PRAXICON: The Development of a Grounding Resource Katerina ... · 1a) PRAXICON word in entry also...
Transcript of PRAXICON: The Development of a Grounding Resource Katerina ... · 1a) PRAXICON word in entry also...
PRAXICON: PRAXICON: The Development of a Grounding The Development of a Grounding
Resource Resource
Katerina Pastra
Institute for Language & Speech Processing(ILSP/Athena R.C.)
Bellagio Meeting6-7 October 2008
OverviewOverview
• The notion of a PRAXICON• Grounding Resources (GR)• Development within the POETICON
project• The COSMOROE approach• Trends in Language Resources (LR)• Outlook
What will the PRAXICON be?What will the PRAXICON be?
a) a “lexicon” with grounded lemmas
word + sensorimotor representationword sense + visual & motoric repr.
Andb) a “lexicon” with conceptual/pragmatic
relations between grounded lemmas
WordX REL WordY
Conceptual structures of different levels
(from action-object combinations, action-action combinations to …scripts…?)
What will the PRAXICON be?What will the PRAXICON be?
Why do we need a PRAXICONWhy do we need a PRAXICON
- AV processing- Human-Computer/Robot
interaction
Grounding needed:
- To tie words to sensorimotor experiences (disambiguation)
- To untie sensorimotor experiences from physical specificity (intentionality indication)
Other Grounding Resources (?)Other Grounding Resources (?)“Integration” suggestions:
Word + visual object representationWord + image-region & visual feature-value vector (Bajcsy & Joshi78)
(Word + 3D model of object)+conceptual structures (Jackendoff83)
Small-scale implementation:
• In AI applications: words + object/action repr.
+ link to conceptual structures • Multimedia Thesauri (e.g. Benitez et al. 2000)
• Multimedia Ontologies (e.g. Zinger 2005 – OntoImage)
• Annotated Corpora (collections of labelled images, e.g. Lin et al. 03)
f-v vectors and/or actual images,
wireframe models, 2D drawings, motion trajectories etc.
ontologies, domain models,
frames etc.
Other Grounding Resources (2)Other Grounding Resources (2)Issues:- Never done systematically, i.e. there is no grounding resource of any scale beyond ad hoc system development- Content (blobs or mini-worlds, not always link to conceptual structures…)- Methodology (manual association & visual abstraction)- Scaling, extending or even basic questions related to the development of such resource not posed
POETICON SuggestionPOETICON Suggestion
PRAXICON developmentPRAXICON development
COSMOROE annotated corpus
POETICON Cognitive Experiments
AssociationAssociationv in everyday interaction v through cognitive categorization
“trolley”
“walk”
What and How?
Naming Strategies
PRAXICON experimentationPRAXICON experimentation
“the yellow taxi-boats…”
Get associations,
create new entries/expand etc.
COSMOROE annotated corpus
COSMOROE RelationsCOSMOROE Relations
see: Pastra 2008, Multimedia Systems Journal
“… helmet for safety...”
TypeType--token equivalencetoken equivalence
“The city, of course, is Athens , and it is here that I will begin my exploration of modern
Greece.”
MetonymyMetonymyThe two referents come from the same domain, have same array of associations, there is no transfer of qualities from one to another – the two modalities refer to different entities but the user intends the two modalities to be considered semantically equivalent
Essential ExophoraEssential Exophora A pragmatic “anaphora” case
“…[pollution has taken its toll] on this..”
COSMOROE Annotated CorpusCOSMOROE Annotated Corpus
Annotation particulars:
• TV travel programmes• 3 hours EL - 1.5 hours EN • validated• 0.88 inter-annotator agreement• tools used: Transcriber (Baras et al. 1998) and ANVIL (Kipp 2004)
• Annotation scheme: multi-facetedannotation comprises of indicating the time offsets of different modalities and the relation into which they participate
COSMOROE annotationCOSMOROE annotation
Annotation ByAnnotation By--ProductsProducts- manual speech transcriptions- optical character transcriptions- acoustic events gold data- audiovisual topics gold data- body movements & gestures gold data- shots gold data- object and event identification gold data
è Important for training and evaluation of corresponding technologiesè Further refinements for specific purposes can be done with less cost and effort (e.g. fully word-level transcription or annotation of gesture phases etc.)
PRAXICON experimentationPRAXICON experimentation
Use cases for PRAXICON extension mechanism:
A) Input: action X or/and object Y Output needed: name the action or object
1. similar visual info in PRAXICON entry = LUCKY
1a) PRAXICON word in entry also present in Utterance = REALLY LUCKY
1b) PRAXICON word in entry not present in Utterance (due to e.g. synonymy, metaphor, antithesis, complementarity, independence etc.) => need COSMOROE rel identification mechanism (trained on CM corpus) à new entry creation
2. no similar visual info in PRAXICON
2a) word in utterance exists in PRAXICON with different visual info (different sense) à new entry creation using CM algorithm
2b) word in utterance not present in PRAXICON => need CM algorithm to decide on new entry to be created and its relation to others
QuestionsQuestions
1) What will grounded lemmas be like? 1-word? Multi-word? Word-centric? sensorimotor repr. – centric?
2) How will they be organised?
3) How specific or general should they be?
4) What kind of relations between entries/lemmas ???
QuestionsQuestions
What information should be included? - Morphological info? (inflection, POS)- Syntactic info? (subcategorisation info)- Syntactico-semantic info?
(thematic roles, selectional restrictions)- Morpho-semantic info? (derivational links)- Lexical semantic relations
(synonymy, antonymy, meronymy…)- Conceptual relations (time e.g. temporal inclusion, manner e.g. troponymy, causation…)?- Will it include facts?
Turning to Language ResourcesTurning to Language Resources
a) Can we tune one to get it grounded?b) Can we interface one with the PRAXICON at some conceptual level?c) Can we use one to develop a mechanism for extending the PRAXICON?
TrendsTrends
• most resources get extended so that they cover more types of conceptual information/relations, going also down to the level of specific instances and facts• the resources get mapped to each other for greater usability• there is a constant search for automatic or semi-automatic mechanisms for extending the coverage of the resources, and• there is a growing development of the resources in different languages, all mapped to each other
Outlook Outlook
è AI Quest for primitive concepts or features – feature bundles to describe the world in a universal way
è categorisation and story-telling to organise and talk about the world
But why not use the sensorimotor experience itself ?
Based on LR lessons, implement this new perspective
We envision PRAXICON We envision PRAXICON ……(A) To be a sensorimotor-centric resourceEntries beyond lexicalization, e.g. different visual representation of
toe and finger entries in the PRAXICON, though single lexicalization in some languages
This implies:- Different organisation than currently available in LRs- Greater granularity in concept analysis (meaning decomposition
down to the level inferences are minimal)
(B) To be a resource that goes beyond the readily lexicalised conceptual level, to the one at which inferences come into play (i.e. the meeting point of symbolic and sensorimotor representations)
(C) To let sensorimotor experience deal with the specifics of everyday interaction, &let language do what it does best: abstraction + interpretation