In this work, we consider the problem to personalize user's Web searches for improving the quality of results. To this aim, we propose a preliminary methodology that allows to de¯ne a conceptual user pro¯le based on the YAGO ontology.
To overcome the limitations of the \one size ¯ts all" approach of search engines,
personalized approaches to Information Retrieval have been proposed.
Personalized search is based both on modeling the user's context by a user's pro¯le that
represents the user's preferences, and on the de¯nition of processes that exploit
the knowledge represented in the user pro¯le to tailor the search outcome to
users' needs. The accurate de¯nition of a user pro¯le plays then a central role
to de¯ne e®ective approaches to personalization. Up to now, bags of words, and
vectors or graph-based representations have been mainly used to de¯ne users'
pro¯les. To improve the quality of the knowledge represented in user pro¯les, in
some recent works, external knowledge sources (i.e., WordNet [
A user pro¯le is de¯ned based on the analysis of the information characterizing
the user's interests and preferences. Elicitation of user's interests and preferences
is not the focus of the research reported in this paper. Numerous approaches have
been proposed in the literature to this aim [
The novelty of the research reported in this paper is to employ the YAGO [
USER
Phase 1 Documents
Phase 2 Analysis and
Terms Extraction
Phase 3 Personalized Knowledge Extraction from YAGO
Phase 4 Conversion to RDFS and Editing with
Protégé Phase 1. This ¯rst phase consists in individuating the user's knowledge that has to be considered to extract the user's interests. In this speci¯c case, we analyzed a set of documents collected by the user and stored in his/her personal computer. Phase 2. Each document is analyzed in two steps: (1) document preprocessing and (2) term frequency analysis, respectively. In the ¯rst step, standard text processing techniques are applied such as stop-word removal, and stemming. In the second step the open source software Lucene is used for indexing the documents; a standard normalized Tf-Idf formula is adopted to compute the index terms weights, but other approaches will be taken into account for further investigations.
Phase 3. The outcome of the previous phase is a list of interest-terms with index terms weight over a given threshold ®. To enrich the knowledge of the user's interests a process of knowledge extraction from the YAGO ontology is performed. This process is articulated in 3 sub-phases: (1) individuals and facts extraction, (2) direct concepts extraction and expansion to their child nodes, and (3) addition of new synonyms, respectively.
The fact extraction process is logically divided into non-taxonomic and taxonomic relations extraction. Non-taxonomic relations are de¯ned in the YAGO ontology over entities which are referred to as individuals, while taxonomic relations can hold between an individual and its parent concept (class), or between two concepts. As previously stated a fact is a triple de¯ned as < entity; relation; entity >, so the ¯rst step of the algorithm consists in locating the facts where an interest-term (obtained based on phase 2) matches with an entity. The outcome of this step is constituted by a set of facts and entities extracted from YAGO. From the analysis of the taxonomic relation di®erent considerations have been made. In fact, it is possible that some facts based on the taxonomic relation SubClassOf do not report useful information with respect to the considered term. For example, the fact < relational database systems; SubClassOf; database systems > contains the knowledge that \relational database systems" are sub-class of \database systems", which is not very informative. For this reason, in case of a direct concept match, the algorithm takes all the ¯rst level children (individuals) of the matched concept. Referring to the previous example, for the term \relational database systems" the following instances MySQL, Oracle, PostgreSQL etc., will be added in the user's pro¯le. A possibility is that the term is not found in YAGO. When this happens, our algorithm analyzes WordNet for checking the existence of synonyms. In case multiple synset exist, we adopt the methodology used by the authors of YAGO, where the most probable synset (i.e., the synset having higher probability of occurrence) is selected.
Phase 4. At this step, the resulting personal ontology is converted into the
ontological language RDFS 1, and its graph portions are visualized by the ontology
editor Prot¶eg¶e [
Preliminary Experiment A preliminary analysis has been made for de¯ning a conceptual user pro¯le based on the YAGO ontology by considering 35 documents. This set of documents is related to several user interests such as art, literature, music, cinema and work. The second phase of the proposed methodology has been conducted by using Lucene, and the threshold for scoring was 1 http://www.w3.org/TR/PR-rdf-schema set to 0:5, thus obtaining 306 terms. At the end of phase 3, 578950 entities (i.e., individual plus concepts) have been counted in the user pro¯le, where 11 new terms are added from WordNet. The last phase has consisted in converting the obtained pro¯le in an ontological language (i.e., RDFS) in order to improve it by, for example, reducing noise or adding relations between terms. For example, if a term was related to the actor \Brad Pitt", all the corresponding information de¯ned in YAGO are extracted such as categories it belongs to (i.e., Action ¯lm actors, American male model, . . . ), as well as its non-taxonomic relations (i.e., hasWonPrize, produced, actedIn, . . . ). By editing this ontological pro¯le in Prot¶eg¶e, the user is allowed to delete non relevant information, for example the ones related to Brad Pitt as a model. 3
The aim of this work is to create user pro¯les based on the YAGO general purpose ontology, to the aim of Web search personalization. We believe that ontologies are worth to be investigated as an interesting support for structuring knowledge in user pro¯les. To this aim, in a ¯rst preliminary application, the documents collected by a user are considered as the evidence of his/her interests. We plan to improve the methodology presented in this paper by following three main directions: the ¯rst is to automatically remove some noise from the pro¯le (e.g., by deleting non relevant entities and the relations involving them), the second is to add new relations and facts between terms not de¯ned in YAGO, and the last one is to consider other sources of information (i.e., past user's queries) to extract user's interests. Furthermore we will test the obtained YAGO-based user pro¯le for expanding the user's queries to contextualize his/her Web searches.