Second day of the FOIS conference. At the beginning there was an opening on how to make the semantic web more semantic by Peter Gandenfors. Nice ideas on how to use conceptual spaces in a practical context. On the forth session there were two interesting contributions: the first by Martin Raubal, on an applied approach on building an ontology with a formalization of a conceptual space and applied to pedestrian navigation. The second by Massimo Poesio on an empirical prospective for shaping an ontology parsing text from google or wordnet.

The running notes in the extended.

TITLE OF PAPER: FOIS Formal Ontologies in Information Systems
PRESENTED BY: misc

CONFERENCE:FOIS
DATE: November, 5, 2004
LOCATION: Torino Incontra, Centro conferenze, Torino, Italy

—
REAL-TIME NOTES / ANNOTATIONS OF THE PAPER:

Invited speaker II: Peter Gardenfors
How to make the semantic web more semantic

The semantic web is an extension of the current web in which information is given well-defined meaning, better enabling computer …

The semantic web is not semantic: at best it is ontological; taxonomy +inference rules; Berners-Lee, Hendler and Lassila: fortunately most of the information we want to express is along the lines of a hex-head bold; unfortunately is not. OWL adds more vocabulary for describing properties and classes, among other relations between classes, cardinality, equality … That is exactly what is to be expressed of a language with the expressivity of first order logic where concepts are defined in terms of sets of objects. The symbols are not grounded.

There is more to semantics: humans often categorise objects according to their similarity. Similarity is not easily expressed in existing web ontology languages. Concepts show prototype effects. The mechanics of concept combination do not follow principles of first order logic ->

Intersection of extension is not sufficient. Non-monotonicity of concept combinations.

Metaphors and metonymies are ubiquitous.

Conceptual Spaces as tool to analyse this: information is organised by quality dimension, that are sorted into domains (shape, time, temperature, …), dimensions within domins are integral (they comes in groups). Domains are endowed fy a certain topology or metrics. Similarity is represented by distance in a conceptual space.

The notion of domain is central to me and replace the concept of category.

Properties vs. concepts -> Properties: A convex region in a single domain. Concepts: A set of convex regions in a number of domains and informations about the regions in different domains are correlated. Regions can have fuzzy boundaries. Convexity: anything is between two points is the same.

Categorization in conceptual spaces: Voronoi tessellation around prototype objects divides conceptual spaces into categories based on the nearest neighbour rule, i.e. each object is associated with the prototype closest to it.

Conceptual spaces as tool for the semantic web: conceptual hierarchies (taxonomies) will emerge from the inclusion structure of the concept regions of a domain (e.g., via Voronoi tessellation). Identities of concepts corresponds to identities of regions. Identities of names given by identities of points (more complex when vectors are incomplete). Properties characteristics (e.g., transitivity and symmetry) are implicit in the topology or geometry of the domains. Explicit inferences based on additional rules become superfluous.

The role of similarity for the Semantic Web: if a search for aan item in a database fails you want to find things that are similar. But similarity is dependent on the user’s interests and on context. This is difficult to handle in taxonomy.

Concepts are sensitive to context: “heavy book” – “heavy smoker”. The effect of contrast classes: red-of the color of fresh blood, rubies, human lips, the tongue, maple leaves in the autumn.- We need a contrast class to define combined word. We are very good in doing the mapping between two classes.

Computational issues: traditional methods involve rule following based on recursive procedures on tree-like structures. Conceptual spaces are based on vectorial representations …

Help needed!: find small domain with identifiable dimensions suitable for web application; describe concepts …

The problems are not new: men are far enough from … John Locke.

Session 4: Semantics and Cognition
<> Martin Raubal (ifgi) -> Formalising Conceptual Spaces.
Case study: city of vienna, finding your way. The assumption is that the landmark system is the same of the landmarks of the user. Needed more adaptation to the user’s semantics.

Cognitive semantic: effort to solve semantic interoperability problems => realist semantics. Problems: learning as extension of conceptual space through new quality dimension.

Formalization: conceptual vector space is the set of vectors representing quality dimensions. Ideally a basis but hard to achieve. Multi-domain concepts. Semantic distances and weights: Euclidean distances between points. I use methods derived from statistics.

Case study: Facades of building as landmarks; concepts of facades represented by different variables; utilise conceptual vector spaces => capture difference the system and the user’s view of the user.

The system view contains different factors: area, shape factor, shape deviation, color, visibility, cultural importance, identifiability by signs. In the user view the cultural importance factor does not have the same weight. During night time, we found that people select different landmarks. Weights of factor changed.

Final goal is bridging semantic gap between system’s view and user’s concepts. We can define a mapping between the user view and the system view. Contribution to formal representation of cognitive semantics.

<> Nicolas Asher -> A case study: the genetive construction.
Natural Language Metaphysics: -natural language examples reveal ontological commitments; -lexical semantics and composition logic as tupe driven; – by paying attention to issues like logical metonymy and predications, we found commitments to a wealth of types under ‘e’ (the type of entity) and to complex types beyond the functional tupes of standard type theory.

<> Massimo Poesio -> Feature-based vs. Property-based KR: An empirical perspective.
Linguistic evidence and concept representation: work on lexical acquisition/concept clustering … Is the notion of feature or attribute useful for concept description?
Concept clustering (aka: automatic taxonomy discovery). The representation of nominal modifiers: in formal semantics nominal modifiers such as adjectives are typically represented in a uniform fashion as … Using text patterns to build concept descriptions, II: attributes. He used an interesting methodology to query Google and look for clusters of word features. He reported several examples of other studies that went in that direction. In a second set of experiments he used Wordnet for interrogations.

You need a methodology to define which of the words you get from this techniques can be defined as an attribute.

<> Olivier Bodenreider (national library of medicine) -> The ontology-Epistemology Divide: A Case study in Medical Terminology.

Biomedical terminology collects the names of substances, qualities and processes. There are many biomedical vocabulary: International Classification of Disiases; SNOMEST CT; Gene Ontology.

Terminology vs. Ontology: various kinds of structures (simple list of mames, thesauri, ontologies). To draw the line between ontologies and epistemology: ontology contains invariants or a theory of reality. The epistemology the knowledge about such entities, therefore the perception of reality. The aim of this talk is that when we look at names in these systems biomedical terminology we found lots of things that were ment to designate non-ontological features.