Embodied Construction Grammar in language (acquisition and) use Jerome Feldman (jfeldman@icsi.berkeley.edu) Computer Science Division, University of California, Berkeley, and International Computer Science Institute State of the Art • Limited Commercial Speech Applications transcription, simple response systems • Statistical NLP for Restricted Tasks tagging, parsing, information retrieval • Template-based Understanding programs expensive, brittle, inflexible, unnatural • Essentially no NLU in HCI, QA Systems What does language do? A sentence can evoke an imagined scene and resulting inferences: “Harry walked to the cafe.” CAFE – Goal of action = at cafe – Source = away from cafe – cafe = point-like location “Harry walked into the cafe.” CAFE – Goal of action = inside cafe – Source = outside cafe – cafe = containing location Language understanding (Utterance, Situation) Conceptual knowledge Linguistic knowledge Analysis Interpretation Language understanding: analysis & simulation “Harry walked to the cafe.” Utterance Lexicon Constructicon General Knowledge Belief State Analysis Process Schema walk Trajector Harry Cafe Goal cafe Semantic Specification Simulation Interpretation: x-schema simulation Constructions can • specify which schemas and entities are involved in an event, and how they are related • profile particular stages of an event • set parameters of an event walker at goal energy goal=home walker=Harry Harry is walking home. Traditional Levels of Analysis Pragmatics Semantics Syntax Morphology Phonetics “Harry walked into the cafe.” Pragmatics Semantics Utterance Syntax Morphology Phonetics Construction Grammar A construction is a form-meaning pair whose properties may not be strictly predictable from other constructions. (Construction Grammar, Goldberg 1995) Form Meaning block walk to Source Trajector Path Goal Form-meaning mappings for language Linguistic knowledge consists of form-meaning mappings: Form Meaning phonological cues word order intonation inflection event structure sensorimotor control attention/perspective social goals... Cafe Constructions as maps between relations Complex constructions are mappings between relations in form and relations in meaning. Form Mover + Motion before(Mover, Motion) “is” + Action + “ing” before(“is”, Action) suffix(Action, “ing”) Mover + Motion + Direction before(Motion, Direction) before(Mover, Motion) Meaning MotionEvent mover(Motion, Mover) ProgressiveAction aspect(Action, ongoing) DirectedMotionEvent direction(Motion, Direction) mover(Motion, Mover) Embodied Construction Grammar (Bergen and Chang 2002) • Embodied representations – active perceptual and motor schemas – situational and discourse context • Construction Grammar – Linguistic units relate form and meaning/function. – Both constituency and (lexical) dependencies allowed. • Constraint-based (Unification) – based on feature structures (as in HPSG) – Diverse factors can flexibly interact. Representing image 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. Inference and Conceptual Schemas • Hypothesis: – Linguistic input is converted into a mental simulation based on bodily-grounded structures. • Components: – Semantic schemas • image schemas and executing schemas are abstractions over neurally grounded perceptual and motor representations – Linguistic units • lexical and phrasal construction representations invoke schemas, in part through metaphor • Inference links these structures and provides parameters for a simulation engine Early Example Understanding News Stories France fell into recession. Pulled out by Germany In1991, India set out on a path of liberalization. The Government started to loosen its stranglehold on business and removed obstacles to international trade. Now the Government is stumbling in implementing the liberalization plan. Task • Interpret simple discourse fragments/blurbs – France fell into recession. Pulled out by Germany – Economy moving at the pace of a Clinton jog. – US Economy on the verge of falling back into recession after moving forward on an anemic recovery. – Indian Government stumbling in implementing Liberalization plan. – Moving forward on all fronts, we are going to be ongoing and relentless as we tighten the net of justice. – The Government is taking bold new steps. We are loosening the stranglehold on business, slashing tariffs and removing obstacles to international trade. I/O as Feature Structures • Indian Government stumbling in implementing liberalization plan Language understanding: analysis & simulation construction WALKED form selff.phon [wakt] meaning : Walk-Action constraints selfm.time before Context.speech-time selfm..aspect encapsulated “Harry walked into the cafe.” Utterance Analysis Process Constructions General Knowledge Semantic Specification Belief State CAFE Simulation Embodied Construction Grammar provides formal tools for linguistic description and analysis motivated largely by cognitive/functional concerns. • Allows precise specifications of structures/processes involved in acquisition of early constructions –Embodied constructions (structured maps between form and meaning); lexically specific and more general –Usage-based processes of learning new constructions to account for co-occurring utterance-situation pairs • Bridge to detailed psycholinguistic and neural imaging experiments Formal Cognitive Linguistics • Schemas and frames – Image schemas, force dynamics, executing schemas… • Constructions – Lexical, grammatical, morphological, gestural… • Maps – Metaphor, metonymy, mental space maps… • Mental spaces – Discourse, hypothetical, counterfactual… Embodied constructions Form Meaning construction HARRY form : [hEriy] meaning : Harry Harry cafe Notation CAFE construction CAFE form : [khaefej] meaning : Cafe Constructions have form and meaning poles that are subject to type constraints. Schema Formalism SCHEMA <name> SUBCASE OF <schema> EVOKES <schema> AS <local name> ROLES < self role name>: <role restriction> < self role name> <-> <role name> CONSTRAINTS <role name> <- <value> <role name> <-> <role name> <setting name> :: <role name> <-> <role name> <setting name> :: <predicate> | <predicate> A Simple Example SCHEMA hypotenuse SUBCASE OF line-segment EVOKES right-triangle AS rt ROLES Comment inherited from line-segment CONSTRAINTS SELF <-> rt.long-side Source-Path-Goal SCHEMA: spg ROLES: source: Place path: Directed Curve goal: Place trajector: Entity Translational Motion SCHEMA translational motion SUBCASE OF motion EVOKES spg AS s ROLES mover <-> s.trajector source <-> s.source goal <-> s.goal CONSTRAINTS before:: mover.location <-> source after:: mover.location <-> goal Construction Formalism CONSTRUCTION<name> SUBCASE OF <construction> CONSTRUCTIONAL EVOKES <construction> AS <local name> CONSTITUENTS < local name> : <construction> CONSTRAINTS // as in SCHEMAs FORM ELEMENTS CONSTRAINTS MEANING // as in SCHEMAs // as in SCHEMAs Representing constructions: TO construction TO form selff.phon [thuw] meaning evokes Trajector-Landmark as tl Source-Path-Goal as spg constraints: tl.trajector spg.trajector tl.landmark spg.goal local alias identification constraint The meaning pole may evoke schemas (e.g., image schemas) with a local alias. The meaning pole may include constraints on the schemas (e.g., identification constraints ). The INTO construction construction INTO form selff.phon [Inthuw] meaning evokes Trajector-Landmark as tl Source-Path-Goal as spg Container as cont constraints: tl.trajector spg.trajector tl.landmark cont cont.interior spg.goal cont.exterior spg.source TO vs. INTO: INTO adds a Container schema and appropriate bindings. Grammatical Construction Example CONSTRUCTION Spatial-PP SUBCASE OF Phrase CONSTRUCTIONAL CONSTITUENTS rel: Spatial-Preposition lm: Referring-Exp CONSTRAINTS rel.case <-> lm.case FORM rel < lm MEANING CONSTRAINTS rel.landmark <-> lm The DIRECTED-MOTION construction construction DIRECTED-MOTION constructional constituents mover : Thing motion : Motion-Process direction : Source-Path-Goal form moverf before motionf motionf before directionf meaning evokes Motion-Event as m m.mover moverm m.motion motionm m.path directionm directionm.trajector moverm motionm.mover moverm Semantic specification The analysis process produces a semantic specification that •includes image-schematic, motor control and conceptual structures •provides parameters for a mental simulation Language Understanding Process Constructional analysis Semantic Specification Language understanding: analysis & simulation construction WALKED form selff.phon [wakt] meaning : Walk-Action constraints selfm.time before Context.speech-time selfm..aspect encapsulated “Harry walked into the cafe.” Utterance Analysis Process Constructions General Knowledge Semantic Specification Belief State CAFE Simulation Simulation-based sense disambiguation Ease of construing nominal as a CONTAINER determines what sense of into is appropriate: • The scientist walked into the laboratory. LAB • The scientist walked into the wall. WALL Bonk!! CONTAINER sense CONTACT sense Simulation-based inference Detailed inferences can result from simulation. Image-schematic content of prepositions must fit with properties of other elements of sentence. • The teacher drifted into the house. – Final location of Trajector = inside cafe – Portal = door • The smoke drifted into the house. – Final location of Trajector = inside (possibly throughout) cafe – Portal = door/window World knowledge informs simulation Physical knowledge of how people and gases interact with houses determines: – Relation between Trajector and Interior The smoke drifted into the house and filled it. ?The teacher drifted into the house and filled it. – Portal for motion across Boundary The smoke drifted into the house because the window had been left open. ?The teacher drifted into the house because the window had been left open. Getting From the Utterance to the SemSpec Johno Bryant • Need a grammar formalism – Embodied Construction Grammar (Bergen & Chang 2002) • Need new models for language analysis – Traditional methods too limited – Traditional methods also don’t get enough leverage out of the semantics. Embodied Construction Grammar • Semantic Freedom – Designed to be symbiotic with cognitive approaches to meaning – More expressive semantic operators than traditional grammar formalisms • Form Freedom – Free word order, over-lapping constituency • Precise enough to be implemented Traditional Parsing Methods Fall Short • PSG parsers too strict – Constructions not allowed to leave constituent order unspecified • Traditional way of dealing with incomplete analyses is ad-hoc – Making sense of incomplete analyses is important when an application must deal with “ill-formed” input (For example, modeling language learning) • Traditional unification grammar can’t handle ECG’s deep semantic operators. Our Analyzer • Replaces the FSMs used in traditional chunking (Abney 96) with much more powerful machines capable of backtracking called construction recognizers • Arranges these recognizers into levels just like in Abney’s work • But uses a chart to deal with ambiguity Our Analyzer (cont’d) • Uses specialized feature structures to deal with ECG’s novel semantic operators • Supports a heuristic evaluation metric for finding the “right” analysis • Puts partial analyses together when no complete analyses are available – The analyzer was designed under the assumption that the grammar won’t cover every meaningful utterance encountered by the system. System Architecture Grammar/Utterance Semantic Chunker Learner Semantic Integration Ranked Analyses The Levels • The analyzer puts the recognizer on the level assigned by the grammar writer. – Assigned level should be greater than or equal to the levels of the construction’s constituents. • The analyzer runs all the recognizers on level 1, then level 2, etc. until no more levels. • Recognizers on the same level can be mutually recursive. Recognizers • Each Construction is turned into a recognizer • Recognizer = active representation – seeks form elements/constituents when initiated – Unites grammar and process - grammar isn’t just a static piece of knowledge in this model. • Checks both form and semantic constraints – Contains an internal representation of both the semantics and the form – A graph data structure used to represent the form and a feature structure representation for the meaning. Recognizer Example Mary kicked the ball into the net. This is the initial Constituent Graph for caused-motion. Agent Patient Action Path Recognizer Example Construct: Caused-Motion Constituent: Agent Constituent: Action Constituent: Patient Constituent: Path The initial constructional tree for the instance of Caused-Motion that we are trying to create. Recognizer Example caused motion.m patient.m agent : 5 agent.m , 3category : , 1category : scene : 4 resolved ref : resolved ref : action : 6 caused motion.cm path.m agent : 15 source : action.m , 2tense : 4 patient : {3}{cm1} ,{7} path : x schema : action : 26 goal : path : {7} trajector : {cm1} Recognizer Example processed Mary kicked the ball into the net. A node filled with gray is removed. Patient Agent Action Path Recognizer Example Construct: Caused-Motion RefExp: Mary Constituent: Action Constituent: Patient Constituent: Path Mary kicked the ball into the net. Recognizer Example caused motion.m patient.m agent : 5 agent.m , 3category : , 1category : Person scene : 4 resolved ref : Mary resolved ref action : 6 caused motion.cm path.m agent : 15 source : action.m , 2tense : 4 patient : {3}{cm1} ,{7} path : x schema : action : 26 goal : path : {7} trajector : {cm1} : Recognizer Example processed Mary kicked the ball into the net. Patient Agent Action Path Recognizer Example Construct: Caused-Motion RefExp: Mary Verb: kicked Constituent: Patient Constituent: Path Mary kicked the ball into the net. Recognizer Example caused motion.m patient.m agent : 5 agent.m , 3category : , 1category : Person scene : 4 resolved ref : Mary resolved ref : action : 6 caused motion.cm path.m agent : 15 source : action.m , 2tense : simpPast 4 patient : {3}{cm1} ,{7} path : x schema : kick action : 26 goal : path : {7} trajector : {cm1} Recognizer Example processed Mary kicked the ball into the net. According to the Constituent Graph, The next constituent can either be the Patient or the Path. Agent Patient Action Path Recognizer Example processed Mary kicked the ball into the net. Patient Agent Action Path Recognizer Example Construct: Caused-Motion RefExp: Mary Verb: kicked Det RefExp: Det Noun Constituent: Path Noun Mary kicked the ball into the net. Recognizer Example caused motion.m patient.m agent : 5 agent.m , 3category : ball , 1category : Person scene : 4 resolved ref : Mary resolved ref : action : 6 caused motion.cm path.m agent : 15 source : action.m , 2tense : simpPast 4 patient : {3}{cm1} ,{7} path : x schema : kick action : 26 goal : path : {7} trajector : {cm1} Recognizer Example processed Mary kicked the ball into the net. Patient Agent Action Path Recognizer Example Construct: Caused-Motion RefExp: Mary Verb: kicked RefExp: Det Noun Spatial-Pred: Prep RefExp RefExp Det Noun Prep Det Noun Mary kicked the ball into the net. Recognizer Example caused motion.m patient.m agent : 5 agent.m , 3category : ball , 1category : Person scene : 4 resolved ref : Mary resolved ref : action : 6 caused motion.cm path.m agent : 15 source : action.m , 2tense : simpPast 4 patient : {3}{cm1} ,{7} path : x schema : kick action : 26 goal : net path : {7} trajector : {cm1} Resulting SemSpec After analyzing the sentence, the following identities are asserted in the resulting SemSpec: Scene = Caused-Motion Agent = Mary Action = Kick Patient = Path.Trajector = The Ball Path = Into the net Path.Goal = The net Chunking L3 ________________________S_____________S L2 ____NP _________PP VP NP ______VP L1 ____NP P_______NP VP NP ______VP L0 D 0 N P D N N V-tns Pron Aux V-ing the woman in the lab coat thought you were sleeping 1 2 3 Cite/description 4 5 6 7 8 9 Construction Recognizers Form Meaning Form Meaning D,N <-> [Cloth “you”<->[Addressee] num:sg] NP You NP want to put NP a cloth Like Abney: One recognizer per rule Bottom up and level-based NP on Form Meaning PP$,N <-> [Hand num:sg poss:addr] NP your hand ? Unlike Abney: Check form and semantics More powerful/slower than FSMs Chunk Chart • Interface between chunking and structure merging • Each edge is linked to its corresponding semantics. You want to put a cloth on your hand ? Combining Partial Parses • Prefer an analysis that spans the input utterance with the minimum number of chunks. • When no spanning analysis exists, however, we still have a chart full of semantic chunks. • The system tries to build a coherent analysis out of these semantics chunks. • This is where structure merging comes in. Structure Merging • Closely related to abductive inferential mechanisms like abduction (Hobbs) • Unify compatible structures (find fillers for frame roles) • Intuition: Unify structures that would have been coindexed had the missing construction been defined. • There are many possible ways to merge structures. • In fact, there are an exponential number of ways to merge structures (NP Hard). But using heuristics cuts down the search space. Structure Merging Example Utterance:You used to hate to have the bib put on . Before Merging: [Addressee < Animate] Bib < Clothing num:sg givenness:def Caused-Motion-Action Agent: [Animate] Patient: [Entity] Path:On After Merging: Caused-Motion-Action Agent: [Addressee] Patient: Bib < Clothing num:sg givenness:def Path:On Semantic Density • Semantic density is a simple heuristic to choose between competing analyses. • Density of an analysis = (filled roles) / (total roles) • The system prefers higher density analyses because a higher density suggests that more frame roles are filled in than in competing analyses. • Extremely simple / useful? but it certainly can be improved upon. 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 – Frees structures to be just structures, used in processing • 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. Issues in Scaling up to Language • Knowledge – Lexicon (FrameNet) – Constructicon (ECG) – Maps (Metaphors, Metonymies) (MetaNet) – Conceptual Relations (Image Schemas, X-schemas) • Computation – Representation (ECG) • expressiveness, modularity, compositionality – Inference (Simulation Semantics) • tractable, distributed, probabilistic concurrent, context-sensitive The Buy schema schema Buy subcase of Action evokes Commercial-Transaction as ct roles self ct.nucleus buyer actor ct.customer ct.agent goods undergoer ct.goods The Sell schema schema Sell subcase of Action evokes Commercial-Transaction as ct roles self ct.nucleus seller actor ct.vendor ct.agent goods undergoer ct.goods Extending Inferential Capabilities • Given the formalization of the conceptual schemas – How to use them for inferencing? • Earlier pilot systems – Used metaphor and Bayesian belief networks – Successfully construed certain inferences – But don’t scale • New approach – Probabilistic relational models – Support an open ontology Semantic Web • The World Wide Web (WWW) contains a large and expanding information base. • HTML is accessible to humans but does not formally describe data in a machine interpretable form. • XML remedies this by allowing for the use of tags to describe data (ex. disambiguating crawl) • Ontologies are useful to describe objects and their interrelationships. • DAML+OIL (http://www.daml.org) is an markup language based on XML and RDF that is grounded in description logic and is designed to allow for ontology development, transfer, and use on the web. The ICSI/Berkeley Neural Theory of Language Project Acquisition of early constructions ECG Probabilistic Relation Inference • Scalable Representation of – States, domain knowledge, ontologies • (Avi Pfeffer 2000, Koller et al. 2001) • Merges relational database technolgy with Probabilistic reasoning based on Graphical Models. – Domain entities and relational entities – Inter-entity relations are probabilistic functions – Can capture complex dependencies with both simple and composite slot (chains). • Inference exploits structure of the domain Status of PRMs • Summer Project – Build the basic PRM codebase/infrastructure • Fall Project – Design Coordinated PRM (CPRM) – Build Interface for testing • Spring/Summer 03 – Implement CPRM to replace Pilot System DBN – Test CPRM for QA • Related Work – Probabilistic OWL (PrOWL) – Probabilistic FrameNet Articulating Projects • FrameNet – NSF (with Colorado, USD) • SmartKom – International Consortium • EDU – European Media Lab • Acquaint – ARDA (with SIMS, Stanford) Conclusion • NLU is essential to large, open domain QA. – Much of the web in unstructured data • Substantial Progress in Enabling Technologies – Knowledge Representation/Inference Techniques • Active Knowledge – X-schemas, Simulation Semantics • Dealing With Uncertainty – PRM’s • Combining Statistics and Structure. • Conceptual Relations – Schemas, Metaphor, ECG – Scaling Up • CYC, Wordnet, Term-bases • FrameNet, Semantic Web, MetaNet • Open Source • The goal of NLU can be realized, perhaps! – Anyway, it’s time to try again.