Obscenity (the use of rude words or offensive expressions) has spread from informal verbal conversations to digital media, becoming increasingly common on user-generated comments found in Web forums, newspaper user boards, social networks, blogs, and media-sharing sites. The basic obscenity-blocking mechanism is based on verbatim comparisons against a blacklist of banned vocabulary; however, creative users circumvent these filters by obfuscating obscenity with symbol substitutions or bogus segmentations that still visually preserve the original semantics, such as writing shit as $h¡;t or s.h.i.t or even worse mixing them as $.h….¡.t . The number of potential obfuscated variants is combinatorial, yielding the verbatim filter impractical. Here we describe a method intended to obstruct this anomaly inspired by sequence alignment algorithms used in genomics, coupled with a tailor-made edit penalty function. The method only requires to set up the vocabulary of plain obscenities; no further training is needed. Its complexity on screening a single obscenity is linear, both in runtime and memory, on the length of the user-generated text. We validated the method on three different experiments. The first one involves a new dataset that is also introduced in this article; it consists of a set of manually annotated real-life comments in Spanish, gathered from the news user boards of an online newspaper, containing this type of obfuscation. The second one is a publicly available dataset of comments in Portuguese from a sports Web site. In these experiments, at the obscenity level, we observed recall rates greater than 90%, whereas precision rates varied between 75% and 95%, depending on their sequence length (shorter lengths yielded a higher number of false alarms). On the other hand, at the comment level, we report recall of 86%, precision of 91%, and specificity of 98%. The last experiment revealed that the method is more effective in matching this type of obfuscation compared to the classical Levenshtein edit distance. We conclude discussing the prospects of the method to help enforcing moderation rules of obscenity expressions or as a preprocessing mechanism for sequence cleaning and/or feature extraction in more sophisticated text categorization techniques.

This paper considers large-scale OneMax and RoyalRoad problems with up to 107 binary variables within a compact Estimation of Distribution Algorithms (EDA) framework. Building upon the compact Genetic Algorithm (cGA), the continuous domain Population-Based Incremental Learning algorithm (PBILc) and the arithmetic-coding EDA, we define a novel method that is able to compactly solve regular and noisy versions of these problems with minimal memory requirements, regardless of problem or population size. This feature allows the algorithm to be run in a conventional desktop machine. Issues regarding probability model sampling, arbitrary precision of the arithmetic-coding decompressing scheme, incremental fitness function evaluation and updating rules for compact learning, are presented and discussed.

Given an $n$-vertex weighted tree with structural diameter $S$ and a subset of $m$ vertices, we present a technique to compute a corresponding $m \times m$ Gram matrix of the pseudoinverse of the graph Laplacian in $O(n+ m^2 + m S)$ time. We discuss the application of this technique to fast label prediction on a generic graph. We approximate the graph with a spanning tree and then we predict with the kernel perceptron. We address the approximation of the graph with either a minimum spanning tree or a shortest path tree. The fast computation of the pseudoinverse enables us to address prediction problems on large graphs. To this end we present experiments on two web-spam classification tasks, one of which includes a graph with 400,000 nodes and more than 10,000,000 edges. The results indicate that the accuracy of our technique is competitive with previous methods using the full graph information.