中科院研究员 刘康 - Event Extraction from Texts

析雅爱

2017/12/18 发布于 技术 分类

2017年12月7-9日,2017中国大数据技术大会(BDTC)在京盛大召开。12月8日,在知识图谱论坛,中科院自动化所模式识别国家重点实验室副研究员刘康做了主题为《Event Extraction from Texts》的演讲,他表示,知识图谱是人工智能和核心基础设施之一,现有知识图谱多关注于以实体为核心的静态知识,缺乏对于以事件为核心的动态知识的刻画和构建。刘康结合研究组近些年的工作,介绍了从非结构化文本中抽取事件知识的基本方法,特别介绍在开放域环境下,面对多种事件类型,在缺乏标注数据的前提下,如何自动进行数据标注,训练鲁棒的事件抽取器的有效方法,同时介绍了在金融事件抽取实践过程中的经验和体会。

文字内容
1. Event Extraction from Texts Kang Liu Institute of Automation, Chinese Academy of Sciences 2017.12.8
2. Texts to Knowledge Texts           Knowledge
3. Statistic Knowledge: Entity-Centric Knowledge Graph (Barack Obama, Spouse, Michelle Obama) Head Entity Relation Tail Entity
4. Dynamic Knowledge: Event-Centric Knowledge Graph <Type: Life, Be-Born> Obama was born on August 4, 1961 Honolulu, Hawaii. Temporal relation In 1961, moved to University of Washington for a year. Person: Obama Time: 1961 Place: Honolulu, Hawaii from 2009 to 2017. Obama, served as president of United states ..… Obama issued executive orders and… <Type: Enter-Position> Person: Obama Time: 2009-2017 Position: President of USA military drawdow n in Iraq. 出⽣生事件 • 出⽣生⽇日期 • 出⽣生地点 • 姓名 结婚事件 • 结婚⽇日期 • 结婚地点 • 男⽅方 • ⼥女女⽅方 离职事件 • 离职⽇日期 • 公司 • 职位 地震事件 • 震中 • 震级 • 震源 • 伤亡⼈人数 • 财产损失 暴暴恐事件 • 地点 • 时间 • 伤亡⼈人数 • 被攻击⽅方 • 实施⽅方 收购事件 • 收购⾦金金额 • 收购⽅方 • 被收购⽅方 • 时间 Event Frame (Script)
5. Extract Events from Texts Barry Diller on Wednesday quit as chief of Vivendi Universal Entertainment. Trigger Arguments Quit (a “Personnel/End-Position” event) Role = Person Barry Diller Role = Organization Vivendi Universal Entertainment Role = Position Chief Role = Time-within Wednesday (2003-03-04)
6. Extract Events from Texts Organization Barry Diller on Wednesday quit as chief of Vivendi Universal Entertainment. Trigger Words Arguments Words Trigger Arguments Quit (a “Personnel/End-Position” event) Role = Person Barry Diller Role = Organization Vivendi Universal Entertainment Role = Position Chief Role = Time-within Wednesday (2003-03-04)
7. Extract Events from Texts Person Organization Time Position Barry Diller on Wednesday quit as chief of Vivendi Universal Entertainment. Trigger Words Arguments Words Trigger Arguments Quit (a “Personnel/End-Position” event) Role = Person Barry Diller Role = Organization Vivendi Universal Entertainment Role = Position Chief Role = Time-within Wednesday (2003-03-04)
8. Task Definition • Definition (ACE) • An event is defined as a specific occurrence involving participants. • Event trigger, Event Type, Event argument, Argument role Barry Diller on Wednesday quit as chief of Vivendi Universal Entertainment. Trigger Arguments Quit (a “Personnel/End-Position” event) Role = Person Barry Diller Role = Organization Vivendi Universal Ente3/r27tainment 1. Event Identification (Trigger Words) 2. Event Type Identification 3. Argument Identification 4. Argument Role Identification Role = Position Role = Time-within Chief Wednesday (2003-03-04)
9. Event Extraction vs. Relation Extraction • Relation Extraction • Identify the relation between two given entities /business/company/founder Steve Jobs was the co-funder of Apple Inc. entity1 • Event Extraction entity2 • Identify the relation between an event and an entity Person Organization Time Position Barry Diller on Wednesday quit as chief of Vivendi Universal Entertainment. Trigger Words Arguments Words
10. Previous Event Extraction Task MUC Message Understanding Conference TDT Topic Detection and Tracking ACE(KBP) Automatic Content Extraction Organizer DARPA DAPRA NIST Period Content 1987-1997 1998-2004 ACE:2000-2008 KBP:2014-2017 抽取指定的事件,包括参与这些 将⽂文本切割为不不同的新闻报道, 指定的源语⾔言数据中发现特定类型 事件的各个实体、属性和关系。 监控其中新事件的报道,并且将 的事件,并且识别出与事件相关的 例例如:MUC-2是从海海军军事情报 同⼀一话题下的分散的报道按照某 信息填⼊入预设的事件模板中。 中抽取事件填⼊入预定义模板中, 种结构有效组织起来。 ACE中共计8⼤大类33个⼩小类的事件 共10个槽 TDT-3:240个topic
11. Challenges in Event Extraction (Open Domain) 33 event types 599 documents 6,000 labeled sentences English data in ACE 2005
12. Challenges in Event Extraction (Open Domain) 33 event types 599 documents How to represent features to in6d,0ic00altaebeelevdesnetnstences English data in ACE 2005
13. Challenges in Event Extraction (Open Domain) 33 event types 599 documents How to represent features to in6d,0ic00altaebeelevdesnetnstences How to acquire sufficient training data for multiple event types/classes English data in ACE 2005
14. Feature Representation
15. Traditional Methods for Feature Representation • Human designed features nsubj -> (cameraman plays the Victim role in die event) ????? -> (cameraman plays the Target role in Attack event) • Too much rely on imprecise NLP tools for feature extraction • Limitations for low-resources languages
16. Dynamic Multi-pooling Convolutional Neural Network • We propose a Dynamic Multi-pooling Convolutional Neural Network(DMCNN) to automatically capture the lexical-level and sentence-level features without POS tagging/Syntactic Parsing/… (ACL-2015) Yubo Chen, Liheng Xu, Kang Liu, Daojian Zeng and Jun Zhao, Event Extraction via Dynamic Multi-Pooling Convolutional Neural Networks, in Proceedings of ACL-IJCNLP 2015, Beijing, China, July, 26-29
17. Dynamic Multi-Pooling Layer • Traditional CNNs use max pooling • We perform Dynamic Multi-Pooling Barry Diller on Wednesday quit as chief of Vivendi Universal Entertainment.
18. Experimental Results • ACE Dataset • 40 newswire articles from ACE 2005 as test set • 30 other documents from different genres as development set • The rest (529) documents for training Compared with the state-of-the-arts Effect of dynamic max-pooling Our method achieves the best performance without the need of the existing NLP tools
19. Event arguments are important to the Event Detection Mohanmad fired Anwar, his former protege, in 1998. Attack or End-Position? • If we consider the argument phrase “former protege” (Role=Position), we will have more confidence to predict it as an End-Position event. Shulin Liu, Yubo Chen, Shizhu He, Kang Liu and Jun Zhao, Leveraging FrameNet to Improve Automatic Event Detection, in Proceedings of ACL 2016, Berlin, Germany, August, 7-12
20. More Attentions on Argument Words • The rp of trigger candidate w • the embedding of w • The rp of the contextual words • The rp of the contextual entities Cw and Ce are contextual words and entities, respectively. α is the attention vector, which is computed as the manner illustrated in the right figure.
21. Attention Supervision • Strategy 1: only pay attention to argument words • Strategy 2: pay attention to both argument words and their surroundings • Step 1: obtaining the raw attention vector in the same manner as S1 • Step 2: creating a new vector α’ with all points initialized with zero • Step 3: for each , we update the new vector: • • Step 4: calculating the final attention vector α* by normalizing α’
22. Regularization in Learning Model • Loss function of attentions • Joint loss function
23. Compared with State-of-the-arts • ACE2005
24. Training Data Generation
25. Generating Labeled Data from Structured KB • Distant (Weak) Supervision in Relation Extraction Knowledge base Relation Entity 1 Entity 2 Founder Steve Jobs Apple … … … Sentence Steve Jobs was the co-founder and CEO of Apple and formerly Pixar. Steve Jobs passed away the day before Apple unveiled iPhone 4S. …
26. The Strategy doesn’t work for Event Extraction • Triggers are not given out in existing knowledge bases RE: ( entity1, relation, entity2) We can use Michelle Obama and Barack Obama to label back EE:(event instance, event type; role1, argument1 ;...; rolen, argumennt) We can not use m:02nqglv and Barack Obama to label back In ACE, an event instance is represented as a trigger word
27. Training Data Generation • Step1: Event Trigger Words Extraction Assumption: The sentences mention all arguments denote such events • Step2: Argument Extraction/Role Identification • using Tigger words and Entities quit Barry Diller on Wednesday as chief of Vivendi Universal Entertainment. Solution: Using Key Arguments Yubo Chen, Shulin Liu, Xiang Zhang, Kang Liu and Jun Zhao, Automatically Labeled Data Generation for Large Scale Event Extraction, To appear in Proceedings of ACL 2017, Vancouver, Canada, July 30-August 4.
28. Arguments Selection for Events • Arguments for a specific event instance are usually mentioned in multiple sentences. Statistics of events in Freebase. Only 0.02% of instances can find all argument mentions in one sentence
29. Our Method for Generating Labeled Data in Event Extraction
30. Our Method for Generating Labeled Data in Event Extraction • Key Argument Detection • Role Saliency: • Event Relevance: • Key Rate: • Trigger Word Detection • Trigger Candidate Frequency: • Trigger Event Type Frequency: • Trigger Rate: we choose verbs with high TR values as the trigger words for each event type
31. Our Method for Generating Labeled Data in Event Extraction • Trigger Word Filtering and Expansion • We propose to use linguistic resource FrameNet to filter noisy verbal triggers and expand nominal triggers
32. Our Method for Generating Labeled Data in Event Extraction • Automatically labeled data generation • We propose a Soft Distant Supervision and use it to automatically generate training data
33. Neural Network of Event Extraction • DMCNN in ACL-2015
34. Employing Multi-Instance Learning to alleviate Noisy • Training Method: Multi-Instance Learning • In stage of argument classification, we take sentences containing the same argument candidate and triggers with a same event type as a bag and all instances in a bag are considered independently.
35. Experiments • Generated Labeled Data • Manual Evaluations of Labeled Data
36. • Main Results Results
37. Summary • Open Domain Event Extraction • Exploiting neural network for feature representation • Employing existing KB (Freebase) to automatically generating training data • This strategy could be expand to many domains and event types
38. Thanks for your attention Questions?