José Luís Devezas

Using the overlapping community structure of a network of tags to improve text clustering

Breadcrumbs is a folksonomy of news clips, where users can aggregate fragments of text taken from online news. Besides the textual content, each news clip contains a set of metadata fields associated with it. User-defined tags are one of the most important of those information fields. Based on a small data set of news clips, we build a network of co-occurrence of tags in news clips, and use it to improve text clustering. We do this by defining a weighted cosine similarity proximity measure that takes into account both the clip vectors and the tag vectors. The tag weight is computed using the related tags that are present in the discovered community. We then use the resulting vectors together with the new distance metric, which allows us to identify socially biased document clusters. Our study indicates that using the structural features of the network of tags leads to a positive impact in the clustering process.

Studying a personality coreference network in a news stories photo collection

We build and analyze a coreference network based on entities from photo descriptions, where nodes represent personalities and edges connect people mentioned in the same photo description. We identify and characterize the communities in this network and propose taking advantage of the context provided by community detection methodologies to improve text illustration and general search.

Studying blog features over link popularity

The study of the blogosphere can provide sociologically relevant data. We analyze the links between blogs in the portuguese blogosphere, in order to understand how they group and interact, to identify clusters and to characterize them. Our data set contains post data for more than 70,000 blogs, with over 400,000 links. The linkage data is represented as a blog graph and partitioned into several slices, according to their in-degree. We then study the evolution of blog features, and observe a consistent pattern of decrease in posting frequency, number of out-links, and post length, as we move from the highly-cited blogs to the less cited ones.

Temporal Visualization of a Multidimensional Network of News Clips

The exploration of large networks carries inherent challenges in the visualization of a great amount of data. We built an interactive visualization system for the purpose of exploring a large multidimensional network of news clips over time. These are clips gathered by users from web news sources and references to people or places are extracted from. In this paper, we present the system’s capabilities and user interface and discuss its advantages in terms of the browsing and extraction of knowledge from the data. These capabilities include a textual search and associated event detection, and temporal navigation allowing the user to seek a certain date and timespan.

The community structure of a multidimensional network of news clips

We analysed the community structure of a network of news clips where relationships were established by the co-reference of entities in pairs of clips. Community detection was applied to a unidimensional version of the news clips network, as well as to a multidimensional version where dimensions were defined based on three different classes of entities: places, people, and dates. The goal was to study the impact on the quality of the identified community structure when using multiple dimensions to model the network. We did a two-fold evaluation, first based on the modularity metric and then based on human input regarding community semantics. We verified that the assessments of the evaluators differed from the results provided by the modularity metric, pointing towards the relevance of the utility and network integration phases in the identification of semantically cohesive groups of news clips.

Index-Based Semantic Tagging for Efficient Query Interpretation

Modern search engines are evolving beyond ad hoc document retrieval. Nowadays, the information needs of the users can be directly satisfied through entity-oriented search, by ranking the entities or attributes that better relate to the query, as opposed to the documents that contain the best matching terms. One of the challenges in entity-oriented search is efficient query interpretation. In particular, the task of semantic tagging, for the identification of entity types in query parts, is central to understanding user intent. We compare two approaches for semantic tagging, within a single domain, one based on a Sesame triple store and another one based on a Lucene index. This provides a segmentation and annotation of the query based on the most probable entity types, leading to query classification and its subsequent interpretation. We evaluate the run time performance for the two strategies and find that there is a statistically significant speedup, of at least four times, for the index-based strategy over the triple store strategy.