Meaning Representation parsing aims to represent a sentence as a structured, Directed, Acyclic Graph (DAG), in an attempt to extract meaning from text. This paper extends an existing 2-stage pipeline AMR parser with state-of-the-art techniques in dependency parsing. First, Pointer-Generator Networks are used for out-of-vocabulary words in the concept identification stage, with an improved initialization via the use of word-and character-level embeddings. Second, the performance of the Relation Identification module is improved by jointly training the Heads Selection and the Arcs Labeling components. Last, we underline the difficulty of end-to-end training with recurrent modules in a static deep neural network construction approach and explore a dynamic construction implementation, which continuously adapts the computation graph, thus potentially enabling end-to-end training in the proposed pipeline solution.

Automatic captioning systems based on recurrent neural networks have been tremendously successful at providing realistic natural language captions for complex and varied image data. We explore methods for adapting existing models trained on large image caption data sets to a similar problem, that of summarising videos using natural language descriptions and frame selection. These architectures create internal high level representations of the input image that can be used to define probability distributions and distance metrics on these distributions. Specifically, we interpret each hidden unit inside a layer of the caption model as representing the un-normalised log probability of some unknown image feature of interest for the caption generation process. We can then apply well understood statistical divergence measures to express the difference between images and create an unsupervised segmentation of video frames, classifying consecutive images of low divergence as belonging to the same context, and those of high divergence as belonging to different contexts. To provide a final summary of the video, we provide a group of selected frames and a text description accompanying them, allowing a user to perform a quick exploration of large unlabeled video databases.

The large amount of unstructured medical documents written in natural language bears a massive quantity of knowledge, whose extraction becomes useful. An automatic relation identification strategy leads to the discovery of relations, (possible unknown) interactions, and associations between medical conditions, investigations and treatments. The current paper introduces a learning based approach for the automatic discovery of relations between medical concepts, entitled REMed. We propose an original list of features, grouped into four categories with the following distribution: lexical - 3, context - 6, grammatical -- 4 and syntactic - 4. We analyzed the influence of each category on the classification performance and determined that the performance of the REMed solution is comparable with similar solutions. We report the overall F-measure as 74.9% that outperforms the best solution reported in the similar systems with 1.2%. This performance was achieved mostly by the features from the lexical and context categories.

Imbalanced classification problems represent a current challenge in data mining research, due to the classifiers' inability to produce sufficiently good models in such situations. We have previously proposed a general methodology for improving the performance of classifiers under imbalance conditions: ECSB -- Evolutionary Cost-Sensitive Balancing. This paper provides an empirical analysis on a distributed approach for ECSB (dECSB). The influence of the number of splits on the quality of the output classification model is studied on several data sets and J4.8 as base classifier. The data sets have been partitioned according to the imbalance ratio and the instances per attributes ratio. We found that the appropriate number of splits is highly dependent on the problem at hand, however, an influence of the two imbalance-related factors is present. The effect of altering the genetic settings has also been investigated, in the attempt to identify several values which constantly yield good results. Again, we found the results to be highly dependent on the problem, with some data sets exhibiting low performance variations due to the genetic settings.

Current important challenges in data mining research are triggered by the need to address various particularities of real-world problems, such as imbalanced data and error cost distributions. This paper presents Distributed Evolutionary Cost-Sensitive Balancing, a distributed methodology for dealing with imbalanced data and -- if necessary -- cost distributions. The method employs a genetic algorithm to search for an optimal cost matrix and base classifier settings, which are then employed by a cost-sensitive classifier, wrapped around the base classifier. Individual fitness computation is the most intensive task in the algorithm, but it also presents a high parallelization potential. Two different parallelization alternatives have been explored: a computation-driven approach, and a data-driven approach. Both have been developed within the Apache Watchmaker framework and deployed on Hadoop-based infrastructures. Experimental evaluations performed up to this point have indicated that the computation-driven approach achieves a good classification performance, but does not reduce the running time significantly, the data-driven approach reduces the running time for slow algorithms, such as the kNN and the SVM, while still yielding important performance improvements.

Perhaps the most utilized and demanded task in data mining is classification. Most existing classification algorithms require all the data used for constructing the model for classification, or at least a good part of it, to be stored in the memory. This makes them limited by the availability of the memory. We present a parallel algorithm, based on the SPRINT decision tree, which eliminates the dependency on the memory available by storing the data to be processed in a database.

This paper focuses on solutions to two NP-Complete problems: k-SAT and the knapsack problem. We propose a new parallel genetic algorithm strategy on the CUDA architecture, and perform experiments to compare it with the sequential versions. We show how these problems can benefit from the GPU solutions, leading to significant improvements in speedup while keeping the quality of the solution. The best performance obtained in terms of speedup is 67 times. The solution presented in this paper suggests a general strategy for finding fast and robust solutions to complex problems.