Question? Leave a message!




Culture & Language in Cognitive Systems

Culture & Language in Cognitive Systems
Intelligent Control and Cognitive Systems brings you... Culture Language in Cognitive Systems Joanna J. Bryson University of Bath, United KingdomOutline What is culture for (computationally) • Why are we social • Why do we communicate • Language as a special case: • Phonetics/phonology/morphology, Syntax, • Semantics, Pragmatics. Natural Language Processing (NLP) •Outline What is culture for (computationally) • Why are we social • Why do we communicate • Language as a special case: • Phonetics/phonology/morphology, Syntax, • Semantics, Pragmatics. Natural Language Processing (NLP) •SocialityWhy not be social Disease parasites. • Competition for food, shelter, mates. • Time spent maintaining social structure. •Traditional Explanation (Galton 1871, Hamilton 1973) Aggregation • as a form of cover seeking.Traditional Explanation (Galton 1871, Hamilton 1973) Aggregation • as a form of cover seeking. Isolation • increases probability of being near a predator.Why not be social Disease parasites. • Competition for food, shelter, mates. • Time spent maintaining social structure. •Traditional Explanation (Galton 1871, Hamilton 1973) Aggregation • as a form of cover seeking Aren’t • predators a form of parasiteCulture – Biological Perspective Culture: Behaviour acquired from • conspecifics by nongenetic means (Richerson Boyd 2005). Neodiffusionist hypothesis: cultural • diffusion of adaptive behaviours more likely than neutral or negative traits (Kashima 2008).Culture as Concurrency If each agent has a 1 chance of • discovering a skill (e.g. making yogurt) in its lifetime and there are 2000 agents, at any instant probably some agents will know the skill. If it is easier to learn the skill from a • knowledgeable agent than by discovery, then selective pressure for culture. Inclusive fitness c b × r • (Hamilton 1964; West et al 2007).What About Selfish Genes How can evolution select traits that help the • community but hurt the individuals Inclusive fitness kin / group selection: • What is transmitted is the replicator. • The unit of selection is the vehicle (or • interactor.) Most current vehicles are composed of • many, many replicators. (Dawkins e.g. The Extended Phenotype)Multiple Levels of Interaction Cooperation boo ha ha Replicator (Gene) Group Rah nyah boo nyah Organism Boo.Strategies for Speeding Search Concurrency • multiple searches at the same time, • only effective if solutions can be • communicated. Pruning • limit search to likely space of solutions •Culture Lets Humans Search Faster Language Built Culture Why Don’t Other Species Use ItThey DoCulture in non human primates Chimpanzees (Whiten, Goodall, McGew, Nishida, Reynolds, Sugiyama, Tutin, Wrangham, Boesch 1999, p . 684). Macaques (de Waal Johanowicz 1993); Capuchins (Perry et al 2003); Orangutans (van Schaik et al 2003).Culture in non human primates Chimpanzees (video from Whiten) Goodall, McGew, Nishida, Reynolds, Sugiyama, Tutin, Wrangham, Boesch 1999, p . 684). Macaques (de Waal Johanowicz 1993); Capuchins (Perry et al 2003); Orangutans (van Schaik et al 2003).Solitary Tortoises Use Culture if It’s Available Social Learning in a NonSocial Tortoise Anna Wilkinson, Karin Künstner Julia Müller Ludwig Huber 2010. 12 12 Left Right 10 10 8 8 6 6 4 4 2 2 0 0 Esme Emily Moses Aldous Molly Quinn Alexandra Wilhelmina Tortoise Tortoise Number of trials in which the tortoise reached the goal Number of trials in which the tortoises reached the goal Even Bacteria Share Info MGEs: e.g. Phages Plasmids One on One ‘speech’ 20 ‘Books’ Images from Bharat Kumar Chimanlal PatelHow Culture is Transmitted Intentionally versus unintentionally • By instruction or by demonstration •(Whiten et al. 2009)Ways to Transmit Culture Intentionally versus unintentionally • By instruction or by demonstration • Language and teaching • Uniquely humanHuman Uniqueness Tool use / built culture • Self concept • Moral sensibility • Culture • Teaching • Language •Human Uniqueness Tool use / built culture • Self concept • Moral sensibility • Culture • Teaching • Language • √Outline What is culture for (computationally) • Why are we social • Why do we communicate • Language as a special case: • Phonetics/phonology/morphology, Syntax, • Semantics, Pragmatics. Natural Language Processing (NLP) •Language: Not Just Communication Phonetics/phonology/morphology: what • words (or subwords) are we dealing with Syntax: What phrases are we dealing with • Which words modify one another Semantics: What’s the literal meaning • Pragmatics: What should you conclude • from what was said How should you actPhonetics / phonology / morphology Understanding a speech (or character) • stream requires decomposing it into the units that have meaning: segmentation. Phonemes are relatively discrete (though • they can be merged in transitions.) Infants babble all() initially then settle on • the ones they hear / in their language.Segmentation Objects in a scene. • Gestures in a video. • Words in speech. • Actions in sequence. • Junqing Chen and Thrasyvoulos Pappas Very, very hard in all domains; better with multiple information sources.Segmentation Objects in a scene. • Gestures in a video. • Words in speech. • HARD Actions in sequence. • Very, very hard in all domains; better with multiple information sources.Speech Recognition Language model speech model noise model http://www.learnartificialneuralnetworks.com/speechrecognition.htmlLots of Machine Learning / Pattern Rec Decision regions formed by a 2layer perceptron using backpropagation training and vowel formant data. (From Huang Lippmann, 1988.)Language Outline Phonetics/phonology/morphology: what • words (or subwords) are we dealing with Syntax: What phrases are we dealing with • Which words modify one another Semantics: What’s the literal meaning • Pragmatics: What should you conclude • from what was said How should you actSyntaxA Brief History of AI Founded in the 1950s. • Funded in the 1960s by promising machine • translation (esp. Russian). Theory: Solve syntax as a program, lookup semantics in dictionary. By 1980s, funders restless. Theory: • Semantics requires grounding in an embodied system (Harnad 1990, Brooks 1991). 1990s: Robots for Language. •What AI Thought Language Was Phonetics/phonology/morphology: what • words (or subwords) are we dealing with Syntax: What phrases are we dealing with • Which words modify one another Semantics: What’s the literal meaning • Pragmatics: What should you conclude • from what was said How should you actThe Plan For Translation Build something that parses and generates • individual language syntax. Automatically morph sentences between • languages’ syntaxes. Use dictionaries to look up replacement • words (semantics). Warning: almost totally doesn’t workSyntax: Chomsky’s Grammar(s) Vocabulary: S →NP + VP terminal symbols • NP →N D + NP ADJ + N PN closed classes • VP →IV AUX + VP TV + NP • IV →laughed cried ... • AUX→can will shall ... • TV→throw catch ... • N→dog peacock justice ... • D→the a an • PN→ he she they ... • English e.g. SVO vs SOVWhat to Do With a Grammar: Parse Use it to parse a sentence. • Ambiguous sentences have multiple parse • trees. Ambiguity can came from multiple • definitions (remember, plug in semantics last – often FOPL). Other words or context may resolve. • The farmer pulls the cow on the barn.What to Do With a Grammar: Generate Use it to generate a sentence. • Associate a probability with every option. • Throw dice. • Automatic language •Example S →NP + VP • NP →N D + NP ADJ + N PN • VP →IV AUX + VP TV + NP • IV →laughed cried ... • AUX→can will shall ... • Dog TV→threw caught ... • will N→dog peacock justice ... • D→the a an catch • an peacock .Is Language Uniquely Human Tool use / built culture • Self concept • Moral sensibility • Culture • Teaching • Language • √Compositionality / Recursion S →NP + VP • NP →N D + NP ADJ + N PN • VP →IV AUX + VP TV + NP • IV →laughed cried ... • Allows language to be AUX→can will shall ... • infinitely productive. TV→threw caught ... • N→dog peacock justice ... • D→the a an What no animal language • learner has shown. (c.f. Hauser, Chomsky Fitch 2002)Chomsky on Cognition Language is for computation / thought, not • communication. Grammars can tell you the limits of human • intelligence (e.g. CFG)Chomsky’s Universal Grammar Hypothesis: every human is born with the • universal grammar capacity. Learns to set parameters from listening • (know this is true of phonemes). Evidence: Poverty of the stimulus – children • don’t hear enough negative examples to learn language from scratch.Critiques of Universal Grammar You can learn a stochastic grammar model • without many negative examples (Chomsky assumed a deterministic one, Chater Manning, 2006). Many characteristics of the UG evolve in the • language naturally in simulation – necessary characteristics of something learnable (Kirby 1999). Dual replicator theory: Culture biology both evolve at the same time under each other’s influence.Language Outline Phonetics/phonology/morphology: what • words (or subwords) are we dealing with Syntax: What phrases are we dealing with • Which words modify one another Semantics: What’s the literal meaning • Pragmatics: What should you conclude • from what was said How should you actPragmaticsPragmatics What you really mean– requires context. • Much elaborate work on reference. e.g. • “They thought I was going to town but that wasn’t what I meant.” Still doesn’t get you to “uh”→ /no don’t go • in there keep going straight/ (Agre Chapman 1988).Semantics and GroundingA Brief History of AI Founded in the 1950s. • Funded in the 1960s by promising machine • translation (esp. Russian). Theory: Solve syntax as a program, lookup semantics in dictionary. By 1980s, funders restless. Theory: • Semantics requires grounding in an embodied system (Harnad 1990, Brooks 1991). 1990s(–now): Robots for Language. •Embodiment Hypothesis: NLP has failed so far because • semantics isn’t grounded in humanlike experience. E.g. life career are understood via a • metaphor to path which you learn about the hard way in your first few years. (Lakoff Johnson 1999) Funding argument for humanoid robotics. • Not much positive evidence. •Syst em sof t ware ( 0 t h) Syst em sof t ware ( commercial processor) Periperhal Mot ion Vergence Ullmanesque Physical schema Saccades based st ereo visual rout ines based obj. recog. “Building VOR Smoot h pursuit Face popout s Face remembering Face recognit ion Brains Head/ body/ eye/ coord Head/ eye coord Gest ure recognit ion Facial gest ure recog. Body mot ion recog. for Own hand t racking Specif ic obj. recog. Generic object recog. Bodies”, Bring hands Hand Grasping, Bodybased met aphors midline linking t ransfer Brooks DOF reduct ion DOF reduct ion Bat t ing st at ic Stein (specific coords) (generic coords) object s Body st abilit y, Body+arm reaching Body mimicry (1993), leaning, rest ing Manipulat ion t urn t aking MIT AI Sound localizat ion Soundbased manip. Voice/ face assoc lab tech Sound/ mot ion correl Human voice ext ract ion Prot o language report Tone ident ificat ionVoice t urn t aking Visual imagery Symbolizat ion 1439. Ment al rehearsal Imaginat ion Mult ipledraf t s emergence Sept 1 Sept 1 Sept 1 Sept 1 Sept 1 1993 1994 1995 1996 1997Alternative: Large Corpus Linguistics Do pattern recognition across many texts. • The more one word is used like another • word, the more they mean the same thing. Mathematically related to the way web • pages are indexed (Lowe 2001). Wikipedia: Latent Semantic AnalysisLarge Corpus Linguistics Human semantics can be replicated by statistical • learning on large corpra (Finch 1993, Landauer Dumais 1997, McDonald Lowe 1998). Only information gathered on each word’s • ‘meaning’ is what words occur in a small window before and after it. Normally just choose 75 fairly frequent words to • watch out for.Data to Be Matched Semantic Priming – reaction times showing • how similar people consider words’ meanings to be. How quickly you are able to tell that a • collection of letters is a real word is dependent on how similar the word’s meaning is to words / concepts you have recently been exposed to. salt circle gold month measles Semantic Priming silver square Replication, year lightning visualised with a 2D projection star sister cabbage latin (Lowe 1998). Analysis for lettuce dog comparison to greek queen cat human data soldier uses similarity thunder measured using black moon sailor 75D cosines. white mumps king brotherEvolution of moral agency terms (Bilovich Bryson 2008) terms from the implicit bias task (Banaji Greenwald 1994) text: BibleBilovich 2006 text: ShakespeareBilovich 2006 text: British National Corpus (contemporary word use)Humanlike Biases in Corpus Semantics Bilovich I did not replicate Banaji (2003). • Nearest miss was Shakespeare – (nearly) • single author Macfarlane I (in prep.) have found matches. •Macfarlane (2013) Results Life terms more like pleasant Death terms • more like unpleasant words. Elderly Youth did not go as per Banaji on • pleasantness, though did on competence. Male terms more like Career Female terms • more like Family. In preparation; also University of Bath Computer Science technical report.Traditional Theory of Semantics justice justice chair chair table table run run phobia phobia i ii iii Ontology e.g. Deacon (1997) The Symbolic SpeciesCorpus Semantics Allows... justice justice chair chair table table run run phobia phobia i ii Ontology more plausibleOutline What is culture for (computationally) • Why are we social • Why do we communicate • Language as a special case: • Phonetics/phonology/morphology, Syntax, • Semantics, Pragmatics. Natural Language Processing (NLP) •What AI Used to Think Language Was Phonetics/phonology/morphology: what • words (or subwords) are we dealing with Syntax: What phrases are we dealing with • Which words modify one another Semantics: What’s the literal meaning • Pragmatics: What should you conclude • from what was said How should you actNgrams Large corpus technique for both language • generation and speech recognition. Given previous N words, what is a probable • following term Memorise a sliding window through text. Recognition: disambiguates parses. • Generation: just press go. • http://johno.jsmf.net/knowhow/ngrams/Speech Recognition Ngram Language model speech model noise model http://www.learnartificialneuralnetworks.com/speechrecognition.htmlaccepted to the World Multiconference on Systemics, Cybernetics and Informatics, 1995. http://pdos.csail.mit.edu/scigen/Note: probably more about a) reviewing b) “academic” incentives esp. in China than NLP.Generally, Still Need ‘Real’ Natural Language Processing (NLP) Negation. • Referents for “this” and “that”. • Recognising multiple meanings for single • word. Motivation, meaning tracking, turn taking. • Cognitive SystemsJeopardy vs Watson IBM April, 2011 Videos via Bath graduate, Dale Lane(Ferrucci et al., AI Magazine 2010)Summary Culture is a powerful process for sharing • intelligence / the output of cognition. Language is particularly effective at that. • NLP is hard, but getting there. • AI can use our culture / exploit our cognition. ⟹ cf. ethics consciousness lectures. •Reminder: NLP in Games Template matching. • Mentioned in Believability lecture: play • with Eliza as homework (Mx doctor on emacs) Dialog in narrative context (story telling). • paper in AAAI 2013
Website URL
Comment