版权说明:本文档由用户提供并上传,收益归属内容提供方,若内容存在侵权,请进行举报或认领
文档简介
1、A Linear Programming Formulation for Global Inference in Natural Language Tasks,Dan RothWen-tau Yih Department of Computer Science University of Illinois at Urbana-Champaign,Page 2,View of Solving NLP Problems,Page 3,Weaknesses of Pipeline Model,Propagation of errors Bi-Directional interactions betw
2、een stages,Occasionally, later stage problems are easier.,Upstream mistakes will not be corrected.,Page 4,Global Inference with Classifiers,Classifiers (for components) are trained or given in advance. There are constraints on classifiers labels (which may be known during training or only known duri
3、ng testing). The inference procedure attempts to make the best global assignment, given the local predictions and the constraints,Page 5,Ideal Inference,Page 6,Inference Procedure,Inference with classifiers is not a new idea. On sequential constraint structure: HMM, PMM, CRFLafferty et al., CSCLPuny
4、akanok xP2=NP+ xP2=1 xP3=NP+ xP3=1; xP4=NP+ xP4=1,i, j xPhrase_i = Class_j = 1,Non-overlapping Constraints: xP1= + xP3=1 xP2= + xP3= + xP4=2,Page 11,LP Formulation,Generate one integer linear program per sentence.,Page 12,Entity/Relation Recognition,John was murdered at JFK after his assassin, Kevin
5、 Identify:,John was murdered at JFK after his assassin, Kevin ,location,person,person,Kill (X, Y),Identify named entities Identify relations between entities Exploit mutual dependencies between named entities and relations to yield a coherent global prediction,Page 13,Problem Setting,R12,R21,R23,R32
6、,R13,R31,The relation between each pair of entities is represented by a relation variable; most of them are null.,The goal is to assign labels to these E and R variables.,Constraints: (R12 = kill) (E1 = person) (E2 = person) (R12 = headquarter) (E1 = organization) (E2 = location) ,Page 14,LP Formula
7、tion Indicator Variables,For each variable xE1 = per, xE1 = loc, , xR12 = kill, xR12 = born_in, , xR12 = , 0,1 For each pair of variables on an edge xR12 = kill, E1 = per, xR12 = kill, E1 = loc , , xR12 = , E1 = per, xR12 = , E1 = loc , , xR32 = , E2 = per, xR32 = , E2 = loc , 0,1,Page 15,LP Formula
8、tion Cost Function,Assignment cost cE1 = per xE1 = per + cE1 = loc xE1 = loc + + cR12 = kill xR12 = kill + + cR12 = xR12 = + ,Constraint cost cR12 = kill, E1 = per xR12 = kill, E1 = per + cR12 = kill, E1 = loc xR12 = kill, E1 = loc + + cR12 = , E1 = loc xR12 = , E1 = loc + ,Costs are given by classi
9、fiers.,-,Total cost = Assignment cost + Constraint cost,Page 16,NodeEdge Consistency Constraints,Binary Constraints Unique-label Constraints,LP Formulation Linear Constraints,Subject to:,NodeEdge Consistency Constraints,Page 17,LP Formulation,Generate one integer linear program per sentence.,Page 18
10、,Experiments Data,Methodology: 1,437 sentences from TREC data; 5,336 entities; 19,048 pairs of potential relations.,Page 19,Experimental Results F1,Entity Predictions,Relation Predictions,Improvement compared to the basic (w/o inference) and pipeline (entityrelation) models Quality of decisions is e
11、nhanced No “stupid mistakes” that violate global constraints,Page 20,Decision-time Constraint,Constraints may be known only in decision time. Question Answering: “Who killed JFK?” Find “kill” relation in candidate sentences,Find the arguments of the “kill” relation,Page 21,Computational Issues,Exhau
12、stive search wont work Even with a small number of variables and classes, the solution space is intractable n=20, k=5, 520 = 95,367,431,640,625 Heuristic search algorithms (e.g., beam search)? Do not guarantee optimal solutions In practice, may not be faster than ILP,Page 22,Generality (1/2),Linearc
13、onstraints can represent any Boolean function More components can be put in this framework Who killed whom? (determine arguments of the Kill relation) Entity1=Entity3 (co-ref classifier) Subj-Verb-Object constraints Able to handle non-sequential constraint structure E/R case has demonstrated this pr
14、operty,Page 23,Generality (2/2),Integer linear programming (ILP) is NP-hard. However, an ILP problem at this scale can be solved very quickly using commercial packages, such as CPLEX or Xpress-MP. CPLEX is able to solve a linear programming problem of 13 million variables within 5 minutes. Processing 20 sentences in a second for a named entity recognition task on P3-800MHz,Page 24,Current/Future Work,Handle stochastic (soft) constraints Example: If the relation is kill, the first argument is person with 0.95 probab
温馨提示
- 1. 本站所有资源如无特殊说明,都需要本地电脑安装OFFICE2007和PDF阅读器。图纸软件为CAD,CAXA,PROE,UG,SolidWorks等.压缩文件请下载最新的WinRAR软件解压。
- 2. 本站的文档不包含任何第三方提供的附件图纸等,如果需要附件,请联系上传者。文件的所有权益归上传用户所有。
- 3. 本站RAR压缩包中若带图纸,网页内容里面会有图纸预览,若没有图纸预览就没有图纸。
- 4. 未经权益所有人同意不得将文件中的内容挪作商业或盈利用途。
- 5. 人人文库网仅提供信息存储空间,仅对用户上传内容的表现方式做保护处理,对用户上传分享的文档内容本身不做任何修改或编辑,并不能对任何下载内容负责。
- 6. 下载文件中如有侵权或不适当内容,请与我们联系,我们立即纠正。
- 7. 本站不保证下载资源的准确性、安全性和完整性, 同时也不承担用户因使用这些下载资源对自己和他人造成任何形式的伤害或损失。
最新文档
- 矿山智能安全智能集成技术创新与应用技术创新总结报告
- 土方工程雨季施工方案
- 2026年传染病医学试题及答案
- 工业园区项目防水工程质量通病防治施工方案
- 2026医院医疗卫生法律法规题库和参考答案
- 铁路大型物件运输事故应急预案演练脚本
- 养老院老人肺栓塞应急预案演练脚本
- 仓库管理单位焊工检修维修安全操作规程
- 建筑房屋拆除工程施工组织设计方案
- 锐丰中心地铁通道装修施工方案含土建装修机电样本
- 水工建筑物水下缺陷修复技术导则
- 2026-2030中国特种空调行业盈利动态及供需状况分析报告
- 2026 齐商银行笔试核心高频考点及题库
- 2025年贵州黔南人力资源开发有限责任公司招聘劳务派遣制专职民兵教练员5人笔试备考试题及答案解析
- 药师执业行为规范(2026年版)
- 2026年5月版-安全环境职业健康法律法规、规章、标准文件清单
- 鞋厂针车车间考核制度
- 《林海雪原》专项训练(简答题)解析版-2025-2026学年六年级语文下册整本书阅读(统编版五四学制)
- 《海南省工程勘察设计收费导则(试行)》
- CMA申请介绍教学课件
- 医院中医药临床案例库建设与管理
评论
0/150
提交评论