In machine-learning applications, data selection is of crucial importance if good runtime performance is to be achieved. In a scenario where the test set is accessible when the model is being built, training instances can be selected so they are the most relevant for the test set. Feature Decay Algorithms (FDA) are a technique for data selection that has demonstrated excellent performance in a number of tasks. This method maximizes the diversity of the n-grams in the training set by devaluing those ones that have already been included. We focus on this method to undertake deeper research on how to select better training data instances. We give an overview of FDA and propose improvements in terms of speed and quality. Using German-to-English parallel data, first we create a novel approach that decreases the execution time of FDA when multiple computation units are available. In addition, we obtain improvements on translation quality by extending FDA using information from the parallel corpus that is generally ignored.

The Bidirectional Encoder Representations from Transformers (BERT) model produces state-of-the-art results in many question answering (QA) datasets, including the Stanford Question Answering Dataset (SQuAD). This paper presents a query expansion (QE) method that identifies good terms from input questions, extracts synonyms for the good terms using a widely-used language resource, WordNet, and selects the most relevant synonyms from the list of extracted synonyms. The paper also introduces a novel QE method that produces many alternative sequences for a given input question using same-language machine translation (MT). Furthermore, we use a coreference resolution (CR) technique to identify anaphors or cataphors in paragraphs and substitute them with the original referents. We found that the QA system with this simple CR technique significantly outperforms the BERT baseline in a QA task. We also found that our best-performing QA system is the one that applies these three preprocessing methods (two QE and CR methods) together to BERT, which produces an excellent F1 score (89.8 F1 points) in a QA task. Further, we present a comparative analysis on the performances of the BERT QA models taking a variety of criteria into account, and demonstrate our findings in the answer span prediction task.

Neural handwriting recognition (NHR) is the recognition of handwritten text with deep learning models, such as multi-dimensional long short-term memory (MDLSTM) recurrent neural networks. Models with MDLSTM layers have achieved state-of-the art results on handwritten text recognition tasks. While multi-directional MDLSTM-layers have an unbeaten ability to capture the complete context in all directions, this strength limits the possibilities for parallelization, and therefore comes at a high computational cost. In this work we develop methods to create efficient MDLSTMbased models for NHR, particularly a method aimed at eliminating computation waste that results from padding. This proposed method, called example-packing, replaces wasteful stacking of padded examples with efficient tiling in a 2-dimensional grid. For word-based NHR this yields a speed improvement of factor 6.6 over an already efficient baseline of minimal padding for each batch separately. For line-based NHR the savings are more modest, but still significant. In addition to example-packing, we propose: 1) a technique to optimize parallelization for dynamic graph definition frameworks including PyTorch, using convolutions with grouping, 2) a method for parallelization across GPUs for variable-length example batches. All our techniques are thoroughly tested on our own PyTorch re-implementation of MDLSTM-based NHR models. A thorough evaluation on the IAM dataset shows that our models are performing similar to earlier implementations of state-of-the art models. Our efficient NHR model and some of the reusable techniques discussed with it offer ways to realize relatively efficient models for the omnipresent scenario of variable-length inputs in deep learning.

Data Selection is a popular step in Machine Translation pipelines. Feature Decay Algorithms (FDA) is a technique for data selection that has shown a good performance in several tasks. FDA aims to maximize the coverage of n-grams in the test set. However, intuitively, more ambiguous n-grams require more training examples in order to adequately estimate their translation probabilities. This ambiguity can be measured by alignment entropy. In this paper we propose two methods for calculating the alignment entropies for n-grams of any size, which can be used for improving the performance of FDA. We evaluate the substitution of the n-gramspecific entropy values computed by these methods to the parameters of both the exponential and linear decay factor of FDA. The experiments conducted on German-to-English and Czechto-English translation demonstrate that the use of alignment entropies can lead to an increase in the quality of the results of FDA.

Neural Machine Translation (NMT) models achieve their best performance when large sets of parallel data are used for training. Consequently, techniques for augmenting the training set have become popular recently. One of these methods is back-translation (Sennrich et al., 2016a), which consists on generating synthetic sentences by translating a set of monolingual, target-language sentences using a Machine Translation (MT) model.Generally, NMT models are used for backtranslation. In this work, we analyze the performance of models when the training data is extended with synthetic data using different MT approaches. In particular we investigate back-translated data generated not only by NMT but also by Statistical Machine Translation (SMT) models and combinations of both. The results reveal that the models achieve the best performances when the training set is augmented with back-translated data created by merging different MT approaches.