL 2016 について • 投稿/査読等に関する詳細 • 採択論文の傾向 2016/09/09 ACL2016参加報告(笹野) 16 タイトルに頻出する内容語 10: language model 7: tppic model 10: word embedding 6: word segmentation 5: word representation miningをタイトルに 含む論文は4本、 text miningは0本 2016/09/09 ACL 2016 順位 内容語 1learning 2model 3neural 4word 5language 6network 7embedding 8translation 9machine 10semantic 11representation 12parsing 13text 14topic 15question 16relation 17detection 18classification 19dependency 20deep 割合 15.2% 14.6% 14.0% 12.2% 9.5% 9.2% 8.2% 7.9% 7.3% 7.0% 7.0% 6.7% 5.8% 4.6% 4.3% 4.0% 4.0% 4.0% 3.4% 3.4% ACL2016参加報告(笹野) 5: representation learning 21: neural network 11: neural machine 6: recurrent neural 5: deep neural (?) machine translation 11: (?) = neural 4: (?) = statistical 17 NAACL/EMNLPとの比較 NAACL 2016 順位 内容語 1neural 2learning 3network 4translation 5word 6embedding 7model 8machine 9language 10feature 11classification 12recurrent 13semantic 14sentence 15extraction 16text 17representation 18parsing 19multilingual 20unsupervised 2016/09/09 割合 20.4% 13.8% 13.3% 11.6% 11.1% 10.5% 9.9% 8.8% 8.3% 8.3% 7.2% 6.6% 5.5% 5.0% 5.0% 4.4% 4.4% 4.4% 4.4% 3.9% ACL 2016 順位 内容語 1learning 2model 3neural 4word 5language 6network 7embedding 8translation 9machine 10semantic 11representation 12parsing 13text 14topic 15question 16relation 17detection 18classification 19dependency 20deep EMNLP 2016 割合 15.2% 14.6% 14.0% 12.2% 9.5% 9.2% 8.2% 7.9% 7.3% 7.0% 7.0% 6.7% 5.8% 4.6% 4.3% 4.0% 4.0% 4.0% 3.4% 3.4% ACL2016参加報告(笹野) 順位 内容語 1neural 2learning 3model 4network 5language 6semantic 7parsing 8embedding 9translation 10word 11text 12machine 13question 14sentence 15generation 16sentiment 17extraction 18classification 19relation 20evaluation 割合 21.1% 12.8% 12.4% 10.5% 9.4% 8.7% 7.5% 7.5% 7.1% 6.8% 6.8% 6.4% 5.3% 4.9% 4.9% 4.5% 4.5% 4.5% 4.1% 3.8% 18 Outstanding Papers • Long: 9本 (825の1.1%) 1. 2. 3. 4. 5. 6. 7. 8. 9. • Danqi Chen, Jason Bolton and Christopher D. Manning: A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task Sida I. Wang, Percy Liang and Christopher D. Manning: Learning Language Games through Interaction E.Dario Gutierrez, Roger Levy and Benjamin Bergen: Finding Non-Arbitrary Form-Meaning Systematicity Using String-Metric Learning for Kernel Regression Best Paper Vered Shwartz, Yoav Goldberg and Ido Dagan: Improving Hypernymy Detection with an Integrated Pathbased and Distributional Method Julian Hitschler, Shigehiko Schamoni and Stefan Riezler: Multimodal Pivots for Image Caption Translation Zhiting Hu, Xuezhe Ma, Zhengzhong Liu, Eduard Hovy and Eric Xing: Harnessing Deep Neural Networks with Logic Rules Fermin Moscoso del Prado Martin and Christian Brendel: Case and Cause in Icelandic: Reconstructing Causal Networks of Cascaded Language Changes Pei-Hao Su, Milica Gasic, Nikola Mrkšić, Lina M. Rojas Barahona, Stefan Ultes, David Vandyke, TsungHsien Wen and Steve Young: On-line Active Reward Learning for Policy Optimisation in Spoken Dialogue Systems Best Student Paper Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev, Slav Petrov and Michael Collins: Globally Normalized Transition-Based Neural Networks Short 2本 (463の0.4%) 1. 2. 2016/09/09 Daniël de Kok and Erhard Hinrichs: Transition-based dependency parsing with topological fields Kim Anh Nguyen, Sabine Schulte im Walde and Ngoc Thang Vu: Integrating Distributional Lexical Contrast into Word Embeddings for Antonym-Synonym Distinction ACL2016参加報告(笹野) 19 1. A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task • Danqi Chen, Jason Bolton and Christopher D. Manning: – 最高精度を実現しているが論文の主目的は 問題の分析: 共参照が絡む問題がまだ難しい 2016/09/09 ACL2016参加報告(笹野) 20 2. Learning Language Games through Interaction • Sida I. Wang, Percy Liang and Christopher D. Manning: 2016/09/09 ACL2016参加報告(笹野) 21 3. Finding Non-Arbitrary Form-Meaning Systematicity Using String-Metric Learning for Kernel Regression • 今回のBest Paper • 著者: E.Dario Gutierrez, Roger Levy and Benjamin Bergen cf. Literal and Metaphorical Senses in Compositional Distributional Semantic Models • トピック: 言語の恣意性(の否定的な立場の検証) – 犬を"dog"と呼ぶことには必然性はない – しかし、まったく言葉の形状と意味に関係がないわけではない • 手法: String-Metric Learning for Kenel Regression – 意味的類似度(by SkipGram)を再現するようにWeighted Edit Distance を学習し検証 • 結論: 語の形状と意味間には従来示されているよりも強い関係 2016/09/09 ACL2016参加報告(笹野) 22 4. Improving Hypernymy Detection with an Integrated Path-based and Distributional Method • Vered Shwartz, Yoav Goldberg and Ido Dagan: – 上位下位関係の検出: Actor => Tom Cruise – Path-basedの手法をRNNで高性能化 – Path-basedな手法をDistributionalな手法を 組み合わせ現状の最高精度を大幅に更新 2016/09/09 ACL2016参加報告(笹野) 23 5. Multimodal Pivots for Image Caption Translation • Julian Hitschler, Shigehiko Schamoni and Stefan Riezler: – 画像のキャプションの翻訳に おいて、目的言語のキャプ ションが付与された類似画像 を利用することで精度向上 2016/09/09 ACL2016参加報告(笹野) 24 9. Globally Normalized Transition-Based Neural Networks • Daniel Andor, Chris Alberti, David Weiss, Aliaksei Severyn, Alessandro Presta, Kuzman Ganchev, Slav Petrov and Michael Collins: – 今年5月にGoogleが公開した高性能な依存構造 解析器SyntaxNet関連の論文 – Transition-basedのニューラルネットで依存構造解析 だけでなく、品詞タグ付け、文圧縮でも 最高精度 2016/09/09 ACL2016参加報告(笹野) 25 まとめ • ACL2016における新たな試み 1. 2. 3. 4. 5. 6. Best Paper以外にOutstanding Paperを選出 Longの発表を20分から15分に短縮(+質疑5分) 投稿時に任意の長さのAppendixが追加可能に Area Chairによるメタレビューを導入 Short Paperの締め切りがLong Paperより前に 投稿時のArea(分野)にOtherを新設 • 過去最大のACL – 参加者数: 1500人以上 – 発表論文: 354本 (long + short + TACL) – 特に女性の登壇者が多かった印象 2016/09/09 ACL2016参加報告(笹野) 26