Natural Language Processing: An Introduction. Recent Significant Advances In NLP. Recent Significant Advances In NLP. NLP: The Ultimate Goal (1990)

Transcription

1 NLP: The Ultimate Goal (1990) The Ultimate Goal For computers to use NL as effectively as humans do. Natural Language Processing: An Introducti Natural language wher spoken written or typed most natural means communicati between humans and mode expressi choice for most documents y produce. As computers play a larger role in preparati acquiti transmsi mitoring storage analys and transformati informati endowing m with ability to understand and generate informati expressed in natural languages becomes more and more necessary. CIS Intro to AI 2 NLP: Grand Challenges (1990) The Ultimate Goal For computers to use NL as effectively as humans do. Reading and writing text Abstracting Mitoring Extracti into Databases Interactive Dialogue: Natural effective access to computer systems Informal Speech Input and Output Recent Significant Advances In NLP Web-scale informati extracti & questi answering IBM s Wats Interactive Dialogue Systems Apple s Siri (Microst Cortana) (Google Now) Translati: Input and Output in Multiple Languages CIS Intro to AI 3 CIS Intro to AI 4 Recent Significant Advances In NLP Automatic Machine Translati Xinhua story (Chinese) Google translate 新华网海牙 3 月 24 日电 ( 记者陈贽潘治 ) 第三届核安全峰会 24 日在荷兰海牙举行 国家主席习近平出席并发表重要讲话 介绍中国核安全措施和成就 阐述中国关于发展和安全并重 权利和义务并重 自主和协作并重 治标和治本并重的核安全观 呼吁国际社会携手合作 实现核能持久安全和发展 CIS Intro to AI 5 The Hague March 24 Xinhua (Reporter Chen Zhi Pan Governance ) Third Nuclear Security Summit held in The Hague 24th. State President Xi Jinping attended and delivered an important speech China's nuclear safety measures and achievements elaborated China development and security both rights and obligatis both both independently and in collaborati both temporary and permanent nuclear security ccept called internatial community to work toger nuclear achieve lasting security and development. Early Successes: Human Machine Interfaces SHRDLU (Winograd 1969) A fragile demstrati fundamental vi LUNAR (Woods Webber Kaplan 1971) Answering geologt s questis about Apollo 11 mo rocks CIS Intro to AI 6 1

2 Review: SHRDLU: A demstrati pro LUNAR William Woods 1971 NLP interface to database analyses Apollo 11 mo rocks Examples What average ccentrati aluminum in high alkali rocks? How many breccias ctain olivine? Give me modal analyses those samples for all phases. Handled 78% sentences typed by geologts at 1971 Lunar Rocks cference (12% more with minor fixes ) CIS Intro to AI 7 CIS Intro to AI 8 The Past: Crucial flaws in paradigm These and or later systems worked well BUT 1. Pers-years work to port to new applicatis 2. Very limited coverage Englh Crucially y worked well because a magical fact: The State NLP NLP Past before 1995: Rich Representatis NLP Present: Powerful Stattical Dambiguati People automatically adapt and limit ir language given a small set exemplars if underlying lingutic generalizatis are HABITABLE Th w t handle pre-exting text! CIS Intro to AI 9 CIS Intro to AI 10 A Few Core Technologies Informati Extracti & Named Entity Recogniti 1. Named Entity Recogniti& Informati Extracti 2. Machine Translati 3. Text Summarizati CIS Intro to AI 11 2

3 President Slobodan Milosevic Thursday Associati Banks by its Milos who also general JugoBanka President Slobodan Milosevic Thursday Associati Banks by its Milos who also general JugoBanka Informati Extracti Informati extracti identificati in text specified classes Named Entities + Relatis Events For relatis and events th includes finding participants and modifiers (date time locati etc.). Goal: fill out a data base with given relati or event types: people s jobs people s whereabouts merger and acquiti activity dease outbreaks genomics relati CIS Intro to AI 13 Extracti Example George Garrick 40 years old Ldbased European Informati Services Inc. was appointed chief executive ficer Nielsen Marketing Research USA. USA. Positi Company Locati Pers Status President European Informati Ld George Garrick Out Services Inc. CEO Nielsen Marketing Research USA George Garrick In CIS Intro to AI 14 Named Entity Recogniti The task: identify atomic elements informati in text Flag who where when & how much in text Pers names Company /organizati names Locatis Dates & times Percentages Metary amounts W t simple lts solve problem? too numerous to include in dictiaries changing cstantly appear in many variant forms subsequent occurrences might be abbreviated lt search/matching doesn t perform well CIS Intro to AI 15 CIS Intro to AI 16 Levels BBN Stattical Analys (2005) S VP Name finding Parsing Co-reference S SBAR VP WH VP GPE Pers ORG Pers ORG CIS Intro to AI President Slobodan Milosevic Thursday Associati Banks by its Milos Milosavljevic who also general JugoBanka. Milosavljevic 17 Informati Extracti from Propositis Propositis are normalized cnectis from parse trees. Entities and relatis are extracted stattically from propositis. Pers: Slobodan Milosevic Positi: Organizati: ia Pers: Milos Milosevic Positi: Organizati: Associati Banks subj obj obj subj arg arg Pers: Milos Milosevic Positi: general Organizati: JugoBanka GPE Pers Date Pers ORG ORG Pers Pers ORG CIS Intro to AI 18 Milosavljevic 18 3

4 Stattical Machine Translati Technology Machine Translati Spanh/Englh Bilingual Text Stattical Analys Englh Text Stattical Analys (Several slides from Language Weaver) Spanh Broken Englh Englh Que hambre tengo yo What hunger have I Hungry I am so I am so hungry Have I that hunger CIS Intro to AI 20 I am so hungry How A Stattical MT System Learns Translating a New Document CIS Intro to AI 21 CIS Intro to AI 22 Broadcast Mitoring BBN MAPS & Language Weaver MT CIS Intro to AI 23 CIS Intro to AI 24 4

5 Language Weaver Hybrid Translati Technology Chinese Source Text Sample 1: 车展 一向是衡量一个国家汽车消费现状和市场潜力的 晴雨表 本届北京国际车展有 24 个国家的 1200 余家厂商参展 8 天接待 40 余万名参观者 创下了中国车展的新纪录 让人深切地感受到汽车市场启动的信号 中国是世界最后一个最大的汽车市场 多年来 这句话更多地包含着汽车商人的一种希冀 然而如今 越来越多的事实预示着它正在变为现实 来自本届车展的一组数据很有说服力 北京青年报 的一份现场调查显示 35 岁以下参观者约占 35%;62.1% 的被访者表示 参观车展主要是为近期买车搜集信息 甚至在展会上就有可能购买或预订合适的产品 ;76% 的被访者表示最近两年会购买私家车 今年以来 国内轿车市场的强劲增长让厂家喜上眉梢 据国家统计局公布的数字 前 4 个月 全国共生产轿车 万辆 增长 27.6%; 特别是 4 月份 生产轿车 9 万辆 同比增长 50.5% 创造了十几年来轿车月产增长的最高纪录 从销售看 一季度 全国轿车生产企业共销售轿车 18.8 万辆 同比增长 22% 产销率达 105%; 轿车库存比年初下降 1.1 万辆 下降幅度近 25% Language Weaver Experimental Syntax MT Sample 1 : Text Summarizati The motor show has always been ' barometer' a natial car csumpti and market potential. The Beijing Internatial Auto Show has more than 1200 exhibitors from 24 countries and 8 days receiving more than 40 milli vitors setting a new record in China's auto show are deeply aware automobile market signals. "China e largest automobile market in world. Over years th phrase implies more auto businessmen. But now more and more facts indicates that it to become a reality. Data from Motor Show very cvincing. The Beijing Qingnian Bao Report --spot investigati showed that about 35 percent 35-year-old vitors 62.1 percent respdents said that truck was mainly to buy a car in near future to collect informati even at exhibiti may purchase or suitable products; 76% respdents indicated in past two years to buy private cars. Since beginning th year strg growth domestic car market. According to figures released by Natial Bureau Stattics in first four mths country produced vehicles up 27.6 percent; in particular in April producti vehicles an increase 50.5% over same period last year setting a record high for mthly output growth over past 10-odd years. In terms sales in first quarter manufacturing enterpres in country sold cars up 22 percent over same period last year up 10.5 percent; vehicles dropping by nearly 25 percent lower than beginning year. (For more th topic check out courses taught by Pr. Ani Nenkova) CIS Intro to AI 25 (Includes work by Pr. Nenkova) CIS Intro to AI 27 5