Common Sense Computing: from the Society of Mind to Digital Intuition and Beyond

Common Sense Computing: From the Society of Mind to Digital Intuition and beyond

ErikCambria1,Amir Hussain1, Catherine Havasi2, andChrisEckl3

1 Dept. of Computing Science and Maths, University of Stirling, Scotland, UK 2 MIT Media Lab, MIT, Massachusetts, USA 3 Sitekit Labs, Sitekit Solutions Ltd, Scotland, UK {eca,ahu}@cs.stir.ac.uk,[email protected],[email protected] http://cs.stir.ac.uk/~eca/commonsense

Abstract. What is Common Sense Computing? And why is it so important for the technological evolution of humankind? This paper presents an overview of past, present and future eﬀorts of the AI community to give computers the capacity for Common Sense reasoning, from Minsky’s Society of Mind to Media Laboratory’s Digital Intuition theory, and beyond. Is it actually possible to build a machine with Common Sense or is it just an utopia? This is the question this paper is trying to answer. Keywords: AI, Semantic networks, NLP, Knowledge base management.

1 Introduction

What magical trick makes us intelligent?-Marvin Minskywas wonderingmore than two decades ago-The trick isthatthereis no trick.The power of intelligence stems from our vast diversity, notfromany single, perfect principle [1]. Human brain in fact isavery complex system, maybe the most complexinna- ture.The functions itperforms are the product ofthousands andthousands of diﬀerentsubsystems workingtogether at the same time.Such a perfect system is very hard to emulate:nowadays in fact there are plentyofexpert systems around but none ofthemisactually intelligent, theyjust havetheveneer of intelligence. The aim ofCommon Sense Computing isto teach computers the things weall know about the worldandgivethemthecapacity forreasoning on these things.

2TheImportanceofCommonSense

Communication is one ofthemost importantaspectsofhumanlife. Communi- catinghasalwaysacost in terms ofenergy andtime, since information needs to be encoded, transmitted anddecoded.This is why people,when communi- cating with each other, provide just the usefulinformation andtaketherestfor granted.This ‘taken forgranted’information is what wecall Common Sense: obvious things people normally know andusually leaveunstated.

J. Fierrez et al. (Eds.): BioID MultiComm2009, LNCS 5707, pp. 252–259, 2009. c Springer-Verlag Berlin Heidelberg 2009 ! Common Sense Computing 253

Wearenottalkingabout the kind ofknowledge youcan ﬁnd in Wikipedia but all the basicrelationships among words, concepts, phrases andthoughts that allow peopleto communicate with each other andfaceeverydaylife problems. It’sakind ofknowledge that sounds obvious and natural to us but it isactually daedal andmultifaceted: the illusion ofsimplicity comes fromthefactthat, as each new group ofskillsmatures,webuildmore layers on top ofthemandtend to forget about the previous layers. Today computers lack thiskind ofknowledge.They doonlywhat they are programmed to do: theyonlyhave one way to dealwith a problem and,if somethinggoes wrong, they get stuck.This is whynowadays wehaveprograms that exceed the capabilities of worldexperts but are not oneableto dowhat a three years oldchildcan do. Machines canonlydological things, but meaning isanintuitiveprocess:itcan’tbereducedto zerosand ones. To helpuswork, computers must get to know what our jobs are.Toentertain us, theyneed to know what we like.Totake care ofus, they haveto know how wefeel. To understandus, they must thinkaswethink.We need to transmit computers our Common Sense knowledge oftheworldbecausesoon there won’t be enough humanworkers to perform the necessary tasks for our rapidly aging population. To face this AI emergency, we’ll haveto givethemphysical knowledge ofhow objects behave, social knowledge ofhow people interact, sensory knowledge ofhow things lookandtaste, psychological knowledge about the way peoplethink, andsoon.But havingadatabaseofmillions ofCommon Sense facts will notbeenough:we’ll also haveto teach computers how to handlethis knowledge, retrieve it whennecessary, learn fromexperience,ina word we’ll haveto givethemthecapacity forCommon Sense reasoning.

3TheBirthofCommonSenseComputing

It’s noteasy to state when exactly Common Sense Computing was born. Before Minsky many AI researchers started to thinkabout the implementation ofa Common Sense reasoningmachine: the very first one was maybe AlanTuring when, in 1950, he first raised the question “can machines think?”. But he never managed to answer that question, he just provided a methodto gauge artificial intelligence: the famous Turingtest.

3.1 The Advice Taker The notion ofCommon Sense in AI isactually dated 1958,whenJohnMcCarthy proposed the ‘advice taker’[2], aprogram meantto try to automatically deduce for itselfasufficiently wide class of immediate consequences ofanything it was toldand what italready knew. McCarthy stressed the importance offindinga proper method ofrepresentingexpressions in the computer anddeveloped the idea ofcreatingapropertylist foreachobject in which are listed the specific things peopleusually know about it.It was the first attempt to buildaCommon Sense knowledge base but, more important,it was the epiphany oftheneed of Common Sense to moveforward in the technological evolution. 254 E. Cambria et al.

3.2 The Society of Mind Theory of Human Cognition While McCarthywas more concerned with establishing logical andmathemat- ical foundationsforCommon Sense reasoning,Minskywas more involved with theories ofhow weactually reason usingpattern recognition andanalogy. These theories were organized in 1986with the publication of The Societyof Mind, a masterpiece of AI literature containinganilluminating vision ofhow the human brain works. Minsky sees the mindmadeofmany littlepartscalled ‘agents’, each mindless byitselfbutabletolead to true intelligence whenworkingtogether.These groups ofagents, called ‘agencies’, are responsibleto perform some type offunction, such as remembering, comparing, generalizing, exemplifying, analogizing, simplifying, predicting, andsoon.The most common agents are the so called ‘K-lines’whose task issimply to activate other agents: this isaveryimportant issue since agents are all highly interconnected andactivatingaK-linecan cause a significant cascade ofeffects.ToMinsky, in fact, mental activity ultimately consists in turning individual agents on and off: at any time only some agents are active andtheircombined activity constitutes the ‘total state’ofthemind. K-lines are a very simplebutpowerful mechanism since they allow enteringa particular configuration ofagents that formed a useful societyinapastsituation: this ishow webuildandretrieve our problem solvingstrategies in our mind, this ishow weshoulddevelop our problem solvingstrategies in our programs.

4TowardsProgramswithCommonSense

Minsky’stheorywas welcomed with great enthusiasm by the AI community andgavebirth to many attempts to buildCommon Sense knowledge bases and exploitthemto develop intelligentsystems e.g. Cycand WordNet. 4.1 The Cyc Project Cyc [3] is one ofthefirstattemptsto assembleamassiveknowledge base spanning human Common Sense knowledge.Initially started byDoug Lenat in 1984, this project utilizes knowledge engineers who handcraft assertionsandplace them into a logical framework usingCycL, Cyc’sproprietarylanguage. Cyc’sknowledge isrepresented redundantly at two levels: aframelanguage distinction (epistemologicallevel), adopted for its efficiency, andapredicate cal- culus representation (heuristic level), needed for its expressivepower to represent constraints.While the first level keeps a copyofthefactsin the uniform user language, the second level keeps its own copyindifferent languages anddata structures suitableto be manipulated by specialized inference engines.Knowl- edge in Cyc isalsoorganized into‘microtheories’, resembling Minsky’sagencies, each one with its own knowledge representation scheme andsetsofassumptions. 4.2 WordNet Begunin1985 at Princeton University, WordNet [4]isadatabaseof words, primarily nouns,verbs andadjectives.Ithasbeenone ofthemost widely used Common Sense Computing 255 resources in computationallinguistics andtextanalysisfortheeasein interfacing it with any kind ofapplication andsystem. The smallest unit in WordNet istheword/sense pair,identified by a ‘sense key’. Word/sense pairs are linked by asmall set ofsemanticrelationssuchas synonyms, antonyms,is-asuperclasses, and words connected byother relations such as part-of. Each synonymset,inparticular,iscalled synset:itconsists in the representation ofaconcept,often explained through a brief gloss, andrep- resents the basicbuildingblock forhierarchies and other conceptual structures in WordNet.

5FromLogicBasedtoCommonSenseReasoning

Using logic-based reasoningcan solvesome problems in computer programming. However, most real-worldproblems need methods better at makingdecisions based on previous experience with examples,orby generalizingfromtypes of explanationsthathave worked well on similar problems in the past. In building intelligentsystems wehaveto try to reproduce our wayofthinking: weturnideas around in our mindto examinethemfromdifferentperspectives until wefind onethatworks forus.Since computers appeared,our approach to solveaproblem has alwaysconsisted in first lookingforthebestway to represent the problem, thenlookingforthebestway to representtheknowledge needed to solve itandfinally lookingfor the best procedure forsolving it. Thisproblem-solvingapproach isgood whenwehaveto dealwith a specific problem but there’ssomethingbasically wrong with it:it lead us towrite only specialized programs that cope with solving only that kind ofproblem.

5.1 The Open Mind Common Sense Project Initially born fromanidea of David Stork, the OpenMindCommon Sense (OMCS) project [5] isakind ofsecond-generation Common Sense database: knowledge isrepresented in naturallanguage, rather than usingaformallogical structure, and information is nothandcrafted by expert engineers but sponta- neously inserted byonline volunteers. The reason whyLenat decided to developan ad-hoc language forCyc is that vagueness andambiguity pervade English andcomputer reasoningsystems generally require knowledge to be expressed accurately andprecisely. But, as expressed in the Societyof Mind, ambiguityisunavoidable when tryingto rep- resenttheCommon Sense world.Nosingleargument in fact isalwayscompletely reliable, but if wecombinemultipletypes ofarguments wecanimprovethero- bustness ofreasoningaswell as wecanimproveatablestability by providing it with many small legs in place of just one very big leg.This wayinformation is not only more reliablebutalso stronger:ifapiece of information goes lost we can still access the wholemeaning, exactly as the tablekeepson standingupif wecutout one ofthesmall legs.Diversityis in fact the keyofOMCS project success: the problem is notchoosingarepresentation in spite ofanother but it’s ﬁndingaway forthemtowork together in onesystem. 256 E. Cambria et al.

5.2 Acquiring Common Sense by Analogy In 2003 Timothy Chklovskiintroduced the cumulativeanalogy method [6]: a class ofanalogy-based reasoningalgorithms that leverage existingknowledge to pose knowledge acquisition questionsto the volunteer contributors. Chklovski’s Learner first determines what other topics are similar to the topic the user iscurrently insertingknowledge for, thenitusescumulativeanalogy to generate andpresent new specific questionsabout thistopic.Because each statementconsists ofanobject andaproperty, the entire knowledge repository can be visualized as a large matrix, with every known object ofsome statement beingarow andevery known property beingacolumn. Cumulativeanalogyisperformed by first selectingasetof nearest neighbors, in terms ofsimilarity, ofthetreatedconcept andthen by projectingknown properties ofthissetontonotknown properties oftheconcept andpresenting them as questions.The replies to the knowledge acquisition questionsformulated by analogy are immediately added to the knowledge repository, affectingthe similarity calculations.

5.3 ConceptNet In 2004 HugoLiuand Push Singh, reﬁned the ideas oftheOMCS project in Con- ceptNet [7], asemanticresource structurally similar toWordNet, but whose scope ofcontents isgeneralworldknowledge in the same vein as Cyc.While WordNet is optimised for lexical categorisation andCyc is optimised forformalised logical reasoning, ConceptNet is optimised formakingpractical context-based inferences over real-worldtexts. In ConceptNet,WordNet’s notion of node in the semantic network isextended frompurely lexicalitems (words andsimplephraseswith atomicmeaning) to include higher-order compoundconcepts, e.g.‘satisfy hunger’or ‘follow recipe’, to representknowledge aroundagreaterrange ofconcepts found in everyday life.Most ofthefactsinterrelatingConceptNet’ssemantic network in fact are dedicated to makingrathergenericconnectionsbetween concepts. Thistype ofknowledge can be brought back toMinsky’s K-lines as it increases the connectivityofthesemantic network andmakesitmore likely that concepts parsed out ofatextdocumentcan be mapped into ConceptNet.InConcept- Net version 2.0 a new system for weightingknowledge was implemented,which scores each binary assertion based on how many times it was uttered in the OMCS corpus, and on how well itcan be inferred indirectly from other facts.In ConceptNet version 3.0 [8] users can also participate in the process ofreﬁning knowledge by evaluatingexistingstatements.

5.4 Digital Intuition The best way to solveaproblem isto already know asolution for itbutif we haveto face a problem wehave never met before we need to use our ‘intuition’. Intuition can be explained as the process ofmakinganalogies between the current problem andtheones solved in the past to findasuitablesolution. Minsky Common Sense Computing 257 attributes thisproperty to the so called ‘difference-engines’, aparticular kind ofagents which recognize differences between the currentstateandthedesired state andactto reduce each difference byinvoking K-lines that turnonsuitable solution methods. To emulate this ‘reasoningby analogy’ weuseAnalogySpace [9], aprocess which generalizes Chklovski’scumulativeanalogy. In thisprocess, ConceptNet isfirstmappedinto asparsematrix andthen truncated Singular Value Decom- position (SVD) isapplied over itto reduce its dimensionality andcapturethe most importantcorrelations.The entries in the resultingmatrix are numbers representingthereliabilityoftheassertionsandtheirmagnitude increases log- arithmically with the confidence score.Applying SVD on thismatrix causes it to describe other features that couldapply to known concepts by analogy: ifa concept in the matrix has no value specified forafeatureowned by many similar concepts, then by analogy the concept is likely to havethatfeatureaswell. Thisprocess is naturally extended by the ‘blending’ technique [10], a new methodto perform inference over multiplesources ofdatasimultaneously, taking advantage oftheoverlap between them.Thisenables Common Sense to be used as a basisfor inference in a wide varietyofsystems andapplicationsso that they can achieve DigitalIntuition about their own data.

6ApplicationsofCommonSenseComputing

Weareinvolved in an EPSRC project whose main aim isto further developand apply the above-mentioned technologies in the ﬁeld ofCommon Sense Comput- ingto buildanovelintelligentsoftware enginethatcan auto-categorise docu- ments, andhence enablethedevelopment offuturesemantic web applications whose design andcontentcan dynamically adapt to the user.

6.1 Enhancing the Knowledge Base The key to perform Common Sense reasoning isto findagoodtrade-off for representingknowledge: since in life no two situationsareever the same,norepresentation shouldbetoo concrete,or it will notapply tonew situations, but, at the same time,norepresentation shouldbetoo abstract,or it will suppress too many details. ConceptNet already supports differentrepresentationsby main- tainingdifferent ways ofconveyingthesameidea with redundantconcepts, but weplan to enhance thismultiplerepresentation by connectingConceptNet with dereferenceable URIsandRDF statements to enlarge the Common Sense knowledge base on adifferent level. Wealso plan toimproveConceptNet by givinga geospatial reference to all those pieces ofknowledge that are likely to have one andhence make itsuitableto be exploited by geographic oriented applications. The ‘knowledge retrieval’ is one ofthemain strengths ofConceptNet:weplan to improve itby developinggamesto train the semantic network andby pointing on socialnetworking. Wealso plan toimproveConceptNet on the levelof what Minsky calls ‘self- reflection’ i.e.onthe capabilityofreflectingabout its internal structure and 258 E. Cambria et al. cognitiveprocesses. ConceptNet in fact currently focuses on the kinds ofknowledge that might populate the A-brain ofaSocietyof Mind:itknowsagreatdeal about the kinds of objects, events and other entities that exist in the external world, but itknowsfarless about how tolearn, reason andreflect. Weplan to givethesemantic network the ability to self-check its consistency e.g. bylookingfor words that appear together in waysthatareimplausible statistically andaskusersfor verification orkeepingtraceofsuccessfullearning strategies.The system isalsolikely to remember attempts that led to particularly negativeconclusions in order to avoidunproductivestrategies in the future.To thisend weplan toimprovethe‘negativeexpertise’ofthesemantic network, which is now just partially implemented by askinguserstojudge inferences,in order to givethesystem the capability tolearn from its mistakes.

6.2 Understanding the Knowledge Base

Whenever wetry to solveaproblem,wecontinuously andalmost instantly switch to differentpoints of viewinour mindto findasolution. Minsky argues that our brainsmay use special machinery, that he calls ‘panalogy’ (parallel analogy), that links correspondingaspectsofeachview to the same ‘slot’ina larger-scale structure that issharedacross several differentrealms. ConceptNet’scurrentknowledge representation doesn’tallow us to think about how toimplementsuchastrategyyet, but,once wemanage to give ConceptNet a multiplerepresentation as planned,we’ll haveto start thinking about it.Atthemoment SVD-based methods are used on the graph structure ofConceptNet to buildthevectorspacerepresentation oftheCommon Sense knowledge.Principal Component Analysis (PCA) isanoptimalway to project data in the mean-square sense but the eigenvalue decomposition ofthedataco- variance matrix is very expensiveto compute.Therefore weplan to explore new methods such as Sparse PCA, atechnique consisting in formulating PCA as a regression-type optimization problem, andRandom Projection, amethod less reliablethanPCA but computationally cheaper.

6.3 Exploiting the Knowledge Base

Our primary goalisto buildanintelligentauto-categorization tool which uses Common Sense Computing, together with statistical methods, to make the doc- umentclassification more accurate andreliable.Wealso plan to apply Common Sense Computingforthedevelopment ofemotion-sensitivesystems.Byanalysing users’Facebookpersonal messages, emails, blogs, etc., the engine will be able to extract users’ emotionsandattitudes andusethis information to be ableto better interact with them. In the sphere ofe-games, forexample, such an enginecouldbeemployed to implementconversational agents capableto react to user’sframeofmindchanges andthusenhance players’levelof immersion. In the field ofenterprise 2.0 the enginecouldbeusedto developcustomer care applicationscapableto measure users’levelofsatisfaction. In the field ofe-health, finally, PDA-microblogging Common Sense Computing 259 techniques couldbeapplied to seize clinicalinformation frompatients by study- ingtheirpieces ofchat, tweets or SMS, andan e-psychologist couldbedeveloped to provide helpfor light psychological problems.

7Conclusions

It ishardto measure the total extent ofaperson’sCommon Sense knowledge, but a machinethatdoes humanlike reasoningmight only need a few dozen millions of items ofknowledge.Thus we wouldbetemptedto giveapositive answer to the question “is itactually possibleto buildamachine with Common Sense?”. But, as wesawinthispaper, Common Sense Computing is not just about collectingCommon Sense knowledge:it’sabout how werepresent itand how weuseitto make inferences. Wemadevery goodprogress in performingthissince McCarthy’s ‘advice taker’ but weareactually still scratchingthesurfaceofhumanintelligence.So wecan’tgiveaconcrete answer to that question, not yet. The road to the creation ofamachine with the capacityofCommon Sense reasoning isstill longandtough but wefeel that the path undertaken so far is agood one.And, evenif wefail in makingmachines intelligent,webelieve we’ll be ableto at least teach them whoweareandthusmakethemableto better contribute to the technological evolution ofhuman kind.

References

1. Minsky, M.: The Society of Mind. Simon and Schuster (1986) 2. McCarthy, J.: Programs with Common Sense. In: Proceedings of the Teddington Conference on the Mechanization of Thought Processes (1959) 3. Lenat, D.: Cyc: toward programs with common sense. ACM Press, New York (1990) 4. Fellbaum, C.: WordNet: An Electronic Lexical Database. MIT Press, Cambridge (1998) 5. Singh, P.: The Open Mind Common Sense Project. KurzweilAI.net (2002) 6. Chklovski, T.: Learner: a system for acquiring commonsense knowledge by analogy. K-CAP (2003) 7. Liu, H., Singh, P.: ConceptNet: a practical commonsense reasoning toolkit. BT Technology Journal (2004) 8. Havasi, C., Speer, R., Alonso, J.: ConceptNet 3: a Flexible, Multilingual Semantic Network for Common Sense Knowledge. RANLP (2007) 9. Speer, R., Havasi, C., Lieberman, H.: AnalogySpace: Reducing the Dimensionality of Common Sense Knowledge. AAAI (2008) 10. Havasi, C., Speer, R., Pustejovsky, J., Lieberman, H.: Digital Intuition: Applying Common Sense Using Dimensionality Reduction. IEEE Intelligent Systems (2009)