In the context of Semantic Web and Web 2.0 environments, finding an appropriate content is regarded not only as a problem of information overload but also as a problem of Web personalization [1], which deals with personalizing content retrieval and access with respect to a specific user model. Moreover, this large volume of data makes impractical or even impossible several manual activities such as extracting small portions of relevant information from available contents, or classifying contents according to a specific model of user interests [2]. As a consequence, the gap between the performance of traditional information retrieval tools (e.g. search engines) and the user satisfaction in their use continues to grow. In order to alleviate this issue [3], more sophisticated approaches and tools become necessary for providing personalized content recommendations and classification. Furthermore, in a world of collaborative publishing we have to take into account e-Learning, knowledge management and Web 2.0 as typical application environments. Indeed, we can discover new relevant information by looking the community of people that, for example, share a common set of documents or use the same tags to label them. In this wider setting, automatic text classification remains a significant research field with several challenges such as:
-Associating rich and precise semantics to information contents. For describing an object, people tend to assign to it a very small number of tags, based on their knowledge background; of consequence, same tags, used by different users, do not share a common semantics [4,5]. -Adapting information retrieval strategies to an evolving user model, providing run-time malleability to end-users [6]. Certainly, continuously updating a user profile is more difficult than building a single static representation, and requires the availability of some forms of user feedback to keep synchronized the model. -Finding relationships between contents and using a uniform method to share and reuse tagging data amongst users or communities [7]. The topicality criteria alone may not be sufficient to relate contents when there is no shared semantics for a tag.
Our main goal in building the PIRATES framework is to empower social bookmarking tools, allowing users to easily add new contents in their personal collection of links, automatically supporting them when categorizing by means of keywords (tags) in a personalized and adaptive way. This work is a first step towards the generation and sharing of personal information spaces described in [8]. We have designed PIRATES keeping in mind several applications where it can provide innovative adaptive tools enhancing user capabilities: in e'learning for supporting the tutor and teacher activities for monitoring (in a personalized fashion) student performance, behavior, and participation; in knowledge management contexts (including for example scholarly publication repositories and digital libraries [9]) for supporting document filtering and classification and for alerting users in a personalized way about new posts or document uploads relevant to their individual interests; in online marketing for monitoring and analyzing the blogosphere where word-of-mouth and viral marketing are nowadays more and more expanding and where consumer opinions can be listen. The paper is organized as follows: Section 2 illustrates the overall architecture and operation of PIRATES; Section 3 describes a typical interaction session and Section 4 concludes the paper.
2 The PIRATES framework PIRATES (Personalized Intelligent Recommender and Annotator TEStbed) is a general framework for text-based content retrieval and categorization and exploits social tagging, user modeling, and information extraction techniques. Rather than proposing a rigid classification toolset, we have developed a testbed platform for integrating (and experimenting with) various tools and techniques, providing an interactive environment where users can customize the way they retrieve and classify information on the Web. The main feature of PIRATES concerns a novel approach that automates in a personalized way some typical manual tasks (e.g. content annotation and tagging). The framework operates on a set of input documents stored in the Information Base (IB) repository and suggests for these some personalized tags and other forms of textual annotations (e.g. key-phrases) in order to classify them. The original documents are then annotated with these tags, forming the Knowledge Base (KB) repository. Personalization is achieved exploiting user profiles (which represent the user interests), personal ontologies, personal tags, etc., as discussed in Section 3. Furthermore, PIRATES provides several mechanisms of user feedback that helps to provide personalized adaptive information.
The PIRATES architecture is illustrated in Figure 1. On the left-hand side, all the possible input sources are shown: single textual documents, specific IB repositories which can be contained within an e-learning knowledge management environment, and the Web, with specific (but not exclusive) focus on Web 2.0 portals, social networks, etc.. The right-hand side shows the suggested annotations and the resulting KB repository. The main modules of PIRATES are:
-IEM (Information Extraction Module), which is based on the GATE platform [10] to extract named entities, adjectives, proper names, etc. from input documents, contained in the IB. -SAT (Sentiment Analysis Tool ), which is a specific plug-in for personalized sentiment analysis (typically to be activated for online marketing applications), that is capable of mining consumer opinions in the blogosphere and classify them according to their polarity (positive, negative, or neutral) [11]. -KPEM (Key-Phrases Extraction Module), which implements a variation of the KEA algorithm [12] for key-phrases extraction. KPEM identifies n-gram keyphrases (typically n between 1 and 4) that summarize each input document. This information is provided to the user, and is also given as input to the subsequent modules. -ORE (Ontology Reasoner Engine), which suggests new abstract concepts by navigating through ontologies, classification schemata, thesauri, lexicon (such as WordNet), etc. An abstract concept is identified by looking for a match between the annotations found by the other modules (IEM, KPEM, IFT, and STE) and the concepts stored in ontologies. When a match is found, ORE navigates through the ontology, looking for the common parent node which represents the more abstract term to suggest as annotation. ORE also assists users in creating personal ontologies with techniques similar to those described in [13].
-IFT (Information Filtering Tool ), which evaluates the relevance (in the sense of topicality) of a document according to a specific model of user interests represented with semantic (co-occurrence) networks [14]. -IFT Web Agents, which continuously monitor the Web (and the blogosphere) looking for new information, cooperates with IFT to filter contents according to the user model, and updates the IB repository. IFT and its Web agents form together the Cognitive Filtering module discussed in [8]. -STE (Social Tagger Engine), which suggests new annotations for a document relying on aggregated tags, i.e. the user's personal tags (tags previously exploited) and the more popular tags used by the community of people that classify the same document in social bookmarking sites such as Del.icio.us1 , Faviki2 or Bibsonomy3 . This social information is integrated with contentbased analysis techniques as discussed in [15].
In this section we provide a typical scenario that illustrates a use case for our framework. Consider a user interested to read scientific publications in the area of software engineering. He trains the IFT tool providing the training data (e.g. 2-3 relevant papers in the field, some keywords and a short textual description for the argument) in order to setup the user model. After training, the IFT agents periodically monitor the Web (in our case especially Web 2.0 sites such as Del.icio.us, Bibsonomy, CiteseerX4 , etc.), download new content and scrap selected data from them to filter out irrelevant information (e.g. ads and navigational links). When a relevant content (with respect to the user model) is retrieved, the agents add it to the IB repository and informs the user with a notification (e.g. an e-mail message). This information retrieval workflow has been already discussed in [14,16], so in the rest of the section we concentrate on the classification features added by the PIRATES framework. Indeed, PIRATES aims expressly to support the user in organizing the IB repository, easing the work of classifying new contents by means of personalized tag suggestions. Suppose now that an IFT agent notifies (among the others) the paper "A UML Class Diagram Analyzer" 5 . In order to classify this new content, the user can enable some PIRATES annotator modules, as illustrated in the left side of person's names, organizations, and places (using IEM); keyphrases, i.e. n-grams long three terms at maximum (using KPEM).
With these settings, the framework produces the tag recommendations showed in the right side of Figure 2. In particular, the suggested tags concern persons such as the authors (Tiago Massoni, Rohit Gheyi, and Paulo Borba) and the people acknowledged in the paper (Bordeau, Chang, Augusto Sampaio, Franklin Ramalho and Rodrigo Ramos), locations (Brazil), and organizations cited in the text (the Informatics Center of the Federal University of Pernambuco, the Software Productivity Group, and the NASA). As keyphrases, KPEM provides many terms related to Alloy specification language (Alloy, Alloy Analyzer, snapshots), to UML (UML, UML Class Diagrams, OCL) and to the specification of dependable systems (Critical Systems, Invariants). The tag suggestions provided so far are extracted by the text present in the input document: no personalization is present at all. Suppose now that the user enables also the ORE module which exploits (in our example) a personal ontology6 in the field of software engineering (see left side of Figure 3).
ORE implements a navigation strategy, taking in input the key-phrases extracted by other annotators (KPEM in this case). For four out of the suggested key-phrases (i.e. Alloy, UML, OCL, and Invariants), ORE identifies a corresponding one-to-one match in the ontology (see Figure 4(b)). Starting from these nodes, ORE uses a spreading activation algorithm to find common ancestors representing more abstract subjects. Then both one-to-one ontology mappings and common ancestors are provided by PIRATES as potential tag recommendations, as summarized in Figure 4(a). The ontology navigation process highlighted by the spreading activation algorithm is depicted in Figure 4(b). In conclusion, the ORE module recommends five new tags which are not present in the text (i.e. Software Design Notation, Formal Specification Language, Design by Contract, Formal Specification Techniques, and Software Design) 7 .
These tags represent abstractions of the key-phrases extracted by the other annotators available in PIRATES.
We believe that the presented framework is a promising approach to automatic, personalized classification of Web contents. It is a first step in the direction of automatically organize document repositories into personal concept maps, moving from information to knowledge. The development of PIRATES has been planned in an incremental fashion, interleaved with experimental evaluation. Several modules have been already developed and integrated in a testbed environment: IEM with the sentiment analysis plug-in [16], KPEM with key-phrases extraction capabilities, and the Cognitive Filtering comprising an extended version of IFT capable to monitor Web 2.0 sources (specifically newsgoups, forums, and blogs). The integration of these modules is currently being evaluated. Prototyping and integration of ORE, SAT, and STE within PIRATES are ongoing processes, and evaluation experiments are planned. Moreover, we are working specifically on integrating the PIRATES modules in a Web-based version of the environment, which let us validate each module thoroughly. Finally, we have also planned to implement the conceptual map editor described in [8] in order to completely validate the framework.