Building Text Meaning Representations from Contextually Related Frames

Building Text Meaning Representations from Contextually Related Frames – A Case Study – Aljoscha Burchardt Anette Frank Manfred Pinkal Saarland University and DFKI Saarbrücken

Motivation ² Broad-coverage statistical parsing systems – High demand for more „semantics-based“ processing ² Deep computational semantic processing – Well-studied formalisms for truth-conditional and discourse semantics – Large-scale deep semantic parsing (e. g. , Bos et al. , 2004) – Little emphasis on lexical semantics and concept-based analysis ² Lexical semantic resources – Word. Net(s) used for approximate concept-based analysis – Lexical semantics resources that model predicate-argument structure (e. g. Frame. Net, Prop. Bank) – Automatic semantic role labelling (Con. LL, Senseval) ² Aim: Building partial text meaning representations from frameannotated deep syntactic structures

Overview Frame Semantics for Partial Text Meaning Representation ² Background: Frame. Net „as a Net“ ² Building text meaning representations from frame semantic annotations • A case study • Linking of contextually related frames and frame roles – Based on patterns of lexico-semantic and contextual relations • Generalisation and acquisition of linking patterns ² Towards Automation: current architecture ² Conclusion

Frame. Net ² Frame Semantics (Fillmore 1976, 1977, . . ) – Frame: a conceptual structure or prototypical situation – Frame elements (roles) • Identify participants of the situation • Are local to their frame – Frame evoking elements (verbs, nouns, adjectives) introduce frames – E. g. VERDICT: [The jury]Judge convicted [him]Defentant [on the counts of theft]Charges. On Thursday [a jury]Judge found [the youth]Defendant [guilty of wounding Mr Lay] Finding ² Berkeley Frame. Net Project – Database of frames for core lexicon of English – Current release: 610 frames, about 9000 lexical units

Frame. Net „as a Net“ – Frame-to-Frame Relations – Inheritance relation: a frame inherits all roles of one or more “super” frame(s) Intentionally_act Agent Patient Arrest Authorities Suspect Offense Charges

Frame. Net „as a Net“ – Frame-to-Frame Relations – Subframe relation – Super frame represents complex event – Subframes represent sub-events – Subframes usually inherit some roles of the super frame Defendant Court Defense Judge Jury Offense Criminal process Charge Prosecution Defendant Charge Arraignment . . . Arrest . . . Sentencing . . . Trial

Frame Semantics for Partial Text Meaning Representation ² Probabilistic models for semantic role labeling (Gildea & Jurafsky, 2002) ² Frame semantic projection from deep (LFG) grammar (Frank & Erk, 2003) – No constructional “glue” – Partially connected/embedded lexico-semantic predicateargument structures – Coarse-grained semantic structures ² Challenge: obtain a more densely connected representation – By learning and applying heuristic linking patterns

A Case Study In the first trial in the world in connection with the terrorist attacks of 11 September 2001, the Higher Regional Court of Hamburg has passed down the maximum sentence. Mounir al Motassadeq will spend 15 years in prison. The 28 -year-old Moroccan was found guilty as an accessory to murder in more than 3000 cases.

Local Roles In the first trial in the world in connection with [the [terrorist]Assailant attacks of [11 September 2001]Time]Case, [the Higher Regional Court of Hamburg]Court has passed down the [maximum]Type sentence.

Local Roles [Mounir al Motassadeq]Inmates will spend [15 years]Duration in prison.

Local Roles [The 28 -year-old Moroccan]Defendant was found [guilty]Finding as [an accessory to [murder]Focal. Entity [in more than 3000 cases]Victim ]Charge.

Unfilled Roles Target Frame roles Filler (given vs. Induced) trial TRIAL attacks ATTACK terrorist attacks accessory to murder Higher Regional Court 28 -year-old Moroccan terrorist sentence SENTENCING CASE CHARGE COURT DEFENDANT. . . ASSAILANT VICTIM. . . TIME (exth. ) CONVICT COURT TYPE. . . 11 September 2001 Mounir al Motassadeq Higher Regional Court maximum sentence (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) prison PRISON INMATES. . . DURATION (exth. ) Mounir al Motassadeq 15 years (11) (12) found VERDICT accessory ASSISTANCE terrorist attacks accessory to murder 28 -year-old Moroccan guilty murder KILLING CASE CHARGE DEFENDANT FINDING. . . CO-AGENT FOCAL_ENTITY HELPER. . . KILLER VICTIM. . . (13) (14) (15) (16) (17) (18) (19) (20) (21) murder 28 -year-old Moroccan m. t. 3000 cases

Frames in Context Target Frame roles Filler (given vs. Induced) trial TRIAL attacks ATTACK terrorist attacks accessory to murder Higher Regional Court 28 -year-old Moroccan terrorist sentence SENTENCING CASE CHARGE COURT DEFENDANT. . . ASSAILANT VICTIM. . . TIME (exth. ) CONVICT COURT TYPE. . . 11 September 2001 Mounir al Motassadeq Higher Regional Court maximum sentence (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) prison PRISON INMATES. . . DURATION (exth. ) Mounir al Motassadeq 15 years (11) (12) found VERDICT accessory ASSISTANCE terrorist attacks accessory to murder 28 -year-old Moroccan guilty murder KILLING CASE CHARGE DEFENDANT FINDING. . . CO-AGENT FOCAL_ENTITY HELPER. . . KILLER VICTIM. . . (13) (14) (15) (16) (17) (18) (19) (20) (21) murder 28 -year-old Moroccan m. t. 3000 cases

Linking Frames and Roles in Context ² At the instance level – given frame instances f 1: F 1 and f 2: F 2, where • f 1 and f 2 stand in a contextual relation (syn, sem, discourse) • frame types F 1 and F 2 stand in some frame relation => identify role instances (referents) of f 1 and f 2 (r 1 (= r 0) = r 2) frame relation context-related instances inferred relation

Linking Frames and Roles in Context In the first trial in the world in connection with the terrorist attacks of 11 September 2001, the Higher Regional Court of Hamburg has passed down the maximum sentence. Criminal Process Court Sentencing Trial frame relation

Linking Frames and Roles in Context In the first trial (f 1) in the world in connection with the terrorist attacks of 11 September 2001, [the Higher Regional Court of Hamburg] (r 2) has passed down the maximum sentence (f 2). Criminal Process Court Sentencing Functional Embedding Trial The Higher Regional Court of Hamburg frame relation context-related instances

Linking Frames and Roles in Context In the first trial (f 1) in the world in connection with the terrorist attacks of 11 September 2001, [the Higher Regional Court of Hamburg] (r 2=r 0= r 1) has passed down the maximum sentence (f 2). Criminal Process Court Sentencing Functional Embedding Trial The Higher Regional Court of Hamburg frame relation context-related instances inferred relation

Linking Frames and Roles in Context ² At the type level (more involved) – If instances of frame roles f 1: F 1 and f 2: F 2 are often found coreferent within particular contextual relations => Hypothesize a frame relation between F 1 and F 2 (no) frame relation context-related instances inferred relation

Linking Frames and Roles in Context … the Higher Regional Court of Hamburg has passed down the Maximum sentence. [Mounir al Motassadeq] will spend 15 years in prison. Prison Sentencing Inmates Discourse Relation • New Frame Relation • (Role Binding: Convict=Inmates) Convict (Co-reference) (no) frame relation context-related instances inferred relations

Meaning Postulates („Semantic Control“) ² Example: the Defendant in a VERDICT is the actor of the Frame embedded in the VERDICT Charge role („Charge event“) ² Modeled as „Semantic Control“: – VERDICT embeds some frame Fx under role Charge, – Fx. Rx inherits from INTENT. _ACT. AGENT => VERDICT. Defendant equals Fx. Rx (at the type level) INTENT. _ ACT VERDICT Defendant Charge Agent Fx Semantic Control Rx „[. . . Moroccan]DEF was found guilty as an accessory (Fx) to murder“ => VERDICT. Defendant equals ASSISTANCE. Helper

Generalisation Mounir El Motassadeq (born April 3, 1974) is a Moroccan. In February 2003 he was convicted [. . . ]. As of April 2004 he is the only person to have been convicted in direct relation to the September 11, 2001 attacks. The verdict and sentence were set aside on appeal [. . . ]. A new trial is expected in mid-2004. ² Selected deeper semantic represention to model – referential properties (introduction of new discourse referents) => blocking factors for role identification rules – Temporal sequence and locational properties => deeper contextual semantic relations between frames

Acquisition of Linking Patterns ² Identified patterns for induction of role-linking – Lexico-semantic relations • Subframe and Inheritance relation, Semantic Control – Contextual relations • Syntactic and semantic embedding • Anaphoricity and referential properties • Discourse relations (or surface linearisation) ² Future work – Learning weighted role-linking patterns from annotated texts in restricted domains, to be used as probabilistic inference rules – “[Baragiola]CONVICT / ESCAPEE had previously been convicted of murder in Italy, but had escaped in 1980 and obtained Swiss citizenship. ” => Infer PRISON and PRISON. Inmates = ESCAPE. Escapee

Current Architecture ² LFG-based parsing and syntax-semantics interface – Par. Gram grammars for German and English (Butt et al. 2002) – Frame projection from f-structure (XLE transfer system) – Interfaces to statistical frame assignment (Baldewein et al. 2004) ² Enriching Semantic Representation – Rule-based refinement of semantic representation – Autom. assignment of SUMO/MILO classes (using Word. Net WSD) ² Logical Representation and Reasoning – Frame relations translated to logic programs – Joint work with P. Baumgartner and F. Suchanek, MPI Saarbrücken – First scenario: RTE Challenge

Conclusion ² Combining – Deep syntactic analysis and Frame Semantic role assignments ² Methods – Linking partial frame annotations in context – Generalisation and automation ² Shallow semantic representations – Necessarily partial (focusing on open class categories) – Robust semantic processing for coarse-grained information access – Incremental depth for finer-grained analysis

Frame Exchange Format „fef“

Types of Relations ² Frame. Net Relations – Frame hierarchy: inherits – Subframes ² Contextual Relations between instantiated frames and roles – Syntactic and/or semantic embedding – Discourse relations – Anaphoric relations ² Inferences – On the basis of both

Frame, Contextual, and Inferred Relations CRIMINAL PROCESS PRISON (2) Inmates Duration SENTENCING (1) Convict Type TRIAL (1) Court Defendant Charge Case VERDICT (3) Defendant Finding (1) sentence number Subframe/FE Contextual Relation Inferred Relation KILLING (3) Victim Killer Charge Case ASSISTANCE (3) Helper Focal_entity Co_agent

CRIMINAL PROCESS PRISON Inmates (Motus. ) Duration (15 Y) SENTENCING TRIAL Convict Duration Court (maximum) (Hmbg. ) Defendant VERDICT Defendant (the Moroccan) Charge Case (9/11) Charge Case (accessory) ASSISTANCE Hierarchy/Subframe/FE Contextual Relations Inference KILLING Victim (3000) Goal (murder) Helper Co_agent Killer In the first trial. . the higher Regional Court. . has passed down the maximum sentence. Mounir al Motussadeq will spend 15 years in prison. The 28 -year-old Moroccan was found guilty as an accessory to murder in. . 3000 cases.

More Involved Examples Semantic control A meaning relation between frame roles F 2: R 2 and Fx: Rx – where F 2 embeds Fx (via some role R 3), and – F 2: R 2 semantically controls (is co-referent with) the AGENT role Fx: Rx of Fx frame relation context-related instances inferred relation

More Involved Examples Semantic control Example: „. . . was found guilty as an accessory to murder“ VERDICT found CASE CHARGE DEFENDANT FINDING. . . INTENTIONALLY_ACT AGENT ASSISTANCE accessory CO-AGENT FOCAL_ENTITY HELPER. . . frame relation terrorist attacks accessory to murder 28 -year-old Moroccan guilty (13) (14) (15) (16) murder 28 -year-old Moroccan (17) (18) (19) context-related instances inferred relation

More Involved Examples (? ) ² Semantic Control – „. . . was found guilty as an accessory to murder“ • The CO-AGENT of ASSISTANCE is co-referent with the AGENT of the frame embedded under ASSISTANCE. FOCAL_ENTITY => ASSISTANCE. CO-AGENT coreferent with KILLING. KILLER – SENTENCING. CONVICT co-referent with PRISON. INMATES • CAUSATION. CAUSE embeds SENTENCING CAUSATION. EFFECT embeds PRISON PATIENT(CAUSATION. CAUSE) co-referent with PATIENT(CAUSATION. EFFECT) (semantic control) (where PATIENT(Frame) inherits from INTENTIONALLY_AFFECT. PATIENT) PATIENT(SENTENCING) = SENTENCING. CONVICT => PATIENT(PRISON) = PRISON. INMATES