Aspect-level sentiment analysis aims to predict the sentiment polarity toward a specific aspect in a sentence. Most current approaches are based on deep learning and the attention mechanism. However, these models cannot simultaneously include the context semantic information carried by local words and the global syntactic information possibly carried by remote words. In this paper, we propose a local semantic and global syntactic integration scheme, which employs a local focus mechanism over local context words and exploits improved graph convolutional networks over dependency tree to encode global syntactic information. Moreover, multi-head attention is used to capture both the semantic information and also the interactive information between semantics and syntactic features. Experimental results on five datasets show the effectiveness of our model over a series of latest models.

In this paper, we consider an integrated service system, providing multirate guaranteed services and homogeneous best-effort services. The total system capacity is shared by both service types, while the guaranteed service customers are treated as high priority and are allocated fixed data rate bandwidth units. The best-effort service customers are supported by the remaining capacity leftover by the guaranteed services, in a processor-sharing manner. Admission control on the best-effort service customers is adopted to provide a certain level of quality of service (not guaranteed) to avoid the effect of repeated attempts. The best-effort customer blocking probability is an important metric for network dimensioning. In this integrated service system, we observe that the blocking probability of the best-effort customers is not insensitive to the shapes of their flow-size distributions and the guaranteed customers' holding-time distributions. We obtain here by light computation, for the best-effort customers, blocking probability bounds, and approximations, which possess the insensitivity property. Considering that the distributions for customers' holding times and flow sizes may be unknown, the proposed insensitive bounds and approximations will facilitate system design and network dimensioning with predictable and acceptable performance.

Mode division multiplexing (MDM) over wavelength division multiplexing (WDM) network is studied, particularly focusing on low differential mode group delay (DMGD) two-mode fiber (TMF) link in a wide spectral range. Here, the DMGD is defined as the group delay between the two guided modes, LP01and LP11modes. For the starter, We will revisit an invention of the original TMF in late 70's, followed by the DMGD-compensated TMF link and crosstalk between two mode channels due to the mode coupling effect. Next, in order to explore the feasibility of "real-time" digital signal processing (DSP) for multiple-input multiple-output (MIMO) transmission system for equalizing the mode channel crosstalk, the computational complexity of MIMO DSP will be evaluated. As a consequence, a criterion of the DMGD value to enable the realtime MIMO DSP is derived. Finally, against a backdrop of 560% predicted traffic surge of metro area network (MAN) by 2017, a deployment scenario of MDM over WDM MAN is presented as the solution, and the MDM node architecture will be discussed.

For integrated service systems supporting real-time (RT) and non-real-time (NRT) traffic, evaluating the performance of the NRT traffic is often challenging. An approximation based on a fluid model - the so-called Fluid Approximation (FA) - can be used when RT traffic has much larger successful call arrival rate than the NRT traffic, while otherwise it may introduce errors. In this paper, we propose a Generalized Fluid Approximation (GFA), extending the use and improving the accuracy of FA. It can also be used to trade off the computational complexity and the accuracy. Benefits achieved by GFA are demonstrated by extensive numerical examples.

This paper considers a two-tier femto-macro overlay network architecture, where groups of femtocells are overlaid by macrocells. The capacity of the femtocells is limited, thus redirection of blocked femto calls to the overlaid macrocell is used as a QoS assurance strategy. Similar overflow approaches have been used in telephony; however, extension of this approach to capacity-constrained femto-macro overlay networks poses significant challenges due to the wide variety of bandwidth requirement of Internet services, the large number of femtocells overlaid by a macrocell, and the need for fast computations imposed by practical requirements of dynamic real-time resource allocation. We propose a new computationally efficient numerical method for the loss probability approximation in such two-tier femto-macro overlay networks. The accuracy of the method is verified by exact Markov-chain solutions for small systems and by simulations for larger systems.