Upload
zoe-carson
View
224
Download
0
Tags:
Embed Size (px)
Citation preview
ISO/TC37/SC4/WG2 WD24617-5SemAF - Discourse Structure
2011-01-12, OxfordHASIDA Koiti
[email protected], Japan
Outline
IntroductionDiscourse Graph: Semantic
StructureDiscourse Tree: Syntactic
(Presentational) StructureDiscourse Relation
2
ISO ContextStatus: WD
NWI ballot for 2010-07/-10Experts
Koiti Hasida (Japan): project leaderHarry Bunt (Netherland)Jerry Hobbs (USA)Nancy Ide (USA)Rashmi Prasad (USA)Kiyong Lee (Korea)Roland Hausser (Germany)Claudia Soria (Italy)Eric de la Clergerie (France)Antonio Pareja Lora (Spain)
3
Objective and Scopeannotation, production, translation, etc. of
various types of documentsannotation to video for generating multilingual
subtitlesextension of Twitter to support semantically
structured discussion by introducing semantic relations among tweets
semantic structure of discourse, consisting of eventualities and semantic relations (in particular discourse relations) among themdiscourse semantics not only in linguistic content
but also in (possibly silent) video, hypertext, game, etc.
documents without predefined total temporal ordering of presentation, such as hypertexts and games
syntactic (presentational) structure of discourse, comprising discourse segments (phrases, clauses, sentences, video scenes, and so on)
guideline to develop and maintain minimal set of discourse relations
4
Outline
IntroductionDiscourse Graph: Semantic
StructureDiscourse Tree: Syntactic
(Presentational) StructureDiscourse Relation
5
Terms and Definitions (Semantics)
eventuality*: event (possibly dialogue act), state (including claim, constraint, semantic relation (in particular discourse relation)), process, proposition, or their abstraction (type)
discourse relation: semantic relation among eventualities?
6
Discourse GraphLogical Form of Discoursegraph representing discourse semanticsNodes and links represent eventualities.Links represent semantic relations (in
particular discourse relations) among eventualities.
Representation of nodes may follow any scheme to describe semantic structure of eventualities.ISO 24617-2 (SemAF-DA) recommended for
describing eventualities consisting of dialogue acts, communicative functions, propositional content, etc.
Representation of coreference and anaphora is attributed to other framework on semantic representation. 7
Sample Discourse Graph
8
authoring of content must be easyauthoring of content must be easy
semantic annotation is necessarysemantic annotation is necessary
inferenceinference
inferenceinference
huge amount of content is necessaryhuge amount of content is necessary
retrieval must be quick and easyretrieval must be quick and easy
realize ubiquitous information servicerealize ubiquitous information service
purposepurpose
purposepurpose
inferenceinference
In order to realize ubiquitous information access, huge amount of content is necessary, so that authoring of content must be easy. Also in order to realize ubiquitous information access, retrieval must be quick and easy. So semantic annotation is necessary.
you slipped vegetables under the tableyou slipped vegetables under the table
do you remember?do you remember?
maybemaybe
Sparky lived so longSparky lived so long
causescauses
contentcontent
objectobject
Links May Connect Other Links
because semantic relations are eventualitis.
9
Do you remember you slipped vegetables under the table? Maybe that’s why Sparky lived so long.
thematic rolethematic role
you slipped vegetables under the tableyou slipped vegetables under the table
do you remember?do you remember?
Sparky lived so longSparky lived so long
contentcontent
causescauses maybemaybe
Reification
10
2nd2nd
objectobject
1st1st
authoring of content must be easyauthoring of content must be easy
semantic annotation is necessarysemantic annotation is necessary
inferenceinference
inferenceinference
huge amount of content is necessaryhuge amount of content is necessary
retrieval must be quick and easyretrieval must be quick and easy
realize ubiquitous information servicerealize ubiquitous information service
purposepurpose
purposepurpose
HypernodeA discourse graph may be
embedded (as a hypernode) in a larger discourse graph.
11
hypernode
Outline
IntroductionDiscourse Graph: Semantic
StructureDiscourse Tree: Syntactic
(Presentational) StructureDiscourse Relation
12
Terms and Definitions (Syntax; 1)discourse segment: part of discourse
representing eventuality.discourse modifier: part of discourse
not representing eventuality, comprising a dependent discourse connective and a discourse segment as its object (representing $2 of the associated discourse relation). The discourse segment modified by a discourse modifier represents $1 of the associated discourse relation. For instance, ‘because I’m sleepy’ is a discourse modifier consisting of dependent discourse connective ‘because’ and its object ‘I’m sleepy’. 13
Terms and Definitions (Syntax; 2)discourse connective: part of discourse
representing a discourse relation without its arguments. A discourse connective is not a discourse segment. It need not be continuous.
dependent discourse connective: discourse connective representing (not reified) discourse relation.
independent discourse connective: discourse connective representing eventuality (reified discourse relation).
discourse connective qualifier: part of discourse qualifying a discourse connective. A discourse connective and a discourse connective qualifier qualifying it constitute a larger discourse connective. For instance, ‘maybe that’s why’ is a discourse connective consisting of more basic discourse connective ‘that’s why’ and its qualifier ‘maybe’. 14
Discourse Tree
Parse Tree of Discourseannotated tree structure of linearly-
ordered discourse representing presentational structures of the discourse
addresses presentational aspects including importance (nucleus/satellite distinction).
abstract syntaxstraightforward to encode in LAF
consistent with intra-sentential syntaxharmonization requirements?
with SynAF and other annotation practices15
A minus sign is the inverse operator.A minus sign is the inverse operator.
Sample Discourse Tree
16
[ [Semantic annotation is necessary {-inference because [conjunction [2 retrieval must be quick and easy]
and [3 authoring of content must be easy]]}.]
[2 Retrieval must be quick and easy
{purpose in order to [0 realize ubiquitous information access]}].
[3 Authoring of content must be easy
{-inference because [1 huge amount of content is necessary
{purpose in order to [0 realize ubiquitous information access]}]}.]
]
discourse connectivediscourse connective
discourse relationdiscourse relation
The object of a discourse
connective is $2 of the discourse
relation.
The object of a discourse
connective is $2 of the discourse
relation.
A discourse modifier depends on $1 of the discourse
relation.
A discourse modifier depends on $1 of the discourse
relation.
A pair of curly brackets encloses a discourse modifier.
A pair of curly brackets encloses a discourse modifier.
A pair of brackets encloses a discourse
segment which is not a discourse
modifier.
A pair of brackets encloses a discourse
segment which is not a discourse
modifier.
Intersentential (Anaphoric) Discourse Connective
1 is more general than 2?1.[addition [Tom was tired]. Also [he
was feverish].]2.[1 Tom was tired]. [{addition Also1}
he was feverish.]Another example:[contrast On the one hand, [Tom is
hungry]. On the other, [Mary is thirsty].]
17
Intersentential (Anaphoric) Discourse Connective (cont.)
[{content [1 But a strong level of investor withdrawal is much more unlikely this time around]}, fund managers said.][-inference A major reason1 is that [investors already have sharply scaled back their purchases of stock funds since Black Monday].]
18
Mapping from Discourse Tree to Discourse Graph
dependent discourse connective-> link
discourse modifier= dependent discourse connective
+ discourse segment-> link + $2
discourse segment headed bydiscourse segmentindependent discourse connective-> discourse graph with semantic head
19
Mapping from Discourse Tree to Discourse Graph (cont.)
[ [Semantic annotation is necessary {-inference because [conjunction [2 retrieval must be quick and easy] and [3 authoring of content must be easy]]}.] [2 Retrieval must be quick and easy {purpose in order to [0 realize ubiquitous information access]}.] [3 Authoring of content must be easy {-inference because [1 huge amount of content is necessary {purpose in order to [0 realize ubiquitous information access]}]}.]]
20
authoring of content must be easy
authoring of content must be easy
semantic annotation is necessarysemantic annotation is necessary
inferenceinference
huge amount of content is necessary
huge amount of content is necessary
retrieval must be quick and easy
retrieval must be quick and easy
realize ubiquitous information servicerealize ubiquitous information service
purposepurpose
purposepurpose
inferenceinference
Outline
IntroductionDiscourse Graph: Semantic
StructureDiscourse Tree: Syntactic
(Presentational) StructureDiscourse Relation
21
Minimizing Discourse Relation Set
Concentrate on semantics.abstract away presentational aspects
importance or headedness (nucleus/satellite distinction)
maximize versatility of semantic representation (discourse graph)
Maximally accommodate polymorphism.
22
Discourse Relations Are Semantic
Also elaborations such as detail are semantic relations because they are defined in terms of semantic entailment, etc.
23
This is an old story.
We' re talking about years ago before anyone heard of asbestos having any questionable properties.
detaildetail
ImportanceAbstract importance (nucleus/satellite
distinction) away from discourse relation.Importance (what is semantic head) is
attributed to syntax (presentation) of discourse tree.
[1 {conflict Although [its rooms are small]}, the hotel is large.] [{So1} Tom will stay there].
[2 {conflict Although [the hotel is large]}, its rooms are small.] [{So2} Mary won’t stay there].
the hotel is largethe hotel is large
its rooms are smallits rooms are small
conflictconflict
Tom will stay thereTom will stay there
Mary won’t stay thereMary won’t stay there
inferenceinference
inferenceinference
symmetricsymmetric
24
Importance (cont.)
Unification between inverse relations:means vs. purposecause vs. resultreason vs. conclusionattribution vs. contentgeneral vs. specificwhole vs. part
Any criterion under which to choose names and directions of these relations?e.g., purpose is better than means
because purpose is in the same direction as causal and temporal ordering. 25
Polymorphism, Metonymy, and Projection
Object/Eventualitysimilardissimilargeneral-specificset-memberwhole-partexamplerestatementcomparisonattribution-contentmeans-purposecomment-topic
Instance/Typepurposeconditionalunconditional
Semantics/Pragmaticsenablementinference
Temporal Projectioncircumstancebefore-afteruntilsimultaneous
dom
ain
=ra
ng
e
26
Object/Eventuality
Some relations concern not only eventualities but also objects.comparison
[Tom is taller {than Mary is tall}].attribution-content
[I believe {that he’s right}].[the belief {that he’s right}]
means-purpose[cut it {with this sword}][cut it {by using this sword}]
27
Instance/Type
Some relations may concern both instances and types of eventualities.
[Use the sword {purpose to [cut it]}].
[I worked hard {purpose to [pass the exam]}].
[{conditional If [you’re going to school]}, it’s eight o’clock].
28
Semantics/Pragmaticsenablement
[{-enables Since [here’s coffee]}, it’s possible that [you drink it].]
[1Here’s coffee.] [{-enables So1} drink it].The presence of coffee provides the precondition for the imperative.
enablesenableshere’s coffeehere’s coffee drink it.drink it.dialog actdialog act
enablesenableshere’s coffeehere’s coffee you drink ityou drink it
prop. contentprop. content
29
Temporal Projection
[Tom came {at 8 o’clock}].[Tom came {when Mary came}].
time(semantic
role)
time(semantic
role)
circumstance(discourse
rel.)
circumstance(discourse
rel.)
equality orProjection?equality orProjection?equality orprojection?equality orprojection?
`time’ and `circumstance’ may be unified.
30
Taxonomy
Ted Sanders’ 3 (out of 4) dimensionsadditive vs. causalpositive vs. negativefactual vs. inferential
Cf. the other dimension concerns linear orderbasic vs. non-basic
31
additivepositive
Elaboration: specific, part, step, object, member, example, extraction, minimum, detail, equivalent, definition
Attribution: contentBackground: background, circumstanceComparison: similar, proportionComplement: supplementAdditive: conjunction, additionManner: manner
negativeContrast: contrast, dissimilar, disjunction, substitutionComplement: constraintComparison: comparison, preference
causalpositive
Causality: causes, motivates, triggersEnablement: purpose, enablesInference: inference, explanationEvaluation: evaluation, interpretation, commentCondition: conditional
negativeConcession: conflictCondition: otherwise, unconditional, compromise
32
Relation Naming ConventionSubject and object of transitive verb
are $1 and $1.causes, motivates, triggers, enables
Relational noun indicates $2.part, member, example, detail, content,
constraint, evaluation, background, mannerCollective adjective or noun is
symmetric relation.equivalent, similar, conjunction,
disjunction, contrast, conflictObject of preposition or conjunction is
$2.after, letAlone
This is not only about discourse relations but about any relations. 33