A Best-Fit Approach for Productive Analysis of Omitted Arguments Eva Mok & John Bryant University of California, Berkeley International Computer Science Institute

Simplify grammar by exploiting the language understanding process  Omission of arguments in Mandarin Chinese  Construction grammar framework  Model of language understanding  Our best-fit approach

 Mother (I) give you this (a toy). CHILDES Beijing Corpus (Tardiff, 1993; Tardiff, 1996) ma1+magei3ni3zhei4+ge mothergive2PSthis+CLS  You give auntie [the peach].  Oh (go on)! You give [auntie] [that]. Productive Argument Omission (in Mandarin ) ni3gei3yi2 2PSgiveauntie aoni3gei3ya EMP2PSgiveEMP 4 gei3 give  [I] give [you] [some peach].

Arguments are omitted with different probabilities All arguments omitted: 30.6%No arguments omitted: 6.1%

Construction grammar approach  Kay & Fillmore 1999; Goldberg 1995  Grammaticality: form and function  Basic unit of analysis: construction, i.e. a pairing of form and meaning constraints  Not purely lexically compositional  Implies early use of semantics in processing  Embodied Construction Grammar (ECG) (Bergen & Chang, 2005)

Problem: Proliferation of constructions SubjVerbObj1Obj2 ↓↓↓↓ GiverTransferRecipientTheme VerbObj1Obj2 ↓↓↓ TransferRecipientTheme … SubjVerbObj2 ↓↓↓ GiverTransferTheme SubjVerbObj1 ↓↓↓ GiverTransferRecipient

If the analysis process is smart, then...  The grammar needs only state one construction  Omission of constituents is flexibly allowed  The analysis process figures out what was omitted SubjVerbObj1Obj2 ↓↓↓↓ GiverTransferRecipientTheme

Best-fit analysis process takes burden off the grammar representation Constructions Simulation Utterance Discourse & Situational Context Semantic Specification: image schemas, frames, action schemas Analyzer: incremental, competition-based, psycholinguistically plausible

Competition-based analyzer finds the best analysis  An analysis is made up of:  A constructional tree  A set of resolutions  A semantic specification The best fit has the highest combined score

Combined score that 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 syntactic information, feature agreement  Semantic Fit:  Semantic bindings for frame roles  Frame roles’ fillers are scored

Analyzing ni3 gei3 yi2 (You give auntie)  Syntactic Fit:  P(Theme omitted | ditransitive cxn) = 0.65  P(Recipient omitted | ditransitive cxn) = 0.42 Two of the competing analyses: ni3gei3yi2omitted ↓↓↓↓ GiverTransferRecipientTheme ni3gei3omittedyi2 ↓↓↓↓ GiverTransferRecipientTheme (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 Lexical Unit gei3 Giver (DNI) Recipient (DNI) Theme (DNI) The Transfer Frame Giver Recipient Theme Manner Means Place Purpose Reason Time

Can the omitted argument be recovered from context?  Antecedent Fit: ni3gei3yi2omitted ↓↓↓↓ GiverTransferRecipientTheme ni3gei3omittedyi2 ↓↓↓↓ GiverTransferRecipientTheme Discourse & Situational Context childmother peachauntie table ?

How good of a theme is a peach? How about an aunt? The Transfer Frame Giver (usually animate) Recipient (usually animate) Theme (usually inanimate) ni3gei3yi2omitted ↓↓↓↓ GiverTransferRecipientTheme ni3gei3omittedyi2 ↓↓↓↓ GiverTransferRecipientTheme  Semantic Fit:

The argument omission patterns shown earlier can be covered with just ONE construction  Each cxn is annotated with probabilities of omission  Language-specific default probability can be set SubjVerbObj1Obj2 ↓↓↓↓ GiverTransferRecipientTheme P(omitted|cxn):

16 Research goal  A computationally-precise modeling framework for learning early constructions Language Data Linguistic Knowledge Learning

17 Frequent argument omission in pro-drop languages  Mandarin example: ni3 gei3 yi2 (“you give auntie”)  Even in English, there are often no spoken antecedents to pronouns in conversations Learner must integrate cues from intentions, gestures, prior discourse, etc

18 A short dialogue  bie2 mo3 wai4+tou2 a: #1_3 ! ( 別抹外頭啊 )  NEG-IMP apply forehead  Don ’ t apply [lotion to your] forehead  mo3 wai4+tou2 ke3 jiu4 bu4 hao3+kan4 le a:. ( 抹外頭可就不好看了啊 )  apply forehead LINKER LINKER NEG good looking CRS SFP  [If you] apply [lotion to your] forehead then [you will] not be pretty …  ze ya a: # bie2 gei3 ma1+ma wang3 lian3 shang4 mo:3 e: ! ( 嘖呀啊 # 別給媽媽往臉上抹呃 )  INTERJ # NEG-IMP BEN mother CV-DIR face on apply  INTERJ # Don ’ t apply [the lotion] on [your mom ’ s] face (for mom)  [- low pitch motherese] ma1+ma bu4 mo:3 you:2. ( 媽媽不抹油 )  mother NEG apply lotion  Mom doesn ’ t apply (use) lotion

Goals, refined  Demonstrate learning given  embodied meaning representation  structured representation of context  Based on  Usage-based learning  Domain-general statistical learning mechanism  Generalization / linguistic category formation 19

Towards a precise computational model  Modeling early grammar learning  Context model & Simulation  Data annotation  Finding the best analysis for learning  Hypothesizing and reorganizing constructions  Pilot results 20

Embodied Construction Grammar 21 construction yi2-N subcase of Morpheme form constraints self.f.orth <-- "yi2" meaning evokes RD as rd constraints self.m rd.referent self.m rd.ontological_category

“you” specifies discourse role 22 construction ni3-N subcase of Morpheme form constraints self.f.orth <-- "ni3" meaning evokes RD as rd constraints self.m rd.referent self.m rd.ontological_category rd.discourse_participant_role rd.set_size

The meaning of “give” is a schema with roles 23 construction gei3-V2 subcase of Morpheme form constraints self.f.orth <-- "gei3" meaning : Give schema Give subcase of Transfer constraints inherent_aspect giver recipient theme schema Transfer subcase of Action roles giver recipient theme constraints giver protagonist

Finally, you-give-aunt links up the roles 24 construction ni3-gei3-yi2 subcase of Finite_Clause constructional constituents n : ni3-N g : gei3-V2 y : yi2-N form constraints n.f meets g.f g.f meets y.f meaning : Give constraints self.m g.m self.m.giver n.m self.m.recipient y.m

The learning loop: Hypothesize & Reorganize 25 Linguistic Knowledge Discourse & Situational Context Analysis World Knowledge Context Fitting

26 XIXI INV Discourse Segment addresseespeaker If the learner has a ditransitive cxn meets ni3 Addressee giver Give gei3 recipient Aunt yi2 omitted theme Peach MOT

27 XIXI giver Give INV recipient Discourse Segment addresseespeaker Context fitting recovers more relations meets ni3 Addressee giver Give gei3 recipient Aunt yi2 omitted theme Peach MOT theme attentional- focus

28 giver Discourse Segment addresseespeaker Peach MOT theme attentional- focus recipient XIXI Give INV But the learner does not yet have phrasal cxns ni3 Addressee Give gei3 Aunt yi2 giver recipient meets

29 Context bootstraps learning ni3 Addressee Give gei3 Aunt yi2 meets construction ni3-gei3-yi2 subcase of Finite_Clause constructional constituents n : ni3 g : gei3 y : yi2 form constraints n.f meets g.f g.f meets y.f meaning : Give constraints self.m g.m self.m.giver n.m self.m.recipient y.m giver recipient

30 A model of context is key to learning  The context model makes it possible for the learning model to:  learn new constructions using contextually available information  learn argument-structure constructions in pro-drop languages

31 Understanding an utterance in context Schemas + Constructions Simulation Transcripts Events + Utterances Semantic Specification Analysis + Resolution Context Fitting Context Model Recency Model

32 Context model: Events + Utterances Setting participants, entities, & relations Setting participants, entities, & relations Start Event DS Sub-Event

33 Entities and Relations are instantiated Setting CHI, MOT (incl. body parts) livingroom(incl. ground, ceiling, chair, etc), lotion Setting CHI, MOT (incl. body parts) livingroom(incl. ground, ceiling, chair, etc), lotion Start apply02 applier = CHI substance = lotion surface = face(CHI) apply02 applier = CHI substance = lotion surface = face(CHI) ds04 admonishing05 speaker = MOT addressee = CHI forcefulness = normal ds04 admonishing05 speaker = MOT addressee = CHI forcefulness = normal caused_motion01 forceful_motionmotion caused_motion01 forceful_motionmotion translational_motion03 mover = lotion spg = SPG translational_motion03 mover = lotion spg = SPG

The context model is updated dynamically  Extended transcript annotation: speech acts & events  Simulator inserts events into context model & updates it with the effects  Some relations persists over time; some don’t. 34 Simulation Events Context Model Recency Model

35 Competition-based analyzer finds the best analysis  An analysis is made up of:  A constructional tree  A semantic specification  A set of resolutions Bill gaveMarythe book MaryBill Ref-Exp Give A-GIVE-B-X subj vobj1 giver recipient theme

36 Combined score that 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 syntactic information, feature agreement  Semantic Fit:  Semantic bindings for frame roles  Frame roles’ fillers are scored

37 XIXI giver Give INV recipient Discourse Segment addresseespeaker Context Fitting goes beyond resolution meets ni3 Addressee giver Give gei3 recipient Aunt yi2 omitted theme Peach MOT theme attentional- focus

38 Context Fitting, a.k.a. intention reading  Context Fitting takes resolution a step further  considers entire context model, ranked by recency  considers relations amongst entities  heuristically fits from top down, e.g. discourse-related entities complex processes simple processes other structured and unstructured entities  more heuristics for future events (e.g. in cases of commands or suggestions)

39 Adult grammar size  ~615 constructions total  ~100 abstract cxns (26 to capture lexical variants)  ~70 phrasal/clausal cxns  ~440 lexical cxns (~260 open class)  ~195 schemas (~120 open class, ~75 closed class)

40 Starter learner grammar size  No grammatical categories (except interjections)  Lexical items only  ~440 lexical constructions  ~260 open class: schema / ontology meanings  ~40 closed class: pronouns, negation markers, etc  ~60 function words: no meanings  ~195 schemas (~120 open class, ~75 closed class)

41 The process hierarchy defined in schemas Process State Action State_ Change Complex_Process Proto_Transitive Intransitive_ State Two_Participant_ State Mental_State Joint_Motion Caused_Motion Concurrent_ Processes Cause_Effect Serial_Processes

42 The process hierarchy defined in schemas Action Translational_ Motion Translational_ Self_Motion Motion Intransitive_Action Expression Self_Motion Force_Application Continuous_ Force_Application Agentive_Impact Forceful_Motion

43 The process hierarchy defined in schemas Action Perception Ingestion Communication Transfer Cause_Change Other_ Transitive_Action Obtainment

44 Understanding an utterance in context Schemas + Constructions Simulation Transcripts Events + Utterances Semantic Specification Analysis + Resolution Context Fitting Context Model Recency Model

Hypothesize & Reorganize  Hypothesize:  utterance-driven;  relies on the analysis (SemSpec & context)  operations: compose  Reorganize:  grammar-driven;  can be triggered by usage (to be determined)  operations: generalize 45

Composing new constructions  Compose operation: If roles from different constructions point to the same context element, propose a new construction and set up a meaning binding. 46 ni3 Addressee Give gei3 giver recipient theme XIXI INV Peach MOT

Creating pivot constructions  Pivot generalization: Given a phrasal cxn, look for another cxn that shares 1+ constituents. Line up roles and bindings. Create new cxn category for the slot. 47 ni3 Addressee Give yi2 giver recipient meets ni3 Addressee Give wo3 giver recipient meets

48 Resulting constructions construction ni3-gei3-cat01 constituents ni3, gei3, cat01 meaning : Give constraints self.m.recipient g.m construction wo3 subcase of cat01 construction yi2 subcase of cat01 general construction cat01 subcase of Morpheme

Pilot Results: Sample constructions learned  Composed:  Pivot Cxns: 49 chi1_fan4 ni3_chuan1_xie2 ni3_shuo1 bu4_na2 wo3_qu4 ni3_ping2zi_gei3_wo3 ni3_gei3_yi2 wo3_bu4_chi1 eat rice you wear shoe you say NEG take I go you bottle give me you give aunt I NEG eat ni3 {shuo1, chuan1} ni3 {shuo1, hua4} wo3 {zhao3, qu4} bu4 {na2, he1} {wo3, ma1} cheng2 you {say, wear} you {say, draw} I {find, go} NEG {take, drink} {I, mom} scoop

Challenge #1: Non-compositional meaning  Non-compositional meaning: Search for additional meaning schemas (in context or in general) that relate the meanings of the individual constructions 50 you Addressee Bake bake baker baked Bake-Event CHI Cake MOT a Give-Event

Challenge #2: Function words  Function words tend to indicate relations rather than events or entities 51 you Addressee Bake bake baker baked Bake-Event CHI Cake MOT a for Benefaction

Challenge #3: How far up to generalize  Eat rice  Eat apple  Eat watermelon  Want rice  Want apple  Want chair 52 Inanimate Object Manipulable Objects Manipulable Objects Unmovable Objects Food Furniture Fruit Savory Chair Sofa apple watermelon rice

Challenge #4: Beyond pivot constructions  Pivot constructions: indexing on particular constituent type Eat rice; Eat apple; Eat watermelon  Abstract constructions: indexing on role-filler relations between constituents 53 Schema Eat roles eater agent food patient food Eat catX Schema Want roles wanter agent wanted patient wanted Want catY

Challenge #5: Omissible constituents  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 54

When does the learning stop?  Most likely grammar given utterances and context  The grammar prior is a preference for the “kind” of grammar  In practice, take the log and minimize cost  Minimum Description Length (MDL) 55 Bayesian Learning Framework Schemas + Constructions SemSpec Analysis +Resolution Context Fitting

Intuition for MDL  S -> Give me NP  NP -> the book  NP -> a book  S -> Give me NP  NP -> DET book  DET -> the  DET -> a 56 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

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 57

How to calculate the prior of this grammar  (Yet to be determined)  There is evidence that the lexicalized constructions do not completely go away  If the more lexicalized constructions are retained, the size of grammar is a bad indication of degree of generality 58