Simulation-based language understanding “Harry walked to the cafe.” Utterance Constructions Analysis Process General Knowledge Belief State Schema walk Trajector Harry Cafe Goal cafe Simulation Specification Simulation Simulation specification The analysis process produces a simulation specification that •includes image-schematic, motor control and conceptual structures •provides parameters for a mental simulation NTL Manifesto • Basic Concepts are Grounded in Experience – Sensory, Motor, Emotional, Social, • Abstract and Technical Concepts map by Metaphor to more Basic Concepts • Neural Computation models all levels Simulation based Language Understanding Utterance Discourse & Situational Context Constructions Analyzer: incremental, competition-based, psycholinguistically plausible Semantic Specification: image schemas, frames, action schemas Simulation Embodied Construction Grammar • Embodied representations – active perceptual and motor schemas (image schemas, x-schemas, frames, etc.) – situational and discourse context • Construction Grammar – Linguistic units relate form and meaning/function. – Both constituency and (lexical) dependencies allowed. • Constraint-based – based on feature unification (as in LFG, HPSG) – Diverse factors can flexibly interact. Embodied Construction Grammar ECG (Formalizing Cognitive Linguisitcs) 1. Linguistic Analysis 2. Computational Implementation a. Test Grammars b. Applied Projects – Question Answering 3. Map to Connectionist Models, Brain 4. Models of Grammar Acquisition ECG Structures • Schemas – image schemas, force-dynamic schemas, executing schemas, frames… • Constructions – lexical, grammatical, morphological, gestural… • Maps – metaphor, metonymy, mental space maps… • Situations (Mental Spaces) – discourse, hypothetical, counterfactual… Embodied schemas schema name schema Source-Path-Goal roles source path goal trajector role name schema Container roles interior exterior portal boundary Boundary Source Trajector Interior Portal Goal Path Exterior These are abstractions over sensorimotor experiences. ECG Schemas schema <name> subcase of <schema> evokes <schema> as <local name> roles < local role >: <role restriction> constraints <role> ↔ <role> <role> <value> <predicate> schema Hypotenuse subcase of LineSegment evokes Right-Tri as rt roles {lower-left: Point} {upper-right: Point} constraints self ↔ rt.long-side Source-Path-Goal; Container schema SPG subcase of TrajLandmark roles source: Place path: Directed–Curve goal: Place {trajector: Entity} {landmark: BoundedRegion} schema Container roles interior: Bounded-Region boundary: Curve portal: Bounded-Region Referent Descriptor Schemas schema RD roles category gender count specificty resolved Ref modifications schema RD5 // Eve roles HumanSchema Female one Known Eve Sweetser none ECG Constructions construction <name> subcase of <construction> constituents <name>:<construction> form constraints <name> before/meets <name> meaning: constraints // same as for schemas construction SpatialPP constituents prep: SpatialPreposition lm: NP form constraints prep meets lm meaning: TrajectorLandmark constraints selfm ↔ prep landmark ↔ lm.category Into and The CXNs construction Into subcase of SpatialPreposition form: WordForm constraints orth "into" meaning: SPG evokes Container as c constraints landmark ↔ c goal ↔ c.interior construction The subcase of Determiner form:WordForm constraints orth "the" meaning evokes RD as rd constraints rd.specificity “known” Two Grammatical CXNs construction DetNoun subcase of NP constituents d:Determiner n:Noun form constraints d before n meaning constraints selfm ↔ d.rd category ↔ n construction NPVP subcase of S constituents subj: NP vp: VP form constraints subj before vp meaning constraints profiled-participant ↔ subj Simulation specification The analysis process produces a simulation specification that •includes image-schematic, motor control and conceptual structures •provides parameters for a mental simulation Competition-based analyzer • An analysis is made up of: – A constructional tree – A semantic specification – A set of resolutions Johno Bryant A-GIVE-B-X subj v obj2 obj1 Ref-Exp Give Ref-Exp Ref-Exp Bill gave Mary the book @Man Give-Action Bill giver @Woman Mary recipient theme @Book book01 Combined score determines best-fit • Syntactic Fit: – Constituency relations – Combine with preferences on non-local elements – Conditioned on syntactic context • Antecedent Fit: – Ability to find referents in the context – Conditioned on syntax match, feature agreement • Semantic Fit: – Semantic bindings for frame roles – Frame roles’ fillers are scored 0Eve1walked2into3the4house5 Constructs -------------NPVP[0] (0,5) Eve[3] (0,1) ActiveSelfMotionPath [2] (1,5) WalkedVerb[57] (1,2) SpatialPP[56] (2,5) Into[174] (2,3) DetNoun[173] (3,5) The[204] (3,4) House[205] (4,5) Schema Instances ------------------SelfMotionPathEvent [1] HouseSchema[66] WalkAction[60] Person[4] SPG[58] RD[177] ~ house RD[5]~ Eve Unification chains and their fillers SelfMotionPathEvent[1].mover SPG[58].trajector WalkAction[60].walker RD[5].resolved-ref RD[5].category Filler: Person4 SelfMotionPathEvent[1] .landmark House[205].m RD[177].category SPG[58].landmark Filler:HouseSchema66 SpatialPP[56].m Into[174].m SelfMotionPathEvent[1].spg Filler: SPG58 WalkedVerb[57].m WalkAction[60].routine WalkAction[60].gait SelfMotionPathEvent[1] .motion Filler:WalkAction60 Summary: ECG • Linguistic constructions are tied to a model of simulated action and perception • Embedded in a theory of language processing – Constrains theory to be usable – Basis for models of grammar learning • Precise, computationally usable formalism – Practical computational applications, like MT and NLU – Testing of functionality, e.g. language learning • A shared theory and formalism for different cognitive mechanisms – Constructions, metaphor, mental spaces, etc. • Reduction to Connectionist and Neural levels Productive Argument Omission (Mandarin) Johno Bryant & Eva Mok 1 ma1+ma gei3 ni3 zhei4+ge • Mother (I) give you this (a toy). mother give 2PS this+CLS 2 ni3 gei3 • You give auntie [the peach]. yi2 2PS give auntie 3 ao ni3 gei3 EMP 2PS give 4 gei3 ya EMP • Oh (go on)! You give [auntie] [that]. • [I] give [you] [some peach]. give CHILDES Beijing Corpus (Tardiff, 1993; Tardiff, 1996) Arguments are omitted with different probabilities % elided (98 total utterances) 100.00% 90.00% 80.00% Giver Theme 70.00% 60.00% 50.00% Recipient 40.00% 30.00% 20.00% 10.00% 0.00% All args omitted: 30.6% No args omitted: 6.1% Analyzing ni3 gei3 yi2 (You give auntie) Two of the competing analyses: ni3 gei3 yi2 omitted ↓ ↓ ↓ ↓ Giver Transfer Recipient Theme ni3 gei3 omitted yi2 ↓ ↓ ↓ ↓ Giver Transfer Recipient Theme • Syntactic Fit: – P(Theme omitted | ditransitive cxn) = 0.65 – P(Recipient omitted | ditransitive cxn) = 0.42 (1-0.78)*(1-0.42)*0.65 = 0.08 (1-0.78)*(1-0.65)*0.42 = 0.03 Using frame and lexical information to restrict type of reference The Transfer Frame Giver Lexical Unit gei3 Recipient Giver (DNI) Theme Recipient (DNI) Theme (DNI) Manner Purpose Means Reason Place Time Can the omitted argument be recovered from context? • Antecedent Fit: ni3 gei3 yi2 omitted ↓ ↓ ↓ ↓ Giver Transfer Recipient Theme ni3 gei3 omitted yi2 ↓ ↓ ↓ ↓ Giver Transfer Recipient Theme Discourse & Situational Context child peach table mother auntie ? How good of a theme is a peach? How about an aunt? Semantic Fit: ni3 gei3 yi2 omitted ↓ ↓ ↓ ↓ Giver Transfer Recipient Theme ni3 gei3 omitted yi2 ↓ ↓ ↓ ↓ Giver Transfer Recipient Theme The Transfer Frame Giver (usually animate) Recipient (usually animate) Theme (usually inanimate) The argument omission patterns shown earlier can be covered with just ONE construction % elided (98 total utterances) 90.00% 80.00% Giver Theme 70.00% 60.00% 50.00% Recipient 40.00% 30.00% 20.00% 10.00% Subj Verb Obj1 Obj2 ↓ ↓ ↓ ↓ Giver P(omitted|cxn): 0.78 Transfer Recipient 0.42 0.00% Theme 0.65 • Each construction is annotated with probabilities of omission • Language-specific default probability can be set Leverage process to simplify representation • The processing model is complementary to the theory of grammar • By using a competition-based analysis process, we can: – Find the best-fit analysis with respect to constituency structure, context, and semantics – Eliminate the need to enumerate allowable patterns of argument omission in grammar • This is currently being applied in models of language understanding and grammar Modeling context for language understanding and learning • Linguistic structure reflects experiential structure – Discourse participants and entities – Embodied schemas: • action, perception, emotion, attention, perspective – Semantic and pragmatic relations: • spatial, social, ontological, causal • ‘Contextual bootstrapping’ for grammar learning The context model tracks accessible entities, events, and utterances Discourse & Situational Context Discourse: Discourse01 participants: Eve , Mother objects: Hands, ... discourse-history: DS01 situational-history: Wash-Action Each of the items in the context model has rich internal structure Discourse: Participants: Eve category: child gender: female name: Eve age: 2 Objects: Mother category: parent gender: female name: Eve age: 33 Situational History: Wash-Action washer: Eve washee: Hands Hands category: BodyPart part-of: Eve number: plural accessibility: accessible Discourse History: DS01 speaker: Mother addressee: Eve attentional-focus: Hands content: {"are they clean yet?"} speech-act: question Analysis produces a semantic specification Utterance “You washed them” Discourse & Situational Context World Knowledge Analysis Semantic Specification WASH-ACTION washer: Eve washee: Hands Linguistic Knowledge How Can Children Be So Good At Learning Language? • Gold’s Theorem: No superfinite class of language is identifiable in the limit from positive data only • Principles & Parameters Babies are born as blank slates but acquire language quickly (with noisy input and little correction) → Language must be innate: Universal Grammar + parameter setting But babies aren’t born as blank slates! And they do not learn language in a vacuum! Key ideas for a NT of language acquisition Nancy Chang and Eva Mok • Embodied Construction Grammar • Opulence of the Substrate – Prelinguistic children already have rich sensorimotor representations and sophisticated social knowledge • Basic Scenes – Simple clause constructions are associated directly with scenes basic to human experience (Goldberg 1995, Slobin 1985) • Verb Island Hypothesis – Children learn their earliest constructions (arguments, syntactic marking) on a verb-specific basis (Verb Island Hypothesis, Tomasello 1992) Embodiment and Grammar Learning Paradigm problem for Nature vs. Nurture The poverty of the stimulus The opulence of the substrate Intricate interplay of genetic and environmental, including social, factors. Two perspectives on grammar learning Computational models Developmental evidence • Grammatical induction – language identification – context-free grammars, unification grammars – statistical NLP (parsing, etc.) • Word learning models – semantic representations • logical forms • discrete representations • continuous representations – statistical models • Prior knowledge – – – – primitive concepts event-based knowledge social cognition lexical items • Data-driven learning – basic scenes – lexically specific patterns – usage-based learning Key assumptions for language acquisition • Significant prior conceptual/embodied knowledge – rich sensorimotor/social substrate • Incremental learning based on experience – Lexically specific constructions are learned first. • Language learning tied to language use – Acquisition interacts with comprehension, production; reflects communication and experience in world. – Statistical properties of data affect learning Analysis draws on constructions and context Form Meaning you Context Addressee Eve addressee before washed before them washer washer Wash-Action Wash-Action washee washee ContextElement Hands Discourse Segment attentionalfocus Learning updates linguistic knowledge based on input utterances Utterance Discourse & Situational Context Analysis Partial SemSpec World Knowledge Linguistic Knowledge Learning Context aids understanding: Incomplete grammars yield partial SemSpec Form you Meaning Addressee Context Eve addressee washer washed Wash-Action Wash-Action Discourse Segment washee them ContextElement Hands attentionalfocus Context bootstraps learning: new construction maps form to meaning Form Meaning you Context Addressee Eve addressee before washed before them washer washer Wash-Action Wash-Action washee washee ContextElement Hands Discourse Segment attentionalfocus Context bootstraps learning: new construction maps form to meaning Form Meaning you Addressee YOU-WASHED-THEM constituents: YOU, WASHED, THEM before washed before washer Wash-Action washee form: YOU before WASHED WASHED before THEM meaning: WASH-ACTION washer: addressee washee: ContextElement them ContextElement Grammar learning: suggesting new CxNs and reorganizing existing ones Utterance World Knowledge Discourse & Situational Context reorganize Linguistic Knowledge • • • merge join split reinforcement Analysis hypothesize • Partial SemSpec • map form to meaning learn contextual constraints Challenge: How far up to generalize Inanimate Object • Eat rice • Eat apple • Eat watermelon • Want rice • Want apple • Want chair Manipulable Objects Unmovable Objects Food Furniture Fruit apple Savory watermelon Chair rice Sofa Challenge: Omissible constituents • In Mandarin, almost anything available in context can be omitted – and often is in child-directed speech. • Intuition: • Same context, two expressions that differ by one constituent a general construction with the constituent being omissible • May require verbatim memory traces of utterances + “relevant” context When does the learning stop? Bayesian Learning Framework Gˆ argmax P(G | U , Z ) G argmax P(U | G, Z ) P(G ) G Schemas + Constructions reorganize reinforcement Analysis + Resolution Context Fitting hypothesize SemSpec • Most likely grammar given utterances and context • The grammar prior includes a preference for the “kind” of grammar • In practice, take the log and minimize cost Minimum Description Length (MDL) Intuition for MDL • S -> Give me NP • NP -> the book • NP -> a book • • • • S -> Give me NP NP -> DET book DET -> the DET -> a Suppose that the prior is inversely proportional to the size of the grammar (e.g. number of rules) It’s not worthwhile to make this generalization 51 Intuition for MDL • • • • • • • • • S -> Give me NP NP -> the book NP -> a book NP -> the pen NP -> a pen NP -> the pencil NP -> a pencil NP -> the marker NP -> a marker • • • • • • • • S -> Give me NP NP -> DET N DET -> the DET -> a N -> book N -> pen N -> pencil N -> marker Usage-based learning: comprehension and production discourse & situational context world knowledge utterance comm. intent constructicon analyze & resolve reinforcement (usage) hypothesize constructions & reorganize analysis simulation reinforcement (usage) reinforcement (correction) generate utterance reinformcent (correction) response From Molecule to Metaphor www.m2mbook.org I. Embodied Information Processing II. How the Brain Computes III. How the Mind Computes IV. Learning Concrete Words V. Learning Words for Actions VI. Abstract and Metaphorical Words VII. Understanding Stories VIII. Combining Form and Meaning IX. Embodied Language Basic Questions Addressed • How could our brain, a mass of chemical cells, produce language and thought? • How much can we know about our own experience? • How do we learn new concepts? • Does our language determine how we think? • Is language innate? • How do children learn grammar? • Why make computational brain models of thought? • Will our robots understand us? Language, Learning and Neural Modeling www.icsi.berkeley.edu/AI • Scientific Goal Understand how people learn and use language • Practical Goal Deploy systems that analyze and produce language • Approach Build models that perform cognitive tasks, respecting all experimental and experiential constraints Embodied linguistic theories with advanced biologically-based computational methods Simulation Semantics • BASIC ASSUMPTION: SAME REPRESENTATION FOR PLANNING AND SIMULATIVE INFERENCE – Evidence for common mechanisms for recognition and action (mirror neurons) in the F5 area (Rizzolatti et al (1996), Gallese 96, Boccino 2002) and from motor imagery (Jeannerod 1996) • IMPLEMENTATION: – x-schemas affect each other by enabling, disabling or modifying execution trajectories. Whenever the CONTROLLER schema makes a transition it may set, get, or modify state leading to triggering or modification of other x-schemas. State is completely distributed (a graph marking) over the network. • RESULT: INTERPRETATION IS IMAGINATIVE SIMULATION! Grammar learning: hypothesizing new constructions and reorganizing them Utterance World Knowledge Discourse & Situational Context reorganize Linguistic Knowledge • • • merge join split reinforcement Analysis hypothesize • Partial SemSpec • map form to meaning learn contextual constraints Discovering the Conceptual Primitives 2008 Cognitive Science Conference Cognitive Science is now in a position to discover the neural basis for many of the conceptual primitives underlying language and thought. The main concern is conceptual mechanisms that have neural realization that does not depend on language and culture. These concepts (the primitives) are good candidates for a catalog of potential foundations of meaning. Lisa Aziz-Zadeh, USC - Neuroscience Daniel Casasanto, Stanford – Psycholinguistics Jerome Feldman, UCB/ICSI - AI Rebecca Saxe, MIT - Development Len Talmy, Buffalo,UCB – Cognitive Linguistics Understanding an utterance in context: analysis and simulation Utterance Discourse & Situational Context World Knowledge Linguistic Knowledge Analysis Semantic Specification Simulation Neural Theory of Language (Feldman, 2006)