Semantics reloaded

Steffen Staab Semantics Reloaded 1Institute for Web Science and Technologies · University of Koblenz-Landau, Germany
Web and Internet Science Group · ECS · University of Southampton, UK &
Semantics Reloaded
Steffen Staab
@ststaab
http://west.uni-koblenz.de
http://wais.soton.ac.uk

Steffen Staab Semantics Reloaded 2
• is the linguistic and philosophical study of meaning,
in language, programming languages, formal logics,
and semiotics.
• It is concerned with the relationship
between signifiers—like words, phrases, signs,
and symbols—and what they stand for,
their denotation.
From Wikipedia
Semantics

My Team@Institute for Web Science and Technologies
Deduction
Semantic Web
RDF
OWL
Commonsense
Forgetting
Programming
Induction
Social Media
Text, RDF
Sensors
Transfer Learning
in Networks
Argumentation
Deep Learning
LDA, FCA, TDA,
Responsibility
Interaction
Eye Tracking
Multimodal
Browser
GazeTheWeb
Gaze Mining
Semantic Data Data Analytics Semantic Interaction

What is a cup?

Let‘s grab for semantics
whereever we can find it!
What is a cup?

“For a large class of cases of the employment of
the word ‘meaning’—though not for all—this
word can be explained in this way:
the meaning of a word is its use in the language”
Wittgenstein, Philosophical Investigations

Deduction
With M. Leinberger, R. Lämmel

• is the linguistic and philosophical study of meaning,
in language, programming languages, formal
logics, and semiotics.
• It is concerned with the relationship
between signifiers —like words, phrases, signs,
and symbols—and what they stand for,
their denotation.
Semantics

Waterfall development
Conceptualization
Conceptualization
Database
Code
strings are passed
back and forth
here
What can we learn
about these
„strings“?

∃recorded.Song ⊑ Musician
Actor ⊓ Musician ⊑ MusicalActor
hendrix, machineGun : recorded
machineGun : Song
elvis :Musician
elvis :Actor
hendrix, elvis : influencedBy
T-Box
A-Box
Example ontology and data
We consider a powerful ontology language, thus
we can handle simpler ones, too.

Description Logics
• Atomic elements of a dataset (knowledge base) 𝒦
defined in signature 𝑆𝑖𝑔 𝒦 = (𝒞, 𝒬, 𝒪)
– Concept identifiers, e.g., Musician
– Role identifiers, e.g., recorded
– Object identifiers, e.g., hendrix, beatles
• Concept expressions built using connectives
– Intersection: Musician⊓Painter
– Existential Quantification: ∃recorded.Song
– Inverse roles: ∃recorded−. ⊤
– Concepts through enumeration: beatles

Knowledge base and inference
• Knowledge base 𝒦: Schema and data
– Subsumption: MusicGroup ⊑ Musician
– Concept assertions: beatles : MusicGroup
– Role assertions: hendrix, beatles : influencedBy
• Interpretation-based semantics
– Formula true or false in specific interpretation
– If all formulas of 𝒦 are true: Model of 𝒦
• If formula 𝐹 true in all models: 𝒦 ⊨ 𝐹
– Answering queries: 𝒦 ⊨ ?X ∶ 𝐶 = ? X 𝒦 ⊨ ?X ∶ 𝐶 }

• Mixture of nominal (Musician) and
structural typing (∃recorded.Song).
• Lack of formal conceptualization:
– e.g. influencedBy
• Logical reasoning
• Number of concepts:
>1,148,230 different concepts in Wikidata
Issues arising

Generic representations
• Only types:
– Node
– Edge
– Axiom
– ...
• Implication
– Typing statements are manually programmed, e.g.
• If x:Node and hasType(x,Person) and hasName(x,y)
then print(“Person:“ y)
If x:Node and hasType(x,Company) and hasName(x,y)
then print(“Company:“ y)
Related Work: Generic Representations
No static typing!

Related Work: Mappings
Conceptualization
Database
Code
„strings are
passed back and
forth here“
Code/mapping
generation
ActiveRDF
Liteq
Owl2Java
...
Issues
- Queries imply nominal and
structural typing
- Large number of possible
types
[J. Pan et al, 2013]

Public class Influences {
static RDFNode MUSICIAN = ...
static Property INFLUENCED_BY = ...
private static String getInfluence(Resource r){
return r.getProperty(INFLUENCED_BY).getObject().toString();
}
public static void main(String args []) {
Model model=... ;// load datasource
for(Resource musician :
model.listSubjectsWithProperty(RDF.type,MUSICIAN))
System.out.format(„%s was influenced by %s“,
musician.toString(),
getInfluence(musician));
}}
Jena style program – with error
∃influencedBy
is not a subclass of
Musician !

Erroneous program in JavaDL
import static semantics.util.names; // helper for converting to IRI
public class Influences knows “music.rdf“ {
private static String getInfluences(∃«:influencedBy».⊤ artist) {
return String.join(“ “, names(artist.«:influencedBy»));
}
// Query for all music artists and print their influences
private static void main(String args[]) {
for («:Musician» m : query-for(“:Musician“))
System.out.format(“%s was influenced by %s“, m.getName(),
getInfluences(m));
}
}
Static typing finds that
∃influencedBy is not a subclass of Musician !

Type-checked program in JavaDL
import static semantics.util.names; // helper for converting to IRI
public class Influences knows “music.rdf“ {
private static String getInfluences(«:Musician» artist) {
switch-type (artist) {
∃«:influencedBy».⊤ influencable { return String.join(“ “,
names(influencable.«:influencedBy»));
}
default: “no influence known“
}
}
// Query for all music artists and print their influences
public static void main(String args[]) {
for («:MusicArtist» m : query-for(“:MusicArtist“))
System.out.format(“%s was influenced by %s“, a.getName(),
getInfluences(m));
}
}

Implementation
Compiler
Reasoning
Service
(HermiT)
Semantic
data
Standard
Java
Compiler
Extended
syntactic
forms
queries during
type-checking
Extended
Java Code
loads
compiled with
Standard
JVM
Bytecode
produces
queries during
runtime

𝝀 𝑫𝑳 in a nutshell
and including powerful type inference
Core principles & Language constructs
(Leinberger, Lämmel, Staab; ESOP 2017)

1. Use concept expressions as types
– Extended syntax for types (e.g., MusicArtist ⊓ Painter)
2. Subtype inferences
– Forward subtyping of concepts expressions (𝐶 ⊑ 𝐷) to 𝒦
– E.g., λx:MusicArtist. … (beatles as MusicGroup)
3. Typing queries
– Use concept expression queries
(e.g., query MusicArtist⊓Painter )
– Check for satisfiability
– Queries always return lists
Core principles

• Subtyping: Additional rule for concept expressions
• Abstraction, application, recursion not affected
– Simple, static types with well defined subtyping behavior
• if-then-else, cons, … need join-type (least upper
bound)
– Separation between normal types and concept types
– Straightforward due to disjunction: 𝑙𝑢𝑏 𝐶, 𝐷 = 𝐶 ⊔ 𝐷
– Same for greatest lower bound
𝝀 𝑫𝑳 rules for static typing

• Typing of objects with most specific concept
– Concepts via enumeration make it straightforward:
beatles ∶ {beatles}
• Queries are straightforward: query MusicArtist
• All songs recorded by beatles: beatles.recorded
– Satisfiable, but empty query result
– head nil is exception to type safety
Objects & Queries

• All influences of hendrix: hendrix.influencedBy
– Result type: ∃influencedBy−
.{hendrix} list
• Typecase to allow for down casting:
case (head hendrix.influencedBy) of
type Painter as x → …
type ¬Painter as y → …
default …
• Not known for beatles if painter or not
– Default case necessary to not get stuck
Down casting

Theorem: A well-typed closed term does not get stuck
during evaluation (with common exceptions).
Result for DL
Typing is a safety net,
but does not solve the halting problem
(empty list)

• Type inference
– Not possible in Java, but in modern languages
• SPARQL BGP queries
• Epistemic concept expressions:
K ∃«:influencedBy»
– Class of instances whose influencers are known
• Shape constraints: shacl
Outlook: DFG Project LISeQ granted

Induction
With Jun Sun, Jerome Kunegis
and the previous teams of ROBUST and REVEAL
(Sun, Staab, Kunegis;
Submitted to IEEE Computer Special Issue on Web Science)

• Benefit from Experience with Social Networks
• Early response to
– trolls
– attacks
– spam
• Social networks are easy to ruin!
What do we want: Healthy social networks

• Role: two nodes belong to the same role if they have
similar structural behavior
• Using structural features of nodes for classification
Roles in Social Networks

• Idea: to learn knowledge from a domain (source
domain) and apply it to another domain (target
domain) using power law
• Challenge: feature distributions differ between the
source and target domains
Transfer Learning

Cumulation (Quantiles)
Transfers
Value transfer from one to another power law

Degree Distribution vs.
Transformed Degree Distribution

Transfer Learning Procedure
• Step 1 - Feature Extraction
• Step 2 - Feature Transformation
• Step 3 - Feature Aggregation
• Step 4 - Classification

Related Work
None: Lower Baseline
• Applying source-trained classifier on target without transfer
SVD
• Agirre and De Lacalle (ACL2008) use SVD for feature
transformation for word sense disambiguation in different domains
TrAda
• Dai et al. propose TrAdaBoost using partially labelled data from the
target network.
TraNet
• Our approach
Trad.: Upper Baseline
• Training and evaluating on the target network

Evaluation
Target dataset:
• Software AG ARIS Community user interaction
• 9566 threads and 20538 comments by 4216 people

Evaluation:
Sources to Target Software AG Aris

14 Wiki-talk social networks (different languages)
• Registered uses who discuss with each other
• At least 25 users marked as administrators
Evaluation: Wiki talk data sets

Wiki Talk: A set of user interactions in
different Wikipedias
SVD and TrAda omitted due
to poor performance
Network properties
– power laws –
are key for
transfer learning

Wiki Talk: A set of user interactions in
different Wikipedias
Semantics of a
concept „trusted“
relative to context

• How to better describe each node
– Algebraic topology
• How to apply to RDF and knowledge graphs?
Outlook:
EU Project Cutler just started
EU Project Co-inform about to start

Interaction
With R. Menges, C. Kumar, K. Sengupta

• is the linguistic and philosophical study of meaning, in
language, programming languages, formal logics, and
semiotics.
• It is concerned with the relationship between signifiers —
like words, phrases, signs, and symbols (web
pages!?) —and what they stand for, their denotation.
• Semiotics is the study of meaning-making, the study
of sign process and meaningful communication.... The
semiotic tradition explores the study of signs and
symbols as a significant part of communications. As
different from linguistics, however, semiotics also studies
non-linguistic sign systems.
Semantics

Eyetracking

Experiment setting
A Search task B Result list C Selection Foto
Suche eine
braune Kuh
pictureskeyword

Labeling a region
„brown cow“
Kuh
Kuh
Kuh

Eye tracking data of different subjects
„brown cow“

MAMEM https://youtu.be/42yGmr3NE0k

• What are eye gaze patterns
– In dynamic environments
• Scrolling
• Drop-down menus
• Rotating web banners
• ...
– Over many people?
Outlook: KMU Innovativ GazeMining just started
Digital Imagination Challenge Finale
Berlin, 15. Februar 2018

Find patterns in eye gaze data
– Layers of presentation activities
• Fixed elements
• Carousels
• Canvas
• ...
– Layers of different users
Visualize resulting analysis for intuitive understanding
Outlook: KMU Innovativ GazeMining just started

Conclusion

“For a large class of cases of the employment of
the word ‘meaning’—though not for all—this
word can be explained in this way:
the meaning of a word is its use in the language”
Wittgenstein, Philosophical Investigations

Steffen Staab Semantics Reloaded 54Institute for Web Science and Technologies · University of Koblenz-Landau, Germany
Web and Internet Science Group · ECS · University of Southampton, UK &
Thanks to my team members and all
the other collaborators:
Martin Leinberger, Ralf Lämmel, Raphael
Menges, Daniel Müller, Chandan Kumar,
Korok Sengupta, Jun Sun, Jerome Kunegis,
Tina Walber,...
Project teams:
MAMEM, ROBUST, REVEAL

Semantics reloaded

Recommandé

Recommandé

Contenu connexe

Tendances

Tendances (20)

Similaire à Semantics reloaded

Similaire à Semantics reloaded (20)

Plus de Steffen Staab

Plus de Steffen Staab (20)

Dernier

Dernier (20)

Semantics reloaded