知網(wǎng)hownet-歡迎來到清華大學(xué)自然語言處理與社會人文

1、靜態(tài)的表達與動態(tài)的激活,董振東 dzd@keenage.com WWW.keenage.com 清華 2007-12,提綱,開場白 -- 知網(wǎng)不是什么？知網(wǎng)系統(tǒng)的概貌知網(wǎng)的創(chuàng)新點結(jié)語,開場白 -- 知網(wǎng)不是什么？（1）,在中文方面，也已有了一個類似詞匯網(wǎng)路的資源，叫做知網(wǎng)》（HowNet, http://www.keenage.c

2、om）。《知網(wǎng)》做法的特色是獨樹一幟；不采用英文詞匯網(wǎng)路的架構(gòu)只要采取他自己的架構(gòu)。而且他先把世界知識本體做個定義，在這定義里再去做區(qū)分。這個由上而下的方法，與英語與歐語詞匯網(wǎng)路由下而上的方法不同，當然有其可取之處?？上У氖?，由于當年資源與訊息的限制，并未與世界相關(guān)的研究接軌。基本上跟其他語言的詞匯網(wǎng)路連接，并無架構(gòu)上的基礎(chǔ)，而其上層知識分類，也是兩人的自由心證，不能說錯，卻也缺乏理論的基礎(chǔ)，面臨一些其他系統(tǒng)互通性（inter-oper

3、ability）的問題。,開場白 --知網(wǎng)不是什么？（2）,近年他在另外的場合又說:“HowNet is a database/network of semantic relationships among Chinese words. Conceptually it’s similar to WordNet of English, but the author claims they differ substantially.

4、For one thing, HowNet is NOT free. Well, they are making words A-D free for download, as a teaser.”,開場白 --知網(wǎng)不是什么？（3）,知網(wǎng)不是語義詞典、義類詞典、概念詞典、英漢雙語詞典 – 知網(wǎng)不是詞典知網(wǎng)不是漢化的WordNet、不是WordNet的中文代用品知網(wǎng)不是語言學(xué)研究的產(chǎn)物,知網(wǎng)系統(tǒng)的概貌,數(shù)據(jù)統(tǒng)計系統(tǒng)組成,數(shù)據(jù)統(tǒng)計

5、,Chinese character 7152 Chinese word & expression 92159English word & expression 86141Chinese meaning106591English meaning106731Definition 27877Record172097,,知網(wǎng)的創(chuàng)新點,理論創(chuàng)新知識獲取和表達創(chuàng)

6、新知網(wǎng)的知識力量,理論創(chuàng)新,知識論事件類概念間關(guān)系 – 雙軸論,知識獲取和表達創(chuàng)新,義原的獲取和選擇義原的組織和分類體系的建構(gòu)用結(jié)構(gòu)化語言（KDML）來定義概念，定義由兩種語言詞語表示的概念,義原的獲取和選擇,Sememes 2090Entity150 thing (physical, mental, fact) component (part, fitt

7、ing) time space (direction, location)Event (relation, state; action)810Attribute245AttributeValue 885Secondary feature121,義原的組織和分類體系的建構(gòu),實體 Entity 事件 Event屬性 Attribute 屬性值 Attri

8、buteValue次要特征 Secondary features事件角色 Event roles事件角色的典型演員 Typical actors of event roles公理關(guān)系與角色轉(zhuǎn)換 Axiomatic relations and role shifting反義義原對 Antonymous sememe pairs對義義原對 Converse sememe pairs,知網(wǎng)中概念的定義 （1）,Conc

9、ept definitions in HowNet – “buy”1. {GiveAsGift|贈:manner={guilty|有罪}, purpose={entice|勾引}}2. {buy|買}Cf. Synset definition in WordNet – “buy”1. buy, purchase (obtain by purchase;)2. bribe, corrupt, buy,

10、 make grease palm (make illeagal payment),知網(wǎng)中概念的定義 （2）,Concept definitions in HowNet – “buyer” {human|人:domain={commerce|商業(yè)},{buy|買:agent={~}}}Cf. Synset definition in WordNet – “buyer” buyer, purchaser, e

11、mptor, vendee (a person who buys)哪個 “buy”? -- 在 WordNet中是歧義的；但在 HowNet中是沒有歧義的,知網(wǎng)的知識力量 – 動態(tài)的激活,知網(wǎng)常識推理舉例概念相似度計算概念相關(guān)關(guān)系的建立,知網(wǎng)常識推理舉例,Can a doctor walk?下列句子的省略如何推導(dǎo)的？“我在南京買了幾本很好的詞典，到家發(fā)現(xiàn)全都丟了?！?-- 誰丟？丟什么？,Can a doctor w

12、alk? (1),1. “doctor”的定義DEF={human|人:HostOf={Occupation|職位}, domain={medical|醫(yī)},{doctor|醫(yī)治:agent={~}}}2. “entity”義原分類體系表│ │ │ ├ {AnimalHuman|動物} {animate|生物:HostOf={Sex|性別},{AlterLocation|變空間位置:agent={~}},{Sta

13、teMental|精神狀態(tài):experiencer={~}}} │ │ │ │ ├ {human|人} {AnimalHuman|動物:HostOf={Name|姓名}{Wisdom|智慧}{Ability|能力},{think|思考:agent={~}},{speak|說:agent={~}}},Can a doctor walk? (2),3. “event”義原分類體系表 │ ├ {AlterLocation|變空間位置}

14、 │ │ ├ {SelfMove|自移} │ │ │ ├ {SelfMoveInManner|方式性自移} │ │ │ │ ├ {roam|流浪} │ │ │ │ ├ {walk|走},公理關(guān)系與角色轉(zhuǎn)換 - 1,我在南京買了幾本很好的詞典，到家發(fā)現(xiàn)全都丟了。{buy|買} {obtain|得到} [consequence];

15、 agent OF {buy|買}=possessor OF {obtain|得到}; possession OF {buy|買}=possession OF {obtain|得到}. {obtain|得到} {own|有} [hypernym]; possessor OF {obtain|得到}=possessor OF {own|有}; possession OF {obtain|得到}=posse

16、ssion OF {own|有}.,公理關(guān)系與角色轉(zhuǎn)換 - 2,{lose|失去} {own|有} [precondition]; possessor OF {lose|失去}=possessor OF {own|有}; possession OF {lose|失去}=possession OF {own|有}.{lose|失去} {obtain|得到} [mutual precondition]; poss

17、essor OF {lose|失去}=possessor OF {obtain|得到}; possession OF {lose|失去}=possession OF {obtain|得到}.,概念相似度計算,貪官 學(xué)生0.307692貪官 教師0.355556貪官 校長0.386667貪官 市長0.454545walk run0.144444walk jump0.144444

18、walk swim0.130159walk fly0.124444walk buy0.018605,概念相關(guān)關(guān)系的建立,試比較HowNet關(guān)于WordNet的評述,試比較HowNet,舉例:buy床,關(guān)于WordNet的評述（1）,On WordNet[1] Jordan Bo Boyd-Graber et al., Oct. 2005, Adding Dense, Weighted Co

19、nnections to WordNet (Princeton paper)[2] Rila Mandala et al., ACL W98, 1998, The Use of WordNet in Information Retrieval (TIT paper),關(guān)于WordNet的評述（2）,Princeton paper reads:“1.1 Shortcomings of WordNetNo cross-part-of-

20、speech links [traffic (n) – stop (v)]Too few relations [chopsticks – Chinese restaurant]No weighted arcs [run:jog; run:move]”,關(guān)于WordNet的評述（3）,Princeton paper continues:To address these shortcomings, we are working to

21、enhance WordNet by adding a radically different kind of information. The idea is to add quantified, oriented arcs between pairs of synsets,e.g. from {car, auto} to {road, route}, from {buy, purchase} to {shop, store}, a

22、nd also in the opposite direction. Each of these arcs will bear a number corresponding to the strength of the relationship. We chose to use the concept of evocation – how much one concept evokes or brings to mind the oth

23、er – to model the relationships between synsets.,結(jié)語,知識是關(guān)系的系統(tǒng)；知網(wǎng)是描述概念與概念間的關(guān)系以及概念的屬性與屬性間的關(guān)系的知識系統(tǒng)；知網(wǎng)描述的關(guān)系是可計算的；知網(wǎng)在本質(zhì)上不同于WordNet；知網(wǎng)在發(fā)展。,謝謝！歡迎來到www.keenage.com,附錄 -- 普遍的語義機制,跳：跳河 -- jump into a river (LocationFin)跳

24、樓 -- jump off a high building (LocationIni)跳墻 -- jump over a wall (LocationThru)導(dǎo)：導(dǎo)游 -- 導(dǎo)購 -- 導(dǎo)診托：醫(yī)托 -- 婚托野：野餐 -- 野炊 -- 野營 -- 野游 -- 野泳 / 野浴,附錄 -- 基本數(shù)據(jù)統(tǒng)計,中文 ：06-04-07synset: Set = 13700 (13692) (13

25、463 ) Word Form = 55180 (55150) (54312)antonym: Set = 13154 (13145) (12777)converse: Set = 6803 (6804) (6753) 英文 ：synset: Set = 18622 (18610) (18575) Word Form = 58622 (5858

26、8) (58488)antonym: Set = 12269 (12268) (12032)converse: Set = 6455 (6454) (6442),附錄 -- 1. 事件框架 ~ Verb frame- {event|事件} ├ {static|靜態(tài)} {event|事件} │ ├ {relation|關(guān)系}

27、 {static|靜態(tài)} │ │ ├ {possession|領(lǐng)屬關(guān)系} {relation|關(guān)系} │ │ │ ├ {own|有} {possession|領(lǐng)屬關(guān)系:possessor={*},possession={*}} │ │ │ │ ├ {obtain|得到} {own|有:possesso

28、r={*},possession={*},source={*}} └ {act|行動} {event|事件:agent={*}} ├ {ActGeneral|泛動} {act|行動:agent={*}} └ {ActSpecific|實動} {act|行動:agent={*}}

29、 └ {AlterSpecific|實變} {ActSpecific|實動:agent={*}} ├ {AlterRelation|變關(guān)系} {AlterSpecific|實變:agent={*}} │ ├ {AlterPossession|變領(lǐng)屬} {AlterRelation|變關(guān)系:agent={*},

30、possession={*}} │ │ ├ {take|取} {AlterPossession|變領(lǐng)屬:agent={*},possession={*},source={*}} │ │ │ ├ {buy|買} {take|取:agent={*}, possession={*}, source={*},

31、 cost={*}, beneficiary={*},附錄 -- 2. 事件角色的典型演員 ~ VerbNet│ ├ {buy|買} {take|取:agent={human|人}{group|群體->}, possession={artifact|人工物->}, source={human|人}{InstitutePlace|