Integrating Folksonomies with the Semantic Web

Lucia Specia and Enrico Motta

Abstract

• Tags ?
• serve primarily for indexing purpose.
• facilitates search and navigation of resources.
• yields a collective classification schema when the same tags are used by more than one individual can .

​

• This paper presents an approach for making explicit the semantics behind the tag space in social tagging systems, so that we can define concepts and ontologies.
• focuses on the relationships amongst tags and their mapping into formal concepts in ontologies
• How is it achieved? - shallow pre-processing strategies and statistical techniques along with ontologies available on the semantic web.

​

​

Introduction

• Social tagging/bookmarking systems (Flickr, Del.ici.ous) have resulted in a crowdsourced evolving ontology - Folksonomy.
• Key elements of such a system: Users, Resources, Tags.
• Paper deals with collective purpose of tags and how they relate with the resources, users and themselves.
• Tag disambiguation, Visualization (clustering), Tag suggestion (similarity) are key applications that result from the ontology.

Problems

• ambiguity
• lack of synonymy and
• discrepancies in granularity

Introduction..

• This can also be used for ontology evolution and population.
• ontologies can be used to structure folksonomies semantically
• Dynamic knowledge from folksonomies can be used for ontology evolution.
• How this can be done?(steps)

Related Work

• Search using similar images(Rely on context) (Aurnhammer et al.)
• Subsumption-based model (Schmitz, 2006)
• Probabilistic models (Wu et al.)
• Tripartite model (Mika)
• Association rules (Schmitz et al. as seen earlier)
• Previous methods lack good tag pre-processing methods.
• Better strategies to cleanup tags needed.

Integrating Folksonomies with the Semantic Web - what does it mean?

In short, the paper deals with integrating structure created by the people who use social bookmarks to come up with a clear system that matches content to topics on the basis of the underlying semantics.

​

​

Datasets

• Similar to the paper that introduced the concept of Folksonomy, this paper also uses the data from Flickr and del.icio.us.

​

• Reason: abundance of users, availability of semantically tagged content.

​

​

Methodology

• Unsupervised - does not assume any previously defined mapping/tagging.
• Three key steps:
• Pre-processing,
• Clustering
• Concept/Relation identification.

Pre-processing

• Filter out unusual tags (tags that do not fit a certain definition -- noise removal)
• Use Levenshtein distance to group morphologically similar tags (plurals, misspellings, special characters)
• Filter out infrequent tags - those that have frequency less than a threshold - outlier removal.

Clustering

• Performed to identify groups of similar tags.
• Similarity is defined based on co-occurrence (similar to collaborative filtering)
• Uses angular separation as the distance measure:

Interpreting clustering results

• Similar tags get ranked close to each other and they show some latent semantic similarity (audio,music,mp3 all got clustered together).
• Improving clusters: by removing all subsets, by merging clusters with minimal differences.

Concept and Relation Identification

• Post each possible pair of tags to existing sources like Wikipedia or other search engines.
• Tags that are not found together (those that don’t return any search results) are eliminated.
• If tags are found together use it as evaluation metric. Once tags are found together build ontology - hierarchy, ancestry.
• Personal comment: This use of external resource seems computationally expensive and recursive.

Experiments and Discussion

Experiments and Discussion

• Relationships between tags in a cluster
• Arrows representing subclass relationships

​

Conclusions and Future Work

• Integrated semantic web and folksonomies
• Experiments performed on data from Flickr and del.icio.us
• Meaningful clusters of tags created.
• Better clustering techniques can be used
• Using semantic web search engines to build better models in future.