第15 讲 知识图谱中的开放问题 - Fudan...
20
Transcript of 第15 讲 知识图谱中的开放问题 - Fudan...
重新审视知识图谱技术的发展
应用场景
技术生态
• 从通用 vs 领域／行业应用
• 从搜索延伸至推荐、问答等复杂任务
• 从简单关系发现到深层关系推理
• 从回答what问题到回答why问题
• 从关键词交互到更自然的人机对话式交互
• 机器学习
• 自然语言处理
• 知识图谱
2
变革
从通用到领域／行业应用
3
医学知识库 代码知识库 军事知识库
电信知识库 工商知识库 电商知识库
计算机知识库 网络运维知识库 一带一路知识库
通用百科知识图谱：CN-DBpedia累计API调用近4亿次
观点1： 通用知识库在通用人工智能中扮演重要角色，是未来竞争的战略制高点，不容忽视观点2: 自动化大规模通用KG构建、样本稀疏环境下的自动化领域KG构建仍是瓶颈问题
从搜索延伸至推荐、问答等复杂任务
4
Search
Search keywords 推荐
知识问答（小Cui问答）观点： KG将在更多复杂、多元任务中发挥重要作用
从关键词交互到更自然的人机对话式交互
5
Google Now Apple Siri Amazon Alexa Microsof Cortana
观点： 自然语言交互将成为人机交互的主流方式，将对机器自然语言认知提出更高要求
Question Answering (QA) systems in industries
从回答what问题到回答why问题
6观点： 可解释是未来人工智能发展的核心诉求之一，是人机互信的前提
What, Who, When，Where
How, Why
从简单关系发现到深层关系推理
7
Why baoqiang select Qizhun Zhang as his lawyer? Why A invests B?
观点： 隐式关系发现、深层关系推理将成为智能的主要体现之一
技术生态
机器学习
• 深度学习飞速发展，在数据丰富的场景与任务下取得较好效果
自然语言处理
• 自然语言处理（NLP）的模型与算法，在深度学习模型推动下，发展迅速
知识图谱
• 英文图谱积累迅速，大量高质量手工构建图谱已在应用中发挥作用
8
机器学习
• 小样本学习、无监督学习手段有限；现有模型难以有效利用大量先验知识
自然语言处理
• 总体上离实际应用需求还很远；NLU（自然语言理解）还很初步
知识图谱
• 其他语种图谱缺乏、常识缺乏；有效利用大规模数据驱动的知识图谱手段缺乏
机遇与挑战• 知识图谱构建的有效策略与方案
• 如何充分利用知识的跨语言特性；如何区别对待数据来源的不同结构化程度；基于概念模板的迭代式抽取；基于语义与语法混合模式的抽取
• 《Data Driven Approaches for Large Scale Knowledge Graph Construction》
• 大规模常识的获取与理解
• 当前人工智能普遍缺乏常识理解能力，常识缺乏是人工智能研究的重大制约瓶颈。人工智能的发展必须尽快突破这一瓶颈。常识理解是通用人工智能的核心问题，是人工智能发展的战略制高点。为此，必须尽快开展大规模常识获取与理解的研究。
• 《Large Scale Commonsense Knowledge Acquisition and Understanding》
• 样本稀疏环境下的领域知识获取
• 有着丰富样本的高频知识的获取能够通过常规统计学习模型有效解决。但是出现频次较低的长尾知识的获取一直是知识工程领域的困难问题。如何将高频知识的获取模型有效迁移到样本稀疏的低频知识，是当前知识获取领域面临的重大难题。
• 《电商智能化与知识图谱》
9
机遇与挑战• 数据驱动与知识引导深度融合的新型机器学习模型
• 如何将符号化知识有机融入基于数据的统计学习模型，是当前人工智能的重大问题，是降低深度学习模型的样本依赖，突破机器学习模型效果的天花板的关键所在。
• 《当知识图谱遇见深度学习》，中国人工智能通信
• 基于知识图谱的可解释人工智能
• 符号化的知识图谱具有形象直观的特性，为弥补深度学习在解释性方面的缺陷提供了可能。如何利用知识图谱解释各类机器学习，特别深度学习，以及高层次决策模型的结果，是当前人工智能领域的重大科学问题。
• 《Explainable AI using Knowledge Graphs》
• 《基于知识图谱的可解释人工智能：机遇与挑战》
• 知识获取中的人机协作机制与方法
• 当前人工智能的发展还需要人类的有效指导。人在环中是人工智能发展的基本模式。然而如何有效利用人力，如何有效切分人机合作的边界仍是知识获取过程中人机协作机制的核心问题。
• 《未来人机区分》
10
机遇与挑战• 知识驱动的机器语言理解
• 机器语言理解释构建在机器认知能力基础之上的。大量知识库的涌现为知识驱动下的机器认知能力的实现提供了新的机遇，为机器语言理解提供了大量背景知识，使得机器理解人类语言成为可能。为此，必须深入开展知识驱动的机器语言理解模型与方法的研究
• 《Enabling Machines to Understand Human Language》
• 知识驱动的搜索与推荐• 在应用层面，大量知识库的出现为互联网应用的两个核心问题搜索与推荐带来全新的机遇。如何有效利用符号知识理解搜索意图，改进检索效果，提升推荐精准性是当前互联网应用的核心问题。
• 《Knowledgeable Search and Recommendation》• 《User Understanding with Knowledge Graph》
11
知识图谱构建的有效策略与方案
• 充分利用知识的跨语言特性，但也要注意知识的本地文化特性性
12
English Entity: Andy Lau
刘德华 @zh
Chinese Entity: 刘德华
rdfs:label rdfs:label
rdf:type
Thing Agent Person Artist Actor MusicArtist
步骤一：跨语言
实体链接
步骤二：跨语言
概念传递
Principle： knowledge is independenton languages, but its expression in a specific language is usually biased
知识图谱构建的有效策略与方案
• 先易后难• 结构化-> 半结构化->非结构化
• 避免从零开始• 以通用图谱中的领域图谱作为种子• 问题：如何有效发现领域实体与关系？
• 跨领域迁移• 从邻近领域迁移• 问题：如何迁移具有共性的知识？
13
通用图谱军事
法律
文娱工商
技术
• 增强AI模型的透明(transparency)是人机互信的前提
• Mr. David Gunning• New machine-learning
systems will have the ability to explain their rationale, characterize their strengths and weaknesses, and convey an understanding of how they will behave in the future.
14
基于知识图谱的可解释人工智能
https://www.darpa.mil/program/explainable-artificial-intelligence
• 解释概念• Bachelor->A man, unmarried
(in IJCAI 2016)
• 解释实体集• China, Japan, India-> Asian
Country (in IJCAI 2015)
• 解释推荐• Sumsung S6, Iphone 6->
Huawei P9
15
基于知识图谱的可解释人工智能
解释推荐, in IJCAI2017
• 面向长尾关系的端到端抽取• 高频向长尾迁移
• 基于深度学习的迁移学习• 富样本向贫样本领域迁移
16
样本稀疏环境下的领域知识获取
top five types ofthe entity thathas the largestfrequency inD B pedia.Then,w e derive the type representation forentitye by m ax-pooling the type em beddings oftop five types fore:
T e = M axP ool(T c1 ,T c2 ...) (5)
w here each ci is one of the top five types.O ur m odelw illtrain the em bedding for each type.W e use Te as the typerepresentation for each token in the description text of e.Thatis w e setTt = Te in Eq.6.
M ultiple Predicate E xtractor w ith PriorK now ledge (M PK )The single predicate extractor is trained separately on dataof different predicates. In the case of predicates w ith in-sufficienttraining data,like predicate previousWork andfoundedBy,the single extractors m ay notbe fully trained,w hich im pairs the perform ance.
C 4.Token R epresentation w ith Predicate as Input (D t)N aturally, w e expect to utilize the training sam ples fromother predicates, w hich m otivates us to develop a m ultiplepredicate extractorstructure.The m ultiple predicate extrac-tornetw ork structure isshow n in the leftofFigure 5.W e addthe predicate as an inputas w ell,therefore presenting a uni-fied m odeltrained for allthe predicates.Sim ilar to Vt,w eencode the predicate inputinto a 30 dim ension vector D t.Thereby,the token is represented by:
xt = concat(V t,P t,T t,D t) (6)
M ultiple Predicate Scanning M odel(M PS)The M PK m odel already has the ability to encode predi-cate inform ation.In order to further enhance the m odelforknow ledge extraction purpose,w e need m ore com plicatedprocessing unitfollow ing the predicate input.O ur m otiva-tion here is to learn the representation forhow the predicateis expressed in text and how it interacts w ith the context.Thus,w e regard our m odel as an entia of three parts.W eassum e that one part is in charge of reading and process-ing sentences regardless of the predicate; the another parttackles specific predicates representation, and the last partcom bines the previous results.Tw o inputparts ofthe m odelare the sentence input(togetherw ith priorknow ledge)x andthe predicate inputd.The three functionalcom ponents are:1) L (x) that encodes the sentences into a m iddle vector;2)D (d) thatencodes the predicate into a m iddle vector;3)u(L (x),D (d)) along w ith its follow ing LSTM layer,w hichcom bines allm iddle states and produces the finallabel.Inthissetting,L (x)isindependentofthe predicates,and itcanbe trained w ith allcorpus.Thus,in the training process ofsom e sam ple-rich predicates,the function L (x),i.e.the pro-cess ofthe sentences w illbe learned in com m on.The em bedding of predicates rem ains to be d,follow ed
by tw o fullconnection layers D .The connection details aredefined in Eq.7:
x t = con cat(V t,P t,T t)
L t = L S T M (x t)
D = σ(W d⇤(σ(W d d + bd )) + bd⇤)
u t = W u [L t;D ]+ bu
(7)
Figure 5:N etw ork structures of the m ultiple predicate extractor.The leftM PK m odel,and the rightone isM PS m odel.
Training M ethodologyTraining data generationSim ilarw ith thedistantsupervision setting,w e draw hs,p,oitriplesfrom K now ledge basesand itssentencesfrom thecor-responding text of entity s.A ny occurrence of o in thesesentences are labeled w ith 1 and any other tokens are la-beled w ith 0.W e only label the abstract paragraphs in thefree textto lim itthe search space.Ifthe paragraphs exceed100 w ords,w e truncate them .Ifthere is no occurrence ofo,w e justskip it.A fterw ards,each entity s together w ith thelabeled textbecom e one sam ple forpredicate p.
N egative sam pling A m ajor concern is that the sam plesderived by the above procedure contain only“positive”sam -ples forp.H ow ever,this is notalw ays the case even w hen ooccurs in the text.For exam ple,Hollywood could be theobjectvalue forboth BirthPlace and DeathPlace.W estillneed negative sam ples to train the extractors.Therefore,w e random ly selectX s from otherp w ith the sam e labeledo.W e change the labels ofeach tokens into 0 and take s asw ellas the re-labeled tokens to form a negative sam ple forp.W e repeatthe above sam pling procedure untilthe num berofnegative instances is equivalentto positive instances.N ote that the above labeling procedure m ightencountersom e hard cases caused by the com plexity of natural lan-guage.First,som e sentences thatm ention o m ightexpressrelationships otherthan p.Second,som e o has differentex-pressions in free textagainstits infobox,w hich m akes theexactstring m atching to identify o failin som e cases.Fortu-nately,W ikipedia containstensofm illionsofentities,w hichallow s us to constructa large enough training data to over-com e those difficulties.From the 25 m ostly frequently usedpredicates,w e collectoverall1,523,161 positive sentences.A detailed statistic forpredicates isshow n in Table 2.
K now ledge TransferA sisexplained in the firstSection,the predicates thatresidein the long tailm ay notgetsufficienttraining data.Thisideais m otivated by pre-training and fine-tuning process (Yu etal.2010).
K now ledge transfer from sam ple-rich predicates (-K T )For sam ple-rich predicates, w e can achieve a good re-sultw ith a single predicate extractor.Extractors for differ-ent predicates are not com pletely independent from each
数据驱动与知识引导融合
• 《当知识图谱遇见深度学习》，中国人工智能通信• 统计学习， 特别是深度学习的数据红利即将耗尽• 大量先验知识未得到有效利用
17
Incorporating Complicated Rules in Deep Generative Models, under review
into deep architectures fordiscrim inative tasks such as sen-tim ent analysis and N ER .(G anchev etal.2010) proposeda fram ew ork of using posterior regularization for addingpriori linear constraints into the m odels. (K araletsos, B e-longie,and R atsch 2015)developed a generative m odelw ithsim ilarity know ledge expressed in triplet form at to learnim proved disentangled representations. C om pared to thesew orks, our w ork is novel in 1) adding priori rules intodeep generativem odels;2)accom m odating any com plicatedrules thatspecify w hatisa good resultornot.
M ethodG lobalFram ew ork
Suppose w e have a generative m odelG (z),w hich uses a la-tentvectorz asthe input,and generatesdata sam plesx.H erethe inputz ⇠ pz m ay be a noise input,oran encoding ofacondition,such as the encoded contextvector of the inputsentence in m achine translation.A fterthe training,the out-putdistribution pg ofthe generative m odelG (z)isexpectedto be sim ilarto the realdata distribution pdata.Figure 1 show s our fram ew ork of incorporating com pli-cated priorirules into the generator G .The m ostim portantpartof our fram ew ork is an additional m odel R ,w hich isreferred to as rule discrim inator.W e use R to m odel anycom plicated priorirules.There are tw o m ajorprocedures torun our fram ew ork.The firstprocedure is training the rulediscrim inatorR .Itisa binary classifierthataccepts the fakesam ple generated by G as the input,and its outputis 0 or1 indicating w hether the sam ple is in com pliance w ith therules.The second procedure is using the rule discrim inatorR to guide the data generator G .These tw o procedures areexecuted in turn,and finally the generatorw illconverge to astate generating data in com pliance w ith the rules.N ext,w eelaborate each m ajorcom ponentin ourfram ew ork.
Back propagation
Random noises or encoded contexts
RuleExaminer
If then return 1If then return 0
...
zMany fake
samples
Positive samples
Negative samples
G RP
Back propagation from golden samples
Figure 1:O urfram ew ork
R ulesD iscrim inator RA challenge is,how to letthe m achine understand the pri-orirules.Itis difficultfor m achines to understand the de-scription ofrules.A betterstrategy isdata-driven,i.e.tellingthe m achine w hatsam ples can pass the rules.In general,itis easy to w rite a sim ple program to realize the evaluation.Forexam ple,a program to testw ord-repetition is show n inA lgorithm 1.M any com plicated rules (like Table 1)can bew ritten as a sim ple program .A ctually any rules thatcan beexpressed by a program can be incorporated into deep gen-erative m odels underourfram ew ork.The data to train R com es from the generator G (z).Thegenerated data thatpass the rule exam iner are used as pos-itive sam ples and those not are used as negative sam ples.
A lgorithm 1 W ord repetition exam ination
Input: S :The inputsentence,a w ord array.O utput: A boolean value ofw hetherS hasno w ord-repetition.1: for ifrom 0 to len(S )− 1 do2: for j from i+ 1 to len(S )− 1 do3: ifS [i]= = S [j]and S [i]!= “the”then4: return False5: end if6: end for7: end for8: return True
Table 1:Exam ples ofcom plicated rulesR ules1 Sentences should end w ith‘#’(a specialcharacter).2 Subjectshould appearonce in a sentence.3 N o repeated continuous characters in the sentence.4 Length ofsentence should be m ore than 4 characters.5 The num berofhigh frequency w ords should be lessthanhalfofsentence length.
These positiveand negativesam plesare used to train therulediscrim inatorR .O urrule exam iner(a rule testprogram )canbe treated asa function P (x).P (x)is1 ifx passesthe rules,and P (x) = 0 otherw ise.The loss function to train R is asbelow :
L R = − E x⇠pg [P (x)log(R (x))+ (1− P (x))log(1− R (x))](1)
This version is based on negative log likelihood,w hich isw idely used in binary classification m odels.W e m ightalter-natively use the leastsquare loss:
L R = E x⇠pg [(R (x)− P (x))2] (2)
In the follow ing texts,w e w illfocus on the discussion w henthe loss function is negative log likelihood.C onsidering thegeneratorG ,the loss function w illbe:
L R = − E z⇠pz [P (G (z))log(R (G (z)))
+ (1 − P (G (z)))log(1 − R (G (z)))](3)
Thus,R istrained to approxim ate P .W hile the deep neu-ralnetw ork hasthe greatability to approxim ate any com pli-cated function,R isexpected to havethe sam e behaviorasPsince w e can use sufficientdata generated from G to train R .The largestdifference betw een R and P isthatR isdifferen-tiable,w hich m eans thatw e can use back-propagation (R o-jas 1996) to optim ize G by R .In this w ay,w e incorporaterules specified in P into deep generative m odels G .
U se R to Im prove GW e firstuse G A N ,one ofthe m ostpopulardeep generativem odels,to elaborate how w e train G via R .In our fram e-w ork,therole ofR issim ilarto thediscrim inatorD in gener-ative adversarialnetw orks.W hile the traditionalG A N usesD to train G ,w e use a sim ilar m ethod to train G via R .InG A N ,the lossfunctions ofD and G are:
L D = − (E x⇠pd a ta [log(D (x))]+ E z⇠pz [1− log(D (G (z)))])(4)
Demo地址：http://kw.fudan.edu.cn/ddemos/vcode/API地址：http://kw.fudan.edu.cn/apis/supervcode/
大规模常识获取与理解
• 大规模常识抽取• 从大数据的角落中寻求常识的踪迹
• 大规模常识推断
• 大规模常识判定
18
W hile the off-line solutions have attracted m uch attention,few efforts have been devoted to the on-line ones.In this pa-per,w e propose com m onsense oracles (C O ) as a realizationof on-line approach. A com m onsense oracle is a m odel oralgorithm thataccepts a com m onsense proposition as inputand returns true or false as the result. For exam ple, a m a-chine m ightneed to determ ine w hether the proposition thatpig can fly holds true.W e feed the proposition into a com -m onsense oracle.A good com m onsense oracle w ill returnfalse.C om m onsense oracle has m any advantageous over IEbased off-line solutions.First,w e can avoid the difficultiesin building a com m onsense K B ,such as schem a definition,exhaustive enum eration.C O only needs to handle the inputproposition on dem and.Second,in m any application scenar-ios,deciding w hether a given com m onsense factsatisfies ornotis enough to achieve their tasks.For exam ple,a painfulproblem in deep language m odelis thatitusually generatesentence like A pig is flying in the sky w hich is syntacticallycorrectbutsem antically a nonsense.If a C O is available,w ecan easily check the sentence is sem antically m eaningful.
In this paper,w e discuss tw o generalw ays to build a com -m onsense oracle.T he firstis a graph-based oracle by learningfrom a com m onsense know ledge base.T he oracle has a highaccuracy but w ith lim ited coverage. T he coverage of thisoracle is obviously lim ited by the capacity of the K B . Toim prove the coverage,w e further propose a w eb-based oracleby learning from search engine results.Finally,w e considerthe integration of these tw o oracles.W e realize each oracleas an binary classifier.T he key to build an effective oracleis collecting effective supporting facts.T he graph-based or-acle collect supporting facts from the paths in K B s,w hilethe w eb-based oracle collects supporting facts from searchengine results. W e use these collected facts as features tobuild the binary classifiers,w hich are further used to decidew hethera proposition holds ornot.
B uilding a com plete com m onsense base is still far fromreach,w e focus on the im portant and practicalpart,nam ely,capableO f relation. W e focus on this part because thatknow ing the capability of the real-w orld objects is theprerequisite for agents to figure out how the w orld w orks,w hich can further be used to actagainstthe environm ent.Forexam ple,w hen the autopilot driver observes a sm allobjectin the w ay,itcan beep to getitaw ay if the objectcan m ove(w alk or fly).A n agent is able to save a life if itw itnessesan anim ated object w hich cannot sw im but in w ater, orsom ething cannotfly butin air.O ther functions pow ered bycapableO f com m onsense know ledge,such as predicting afire by know ing w hich could burn,are also very helpful toour hum an beings. In this w ork, w e study the capableO fcom m onsense especially for 14 kinds of capability,w hichare:fly,sleep,die,eat,burn,w alk,grow ,run,w ork,breathe,sw im ,think,read,sing.H ow ever,itis possible to extend thism ethod to generalcapableO f reasoning.
C ontributions.T he contributions of this paper are sum m a-rized as follow s:
- W e propose a lightw eightcom m onsense oracle for on-lineproposition judgm ent;
- W eb-based features for capableO f com m onsense reason-ing,including roundaboutones,are designed and analyzed;
- E xperim ents have been conducted to exam ine the effec-tiveness of the proposed m ethod.T he experim entalresultsshow thatthe proposed com m onsense oracle can achievecom parable accuracy com pared w ith hum an-builtknow l-edge bases.
C om m onsense O racleW e propose an oracle m odelbased on the w eb as illustrated inFigure 1.T he users ask the oracle for arbitrary com m onsensefacts and gettrue or false as response.T he oracle representsthese facts by theirfeatures and learns classification m odelsfrom existing know ledge. Form ally,w e define a com m on-sense oracle as:
f :x(s,r,t) ! ±1,
w here x(s,r,t) represents the features of the triple-shapedproposition (s,r,t),1 for true,and − 1 for false.W e considerr equals to capableO f in this paper.In rest of this section,w e introduce the features derived
from know ledge graph and the w eb.Intuitively,inform ationin K G are m ore prom ising and explicit than the w ebpages,but the latter one provides opportunity to inference factsinvolving unknow n entities.
(s, r, t) ?
Commonsense Oracle
true/false
Training
The Web
KB
Application
ModelData Fetcher
Figure 1:O verview ofthe C om m onsense O racle
G raph-B ased Supporting Facts
T he paths betw een tw o given entities in know ledge graphim ply the directrelations betw een them .W e introduce graphpathsin K G asfeatures to inference these directrelations.G iven concepts and tforourtask,each path in the form of
P (s,t) = sp1−! e1
p2−! e2 ...
pn−! tcorresponds to a feature
dim ension.W e use tw o differentsets of features for this pathasfollow s,w hich are also illustrated in Figure 2.
- R elation Path.For relation paths,w e ignore allthe ei inthe path,and denote P = p1p2 ...pn as P (s,t)’s repre-sentation. Since it ignores all the routing entities, it canrepresentgeneralrelations fordifferentt’s.
- R elation-E ntity Path.For relation-entity paths,w e denoteP = p1e1p2e2 ...pn as P (s,t)’s representation. Since
Common Sense Oracle
Can pig fly？
19
Edward Feigenbaum
Knowledge is Power in AI
即将到来的是一个终点，还是一个起点？ 它会不会是一个起点，其重要性相当于人最初成为人的时候，所不同的只是人现在拥有大量新获得的工具以及在一个新的、更高的水准上的经验能力？——《时代的精神状况》
卡尔•雅斯贝斯
20
Thank YOU！
• Our LAB: Knowledge Works at Fudan University
• http://kw.fudan.edu.cn
