José A. R. Fonollosa

N-gram-based Machine Translation

This article describes in detail an n-gram approach to statistical machine translation. This approach consists of a log-linear combination of a translation model based on n-grams of bilingual units, which are referred to as tuples, along with four specific feature functions. Translation performance, which happens to be in the state of the art, is demonstrated with Spanish-to-English and English-to-Spanish translations of the European Parliament Plenary Sessions (EPPS).

Ordered Eigenvalues of a General Class of Hermitian Random Matrices With Application to the Performance Analysis of MIMO Systems

In this paper, we present a general formulation that unifies the probabilistic characterization of Hermitian random matrices with a specific structure. Based on a general expression for the joint pdf of the ordered eigenvalues, we obtain i) the joint cdf; ii) the marginal cdfs; and iii) the marginal pdfs of the ordered eigenvalues, where ii) and iii) follow as simple particularizations of i). Our formulation is shown to include the distribution of some common MIMO channel models such as the uncorrelated, semicorrelated, and double-correlated Rayleigh MIMO fading channel and the uncorrelated Rician MIMO fading channel, although it is not restricted only to these. Hence, the proposed formulation and derived results provide a solid framework for the simultaneous analytical performance analysis of MIMO systems under different channel models. As an illustrative application, we obtain the exact outage probability of a spatial multiplexing MIMO system transmitting through the strongest channel eigenmodes.

System identification using a linear combination of cumulant slices

A linear approach to identifying the parameters of a moving-average (MA) model from the statistics of the output is presented. First, it is shown that, under some constraints, the impulse response of the system can be expressed as a linear combination of cumulant slices. Then, this result is used to obtain a well-conditioned linear method for estimating the MA parameters of a nonGaussian process. The linear combination of slices used to compute the MA parameters can be constructed from different sets of cumulants of different orders, provided a general framework in which all the statistics can be combined. It is not necessary to use second-order statistics (autocorrelation slice), and therefore the proposed algorithm still provides consistent estimates in the presence of colored Gaussian noise. Another advantage of the method is that while most linear methods give totally erroneous estimates if the order is overestimated, the proposed approach does not require a previous estimation of the filter order. The simulation results confirm the good numerical conditioning of the algorithm and its improvement in performance in comparison to existing methods. >

Blind channel estimation and data detection using hidden Markov models

We propose applying the hidden Markov models (HMM) theory to the problem of blind channel estimation and data detection. The Baum-Welch (BW) algorithm, which is able to estimate all the parameters of the model, is enriched by introducing some linear constraints emerging from a linear FIR hypothesis on the channel. Additionally, a version of the algorithm that is suitable for time-varying channels is also presented. Performance is analyzed in a GSM environment using standard test channels and is found to be close to that obtained with a nonblind receiver.

Character-based neural machine translation

Neural Machine Translation (MT) has reached state-of-the-art results. However, one of the main challenges that neural MT still faces is dealing with very large vocabularies and morphologically rich languages. In this paper, we propose a neural MT system using character-based embeddings in combination with convolutional and highway layers to replace the standard lookup-based word representations. The resulting unlimited-vocabulary and affix-aware source word embeddings are tested in a state-of-the-art neural MT based on an attention-based bidirectional recurrent neural network. The proposed MT scheme provides improved results even when the source language is not morphologically rich. Improvements up to 3 BLEU points are obtained in the German-English WMT task.

Analytical BER performance in spatial multiplexing MIMO systems

In this paper, we investigate the average bit error rate (BER) performance of spatial multiplexing MIMO systems with CSI at both sides of the link. Such systems result from the use of a multiple beamforming strategy that exploits the channel eigenmodes or, in other words, from the use of joint transmit-receive linear signal processing techniques in multiantenna systems. Since an analytical closed-form expression for the average BER of the multiple beamforming scheme is difficult to be found, we derive an approximation of the BER performance for high signal-to-noise ratio (SNR) assuming a Rayleigh flat fading channel. In the high SNR regime, the BER versus SNR curve can be characterized in terms of two key parameters: the array gain and the diversity gain. Thus, the proposed analysis offers a simple approach to determine the average performance of the multiple beamforming scheme.

Blind multiuser identification and detection in CDMA systems

Multiuser detection in code division multiple access systems usually requires either knowledge of the transmitted signature sequences and channel state information or use of a known training sequence for adaptation. We develop a scheme that can be employed for the joint adaptive blind multiuser identification and detection in asynchronous CDMA systems. This scheme relies on a multiuser Viterbi algorithm that incorporates an adaptive estimation of the overall channel impulse responses, given by the convolution of the signature sequences of the users and corresponding physical channels impulse responses. Once the overall channel responses are estimated, the blind multiuser detection algorithm performs like the maximum-likelihood sequence estimator. Results are provided to illustrate the convergence of the blind multiuser approach, near-far resistance and sensitivity to the algorithm initialization.

Application of hidden Markov models to blind channel characterization and data detection

The paper applies the theory of hidden Markov models (HMM) in digital communications to obtain a complete characterization of the channel, convolutional code, and transmitted constellation in a blind environment, i.e., without the help of a training sequence. The HMM formulation leads to a joint maximum likelihood estimation of both the channel and the transmitted sequence. The Baum-Welch (BW) identification algorithm is able to estimate all the parameters of the model, including the constellation and the probability of each symbol. Minor modifications to the algorithm allow one to consider restrictions or known parameters to improve the estimation of the rest of parameters with methods of lower complexity. An LMS-type adaptive version of the BW algorithm is also developed and tested. The simulation results show the fast convergence of the proposed optimal approaches. They also demonstrate that the performance of the decoder is comparable to that obtained in a known environment.<<ETX>>

Pitch determination of noisy speech using higher order statistics

The use of third-order statistics to determine the pitch of a speech signal and how they can eliminate the effect of a wide range of noises, including those generated by periodic sources, are shown. The proposed algorithm is based on the property that higher-order statistics can extract useful information about the statistics of voiced frames, and they can separate speech from noise. Third-order statistics are quite insensitive to most noises (Gaussian, sinusoidal, car noise, etc.) because these noises have a symmetric probability density function, and therefore their third-order cumulants are zero. The algorithm has been tested in noise-corrupted speech, at different levels of signal to noise ratio, and with different kinds of noise. The results show that this new algorithm gives in all the cases a much better estimation of the pitch than the conventional autocorrelation method.<<ETX>>

N-Gram-Based Statistical Machine Translation versus Syntax Augmented Machine Translation: Comparison and System Combination

In this paper we compare and contrast two approaches to Machine Translation (MT): the CMU-UKA Syntax Augmented Machine Translation system (SAMT) and UPC-TALP N-gram-based Statistical Machine Translation (SMT). SAMT is a hierarchical syntax-driven translation system underlain by a phrase-based model and a target part parse tree. In N-gram-based SMT, the translation process is based on bilingual units related to word-to-word alignment and statistical modeling of the bilingual context following a maximum-entropy framework. We provide a step-by-step comparison of the systems and report results in terms of automatic evaluation metrics and required computational resources for a smaller Arabic-to-English translation task (1.5M tokens in the training corpus). Human error analysis clarifies advantages and disadvantages of the systems under consideration. Finally, we combine the output of both systems to yield significant improvements in translation quality.