Unsupervised Bilingual Dictionary Induction methods based on the initialization and the selflearning have achieved great success in similar language pairs, e.g., English-Spanish.But they still fail and have an accuracy of 0% in many distant language pairs, e.g., English-Japanese.In this work, we show that this failure results from the gap between the actual initialization performance and the minimum initialization performance for the self-learning to succeed.We propose Iterative Dimension Reduction to bridge this gap.Our experiments show that this simple method does not hamper the performance of similar language pairs and achieves an accuracy of 13.64∼55.53%between English and four distant languages, i.e., Chinese, Japanese, Vietnamese and Thai.

Deep encoders have been proven to be effective in improving neural machine translation (NMT) systems, but training an extremely deep encoder is time consuming.Moreover, why deep models help NMT is an open question.In this paper, we investigate the behavior of a well-tuned deep Transformer system.We find that stacking layers is helpful in improving the representation ability of N-MT models and adjacent layers perform similarly.This inspires us to develop a shallowto-deep training method that learns deep models by stacking shallow models.In this way, we successfully train a Transformer system with a 54-layer encoder.Experimental results on WMT'16 English-German and WMT'14 English-French translation tasks show that it is 1.4 × faster than training from scratch, and achieves a BLEU score of 30.33 and 43.29 on two tasks.The code is publicly available at https://github.com/libeineu/ SDT-Training.

The wound-field doubly salient generator (WFDSG) is receiving consideration for the field of industrial and aerospace/aviation, since it has the advantages of high robustness with simple structure and adjustable output voltage. In this paper, upon the analysis of generator structures, the features of three rectifiers that are suitable for WFDSG are analyzed in detail. The comparative research of WFDSG with hybrid half-bridge rectifier (HHBR), full-bridge rectifier (FBR), and half-bridge rectifier is presented by the finite-element method (FEM). Then, the load and no-load characteristics of WFDSG under the three rectifiers mentioned above are researched comparatively. The commutation overlap angles and total commutation output voltage losses of three rectifiers are deduced. The experiments of a 12/8-pole single armature-winding WFDSG (SAW-WFDSG) with an FBR and a 12/8-pole dual-armature-winding WFDSG (DAW-WFDSG) with HHBR prototypes have been performed. The FEM and experiment results shows that the SAW-WFDSG with FBR and the DAW-WFDSG with HHBR both have a good application prospect.

Wound-excited doubly salient machine (WEDSM) has a broad application prospect in the field of aircraft due to the advantages of simple and reliable structure, convenient manufacture and flexible adjustment of the air gap flux. However, the disadvantage of large torque ripple has also restricted its scope of application. The cogging torque of the machine is obtained from simulation and analyzed in details, which is proved to be one of the main factors that cause the torque ripple. Based on the back EMF waveforms of the machine, a novel three-phase nine-states control strategy (TPNS) for WEDSM is proposed and investigated in this paper. This novel control strategy can partially reduce the negative cogging torque. The above theoretical analysis is verified by the simulations and experiments with a 18kW 12/8-pole WEDSEM prototype.

Condition monitoring (CM) is of great significance to ensure the reliability of systems. With the wide use of power electronic equipment in multiple fields, the CM technology for power electronics has become more important. Feature parameter extraction is the prerequisite to realize CM of power electronic circuits. A unified method to extract parameters of the Buck converter with different properties of loads, including linear and nonlinear loads, is proposed in this paper. The established model containing target components makes the method more effective to eliminate the accuracy influence of non-ideal factors in practical applications compared to the traditional one based on hybrid system theory. De-noising pretreatment based on wavelet transform (WT) is applied to acquired data to filter non-ideal sampling errors in practice. The method is confirmed suitable for the Buck converter operating at any allowable operation condition. Both the simulation and experiment results validate the effectiveness of the presented strategy.

Dual-buck circuit has higher reliability as it overcomes the straight-leg problem existing in the traditional bridge circuit. However, more magnetic components are needed in this topology, especially in three-phase dual buck inverter. Magnetic integration is a good way to reduce the volume and weight of the magnetic components and improve the density of the system. The coupled-inductor dual-buck circuit utilizing PCDC (partly coupling with dotted-ends connected) method is essentially identical to three-inductor dual-buck circuit, which separates the anti-through and filter inductors and improves the inductor utilization compared to the conventional dual-buck circuit. This paper proposes an optimal design method for the coupled inductors. It not only decreases the magnetics number, but also reduces the total volume in comparison with discrete three inductors. The comparison of volume as well as equivalent filter inductance is carried out to verify the better performance of the coupled inductors. Finally the theory is validated by a 15kVA prototype, the experiment results are in consistent with analysis.

Abstract Shift-reduce parsing has been studied extensively for diverse grammars due to the simplicity and running efficiency. However, in the field of constituency parsing, shift-reduce parsers lag behind state-of-the-art parsers. In this paper we propose a semi-supervised approach for advancing shift-reduce constituency parsing. First, we apply the uptraining approach (Petrov, S. et al. 2010. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing (EMNLP), Cambridge, MA, USA, pp. 705–713) to improve part-of-speech taggers to provide better part-of-speech tags to subsequent shift-reduce parsers. Second, we enhance shift-reduce parsing models with novel features that are defined on lexical dependency information. Both stages depend on the use of large-scale unlabeled data. Experimental results show that the approach achieves overall improvements of 1.5 percent and 2.1 percent on English and Chinese data respectively. Moreover, the final parsing accuracies reach 90.9 percent and 82.2 percent respectively, which are comparable with the accuracy of state-of-the-art parsers.

The sensorless control technology is widely adopted for various applications. In this paper, a novel sensorless control based on coordinate transformation is proposed for doubly salient electromagnetic machine (DSEM). Based on the characteristics of the back electromotive force (EMF), the rotor position signals would be constructed. In order to acquire the commutation instant accurately, two different coordinates are defined and named as original coordinate and transformed coordinate. The detection of three-phase terminal voltages and coordinate transformation are realized in the digital signal process chip of DSP2812, and the estimated position signal is gained by the zero-crossing point of the terminal voltage in the transformed coordinate. The terminal voltage in the transformed coordinate is shifted from the terminal voltage in the original coordinate which is detected by DSP2812. Detailed analysis of DSEM's terminal voltage and the transform angle between two 3-phase coordinates at different speeds are involved in this paper. The sensorless control method is independent with the PWM technologies. A 18kW DSEM prototype is built and the above analysis is verified by the experimental results.

Five-phase fault-tolerant doubly salient electromagnetic generator (FTDSEG) is proposed for high reliability applications in this paper. On the basis of comparative analysis on structure characteristics between five-phase non-fault-tolerant doubly salient electro-magnetic generator (NFTDSEG) and five-phase FTDSEG, the variation of stator and rotor pole number from NFTDSEG to FTDSEG is revealed and the phase width of back electromotive force (EMF) of five-phase FTDSEG approaching 144°is deduced. Then the flux linkage and EMF of five-phase NFTDSEG and five-phase FTDSEG are investigated using finite-element method with two dimensions and four fault modes of five-phase FTDSEG are analyzed in detail. The laws of fault operating influence on external and power characteristic of motor and output voltages of rectifier are researched in four fault modes with different exciting currents. Meanwhile, the reason why the output voltage of rectifier of single phase short circuit fault is bigger than open circuit fault is given. Finally, the fault-tolerant characteristic of five-phase FTDSEG is verified by the experimental results.

The three-phase four-leg rectifier can be obtained by adding a bridge into the traditional three-phase three-leg rectifier. The cycle by cycle switch average model (CCA) based on three-phase four-leg rectifier is presented in the paper, and the model shows the rectifier can de decoupled to three single phase full bridge rectifiers. The SPWM (Sinusoidal Pulse Width Modulation) and SAPWM (Saddle Pulse Width Modulation) control strategy of rectifier are analyzed respectively. Compared with the SPWM control strategy, the SAPWM control strategy is benefit to lower output DC voltage and reduce the voltage level of switches. The three-phase four-leg rectifier has rigid adaptability to the power system whether it is three-phase balanced or unbalanced. Experimental results based on a 5-kW three-phase four-leg rectifier prototype verified that the rectifier achieves excellent performance.

Static and dynamic simulation methods are two commonly adopted techniques in constructing wind turbine simulator (WTS). To better understand the difference between static simulation method and actual wind turbine, their frequency-domain models which have never been discussed before are elaborated in this paper, and the reason why conventional static simulation method can not reproduce the dynamic behavior of a wind turbine and why dynamic simulation method should be adopted are carefully discussed. Then, the relationship between inertia of wind turbine and time constant of LFP are discussed based on the simplified model of dynamic WTS proposed in the paper, which reveals the ways to simulate different wind turbines with one hardware specification. An experimental platform which can accurately simulate both the static and dynamic characteristic of real wind turbines is implemented. The theoretical analysis and proposed strategy is verified by simulation and experimental results performed by a 7.5KW laboratory prototype.

Opinion polling has been traditionally done via customer satisfaction studies in which questions are carefully designed to gather customer opinions about target products or services. This paper studies aspect-based opinion polling from unlabeled free-form textual customer reviews without requiring customers to answer any questions. First, a multi-aspect bootstrapping method is proposed to learn aspect-related terms of each aspect that are used for aspect identification. Second, an aspect-based segmentation model is proposed to segment a multi-aspect sentence into multiple single-aspect units as basic units for opinion polling. Finally, an aspect-based opinion polling algorithm is presented in detail. Experiments on real Chinese restaurant reviews demonstrated that our approach can achieve 75.5 percent accuracy in aspect-based opinion polling tasks. The proposed opinion polling method does not require labeled training data. It is thus easy to implement and can be applicable to other languages (e.g., English) or other domains such as product or movie reviews.

To solve the knowledge bottleneck problem, active learning has been widely used for its ability to automatically select the most informative unlabeled examples for human annotation. One of the key enabling techniques of active learning is uncertainty sampling, which uses one classifier to identify unlabeled examples with the least confidence. Uncertainty sampling often presents problems when outliers are selected. To solve the outlier problem, this paper presents two techniques,sampling by uncertainty and density (SUD)anddensity-based re-ranking. Both techniques prefer not only the most informative example in terms of uncertainty criterion, but also the most representative example in terms of density criterion. Experimental results of active learning for word sense disambiguation and text classification tasks using six real-world evaluation data sets demonstrate the effectiveness of the proposed methods.

This paper presents an unsupervised approach to aspect-based opinion polling from raw textual reviews without explicit ratings. The key contribution of this paper is three-fold. First, a multi-aspect bootstrapping algorithm is proposed to learn from unlabeled data aspect-related terms of each aspect to be used for aspect identification. Second, an unsupervised segmentation model is proposed to address the challenge of identifying multiple single-aspect units in a multi-aspect sentence. Finally, an aspect-based opinion polling algorithm is presented. Experiments on real Chinese restaurant reviews show that our opinion polling method can achieve 75.5% precision performance.

The performance of machine learning methods heavily depends on the volume of used training data. For the purpose of dataset enlargement, it is of interest to study the problem of unifying multiple labeled datasets with different annotation standards. In this paper, we focus on the case of unifying datasets for sequence labeling problems with natural language part-of-speech (POS) tagging as an examplar application. To this end, we propose a probabilistic approach to transforming the annotations of one dataset to the standard specified by another dataset. The key component of the approach, named as label correspondence learning, serves as a bridge of annotations from the datasets. Two methods designed from distinct perspectives are proposed to attack this sub-problem. Experiments on two large-scale part-of-speech datasets demonstrate the efficacy of the transformation and label correspondence learning methods.