深度語法分析在自然語言的應(yīng)用

1、深度語法分析在自然語言的應(yīng)用技術(shù)創(chuàng)新，變革未來Outline人工智能的歷史和現(xiàn)狀簡介：從感知到認(rèn)知此消彼長的經(jīng)驗(yàn)主義和理性主義鐘擺深度解析(Deep Parsing)是什么？NLP 架構(gòu)縱覽核武器應(yīng)用舉例Outline: AI History人工智能的歷史和現(xiàn)狀簡介：從感知到認(rèn)知此消彼長的經(jīng)驗(yàn)主義和理性主義鐘擺NLP Mainstream Since 1990sCourtesy of Prof. Church: “A Pendulum Swung Too Far”/blog-362400-988692.htmlTwo Basic Approaches to NLPComplementary r

2、ather than competingHybrid: Best of both worldsBalance and configurability between precision and recallGrammar Engineering(based onsentence structure)Good for sentence levelHandles short messages wellHigh precisionFine-grained insightsEasy to debugParsing and understandingRequires deep skillsRequire

3、s scale up skillsRequires robustness skillsModerate recall (coverage)Parser development slowApproachProsConsStatistical Learning(based on keywords)Good for document-levelHigh recallRobustEasy to scaleFast development (if data available)Requires large annotationCoarse-grainedDifficult to debugFail in

4、 short messagesOnly shallow NLPNo understandingOutline: What Is Deep Parsing人工智能的歷史和現(xiàn)狀簡介：從感知到認(rèn)知此消彼長的經(jīng)驗(yàn)主義和理性主義鐘擺深度解析(Deep Parsing)是什么？Deep Parsing: Unstructured to StructuresWhy parsing?Limited PatternsSubtree Pattern: Data to IntelligenceSVO Paern:Barack Obama (S) Endorse (V) Hillary Clinton (O)Know

5、ledge GraphDeep Parsing: Unstructured to StructuresSubtree Pattern: Data to IntelligenceInter-Clause Paern:雖然 遺憾無所謂Linear: Innite number of sentencesStructure:Limited paernsmild senGmentData IntelligenceOutline: NLP Architectures人工智能的歷史和現(xiàn)狀簡介：從感知到認(rèn)知此消彼長的經(jīng)驗(yàn)主義和理性主義鐘擺深度解析(Deep Parsing)是什么？NLP 架構(gòu)縱覽NLP Ar

6、chitecture 1: Deep Parser as CoreCascaded FSAs break through Chomskys hierarchy walls Robust, linear, F-measure: scale up to big dataSample Deep Parse Tree (dependency)Sample Deep Parse Tree (PS flavor)NLP Architecture 2: Information ExtractionIncluding sentiment analysis (on subjective language)NLP

7、 Architecture 3: Text MiningNLP Architecture 4: Landing on ApplicationsSample Deep Parse TreeSample Deep Parse TreeOutline: NLP Applications人工智能的歷史和現(xiàn)狀簡介：從感知到認(rèn)知此消彼長的經(jīng)驗(yàn)主義和理性主義鐘擺深度解析(Deep Parsing)是什么？NLP 架構(gòu)縱覽核武器應(yīng)用舉例社媒輿清分析，大數(shù)據(jù)挖掘，智能搜索，對話系統(tǒng) Sentiment AnalysisWhy deep parsing, not deep learning?Learning wi

8、thout parsing does not work for social media sentimentSocial media is dominated by short messagesStatistical learning breaks in short messages: no sufficient data pointsDeep parsing enables linguistic analysis for best precisionDeep parsing enables insights mining 2 magnitudes more efficientparsing-

9、supported rule has power of about 100 ngram rulesDeep learning is a great algorithm but still delinked from parsingParsers trained by deep learning are all research systemsdifficult to adapt to real life text of social media (or other genres)knowledge bottleneck: domains where labeled data are insuf

10、ficientReal life deep learning systems are mostly end-to-end, still no structuresanalysisSentiment Analysis: Bag of Words vs. ParsingKEYWORD CHALLENGEThe iPhone has never been good.The iPhone has never been this good.ASSOCIATION CHALLENGEAnother reason to switch from Visa to MasterCard I prefer Mast

11、erCard over Visa.MasterCard is way better than Visa.CLASSIFICATION CHALLENGEI had a wonderful day today. Even my instant coffee tastes great. However my Dell laptopdoesnt boot again. Maybe I should check out the MacBook. It MacBook seems so easy to use.FCionaer-sgera-ginraeidneAdnaClylassissiuficnac

12、toiovnertshu“wmhbys”-buephainndd sdeonwtinm: eonvtesr:all tone positive (3 vs 1)Instant coffee / tastes greatDell Laptop / does not bootMacbook / easy to useDeep Parsing Supports Deep SentimentsSentiment analysis has different layerssentiment classification: thumbs-up and down (or neutral)sentiment

13、association: to associate a sentiment with a topic or brand as its objectdeep sentiment insights:who has the sentiment?how intense?why?Evaluations, comparisons and contrasts;needs and wish-list;positive/negative actions (e.g. adopt / abandon);purchase intent;pros and consMost learning systems stop a

14、t 1 and sometimes at 2.All 3 can be done via deep parsing.Illustration: Real-time PollsChallenges observed:economy topicat 6:55pm;China topic at 7:30pmIllustration: Stock Market TrendsTopic: HTCData 1: Stock Market PerformanceData 2:Chinese social media (Weibo, Tianya, Facebook, Twitter)Time range:2

15、013/08 2014/08Strong correlation observedBig Data Mining: Who benefits?For businesses: social listeningConsumer in:sights: sentiments and why Brand image: trendsCompetitive research: where do we standFor consumersPurchase decision Personalized serviceFor governmentElection campaignPublic opinions on

16、 policies and social topicsOthers?Hot topics or anywhere public opinions are involved Stock market trends correlationFor consumers: Purchase DecisionFor consumers: Purchase DecisionFor consumers: Purchase DecisionIntelligent Search and ChatbotsA mixture/convergence of 3 is possibleThree types of Cha

17、tbots:Domain knowledge QA:e.g. customer service;Open domain knowledge QA:e.g. Who won Nobel Prize in 2015?Interactive chatting: e.g. just for fun (killing time);in time, for comfort (senior people); for mental health counsellingQ:questions are a subset of language, tractablefor decoding intent, asking p