In dialogue state tracking (DST), the exploitation of dialogue history is a crucial research direction, and the existing DST models can be divided into two categories: full-history models and partial-history models. Since the “select first, use later” mechanism explicitly filters the distracting information being passed to the downstream state prediction, the partial-history models have recently achieved a performance advantage over the full-history models. However, besides the redundant information, some critical dialogue context information was inevitably filtered out by the partial-history models simultaneously. To reconcile the contextual consideration with avoiding the introduction of redundant information, we propose DICE-DST, a model-agnostic module widely applicable to the partial-history DST models, which aims to strengthen the ability of context exploitation for the encoder of each DST model. Specifically, we first construct a teacher encoder and devise two contextual reasoning tasks to train it to acquire extensive dialogue contextual knowledge. Then we transfer the contextual knowledge from the teacher encoder to the student encoder via a novel turn-level attention-alignment distillation. Experimental results show that our approach extensively improves the performance of partial-history DST models and thereby achieves new state-of-the-art performance on multiple mainstream datasets while keeping high efficiency.

The EAE task extracts a structured event record from an event text.Most existing approaches train the EAE model on each dataset independently and ignore the overlap knowledge across datasets.However, insufficient event records in a single dataset often prevent the existing model from achieving better performance.In this paper, we clearly define the overlap knowledge across datasets and split the knowledge of the EAE task into overlap knowledge across datasets and specific knowledge of the target dataset.We propose APE model to learn the two parts of knowledge in two serial learning phases without causing catastrophic forgetting.In addition, we formulate both learning phases as conditional generation tasks and design Stressing Entity Type Prompt to close the gap between the two phases.The experiments show APE achieves new state-of-the-art with a large margin in the EAE task.When only ten records are available in the target dataset, our model dramatically outperforms the baseline model with average 27.27%F1 gain.

Relation Classification (RC) is a basic and essential task of Natural Language Processing. Existing RC methods can be classified into two categories: sequence-based methods and dependency-based methods. Sequence-based methods identify the target relation based on the overall semantics of the whole sentence, which will inevitably introduce noisy features. Dependency-based methods extract indicative word-level features from the Shortest Dependency Path (SDP) between given entities and attempt to establish a statistical association between the words and the target relations. This pattern relatively eliminates the influence of noisy features and achieves a robust performance on long sentences. Nevertheless, we observe that majority of relation classification processes involve complex semantic reasoning which is hard to be achieved based on the word-level statistical association. To solve this problem, we categorize all relations into atomic relations and composed-relations. The atomic relations are the basic relations that can be identified based on the word-level features, while the composed-relation requires to be deducted from multiple atomic relations. Correspondingly, we propose theAtomic RelationEncoding andReasoningModel (ATERM). In the atomic relation encoding stage, ATERM groups the word-level features and encodes multiple atomic relations in parallel. In the atomic relation reasoning stage, ATERM establishes the atomic relation chain where relation-level features are extracted to identify composed-relations. Experiments show that our method achieves state-of-the-art results on the three most popular relation classification datasets – TACRED, TACRED-Revisit, and SemEval 2010 task 8 with significant improvements.

In dialogue state tracking, dialogue history is a crucial material, and its utilization varies between different models.However, no matter how the dialogue history is used, each existing model uses its own consistent dialogue history during the entire state tracking process, regardless of which slot is updated.Apparently, it requires different dialogue history to update different slots in different turns.Therefore, using consistent dialogue contents may lead to insufficient or redundant information for different slots, which affects the overall performance.To address this problem, we devise DiCoS-DST to dynamically select the relevant dialogue contents corresponding to each slot for state updating.Specifically, it first retrieves turn-level utterances of dialogue history and evaluates their relevance to the slot from a combination of three perspectives: (1) its explicit connection to the slot name; (2) its relevance to the current turn dialogue; (3) Implicit Mention Oriented Reasoning.Then these perspectives are combined to yield a decision, and only the selected dialogue contents are fed into State Generator, which explicitly minimizes the distracting information passed to the downstream state prediction.Experimental results show that our approach achieves new state-of-the-art performance on MultiWOZ 2.1 and MultiWOZ 2.2, and achieves superior performance on multiple mainstream benchmark datasets (including Sim-M, Sim-R, and DSTC2). 1

Jinyu Guo, Kai Shuang, Jijie Li, Zihan Wang. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Inspired by the success of BERT, several multimodal representation learning approaches have been proposed that jointly represent image and text. These approaches achieve superior performance by capturing high-level semantic information from large-scale multimodal pretraining. In particular, LXMERT and UNITER adopt visual region feature regression and label classification as pretext tasks. However, they tend to suffer from the problems of noisy labels and sparse semantic annotations, based on the visual features having been pretrained on a crowdsourced dataset with limited and inconsistent semantic labeling. To overcome these issues, we propose unbiased Dense Contrastive Visual-Linguistic Pretraining (DCVLP), which replaces the region regression and classification with cross-modality region contrastive learning that requires no annotations. Two data augmentation strategies (Mask Perturbation and Intra-Inter-Adversarial Perturbation) are developed to improve the quality of negative samples used in contrastive learning. Overall, DCVLP allows cross-modality dense region contrastive learning in a self-supervised setting independent of any object annotations. We compare our method against prior visual-linguistic pretraining frameworks to validate the superiority of dense contrastive learning on multimodal representation learning.

Multi-modality fusion technologies have greatly improved the performance of neural network-based Video Description/Caption, Visual Question Answering (VQA) and Audio Visual Scene-aware Dialog (AVSD) over the recent years. Most previous approaches only explore the last layers of multiple layer feature fusion while omitting the importance of intermediate layers. To solve the issue for the intermediate layers, we propose an efficient Quaternion Block Network (QBN) to learn interaction not only for the last layer but also for all intermediate layers simultaneously. In our proposed QBN, we use the holistic text features to guide the update of visual features. In the meantime, Hamilton quaternion products can efficiently perform information flow from higher layers to lower layers for both visual and text modalities. The evaluation results show our QBN improve the performance on VQA 2.0, furthermore surpasses the approach using large scale BERT or visual BERT pre-trained models. Extensive ablation study has been carried out to examine the influence of each proposed module in this study.

Neuron-level structured pruning is a very effective technique to reduce the computation of neural networks without compromising prediction accuracy. In previous works, structured pruning is usually achieved by imposing L1 regularization on the scaling factors of neurons, and pruning the neurons whose scaling factors are below a certain threshold. The reasoning is that neurons with smaller scaling factors have weaker influence on network output. A scaling factor close to 0 actually suppresses a neuron. However, L1 regularization lacks discrimination between neurons because it pushes all scaling factors towards 0. A more reasonable pruning method is to only suppress unimportant neurons (with 0 scaling factors) and simultaneously keep important neurons intact (with larger scaling factor). To achieve this goal, we propose a new regularizer on scaling factors, namely polarization regularizer. Theoretically, we prove that polarization regularizer pushes some scaling factors to 0 and others to a value $a > 0$. Experimentally, we show that structured pruning using polarization regularizer achieves much better results than using L1 regularizer. Experiments on CIFAR and ImageNet datasets show that polarization pruning achieves the state-of-the-art result to date.

The object tracking task requires tracking the arbitrary target in consecutive video frames. Recently, several attempts have been made to develop the template-free models to attain generality. However, current template-free paradigm only estimates the displacement to approximate the motion of the object. The displacement is insufficient to represent complex bounding box transformation, including scaling and rotation. We argue that the coarse-grained representation of object motion limits the performance of current template-free approaches. In this paper, we explore the finer-grained motion estimation to improve the accuracy of the template-free model. In respect of the image space, our method estimates the transformation for each pixel in the image. Concern on the motion representation, we represent the motion by the transformation parameterized by displacement, scaling, and rotation. By applying the differential vector operators on the optical flow, our approach estimates both displacement, scaling, and rotation for each pixel in a unified theory. To the best of our knowledge, we are the first work to model the displacement, scaling and rotation in a unified theory with the optical flow. To further improve the localization accuracy, we develop the appearance branch to introduce the appearance information into our model. Furthermore, to suppress optical flow estimation failure samples during training, we propose a novel loss function Limited L1. The experiment shows our model FGTrack achieves state-of-the-art performance on both NFS and VOT2017 datasets.

Language model (LM) plays an important role in natural language processing (NLP) systems, such as machine translation, speech recognition, learning token embeddings, natural language generation, and text classification. Recently, the multilayer long short-term memory (LSTM) models have been demonstrated to achieve promising performance on word-level language modeling. For each LSTM layer, larger hidden size usually means more diverse semantic features, which enables the LM to perform better. However, we have observed that when a certain LSTM layer reaches a sufficiently large scale, the promotion of overall effect will slow down, as its hidden size increases. In this article, we analyze that an important factor leading to this phenomenon is the high correlation between the newly extended hidden states and the original hidden states, which hinders diverse feature expression of the LSTM. As a result, when the scale is large enough, simply lengthening the LSTM hidden states will cost tremendous extra parameters but has little effect. We propose a simple yet effective improvement on each LSTM layer consisting of a large-scale Major LSTM and a small-scale Minor LSTM to break the high correlation between the two parts of hidden states, which we call Major-Minor LSTMs (MMLSTMs). In experiments, we demonstrate the LM with MMLSTMs surpasses the existing state-of-the-art model on Penn Treebank (PTB) and WikiText-2 (WT2) data sets and outperforms the baseline by 3.3 points in perplexity on WikiText-103 data set without increasing model parameter counts.

This study introduces sparse code multiple access (SCMA) and wavelet packet division multiplexing (WPDM) to logging cable telemetry systems (LCTSs). The authors study the feasibility of SCMA for LCTSs and provide a WPDM‐based system model, as well as the associated optimal waveform design under sampling clock offset (SCO) via a genetic algorithm. The performance advantages of the designed scheme over the conventional WPDM and orthogonal frequency division multiplexing (OFDM) are demonstrated by both analytical and simulation methods. Numerical results show the performance of the SCMA‐based mapping method over quadrature amplitude modulation in terms of improved spectral efficiency. Furthermore, the new scheme outperforms conventional WPDM and OFDM in bit error rate performance and transmission rate under SCO. Such advantages with the flexible design of basis functions make SCMA and WPDM a significant candidate for LCTSs.

Text classification is a foundational task in many natural language processing applications. All traditional text classifiers take words as the basic units and conduct the pre-training process (like word2vec) to directly generate word vectors at the first step. However, none of them have considered the information contained in word structure which is proved to be helpful for text classification. In this paper, we propose a word-building method based on neural network model that can decompose a Chinese word to a sequence of radicals and learn structure information from these radical level features which is a key difference from the existing models. Then, the convolutional neural network is applied to extract structure information of words from radical sequence to generate a word vector, and the long short-term memory is applied to generate the sentence vector for the prediction purpose. The experimental results show that our model outperforms other existing models on Chinese dataset. Our model is also applicable to English as well where an English word can be decomposed down to character level, which demonstrates the excellent generalisation ability of our model. The experimental results have proved that our model also outperforms others on English dataset.

In network virtualization, one of the key problems is to embed a sequence of virtual networks with both node and link constraints onto the physical network, which is known to be NP-hard. When a virtual network arrives, the recent studies focus on designing a solution to minimize the energy cost while maximizing the revenue of the physical network at that time. However, after some time, due to the significant dynamics of the resources of the physical network, the previous solution may become less energy efficient. In this paper, we take a further step and study how to re-optimize the energy cost by leveraging the migration technique. In particular, we first model the pros (e.g., energy saving) and cons (e.g., interruption time and bandwidth waste) of migration. Then we design a heuristic energy aware virtual network migration algorithm called EA-VNM. It answers the following key questions: when performing migration, migrate which virtual nodes to where, and how to perform migration. Extensive simulations show that our algorithm significantly reduces the energy cost by up to 25% over the state-of-the-art algorithm while maintaining similar revenue.

In network virtualization, how to efficiently embed virtual networks with both node and link demands into a shared physical network, namely virtual network embedding, has attracted significant attention. Most of prior studies on this problem have the following two limitations: i) they assumed that the virtual network demands are constant values, which does not hold in real-world network since such demands may vary a lot over time; ii) their primary goal was to generate more revenues for the physical network, with no consideration of the energy cost, which has become a critical issue for the physical network. In this paper, we bridge the gap and study the energy aware virtual network embedding with dynamic demands. Specifically, we first model the dynamics of virtual network demands as a combination of following Gaussian distribution and exhibiting daily diurnal pattern. We then design an efficient heuristic algorithm by leveraging the dynamic characteristic of virtual network demands to minimize the energy consumption while keeping high revenue for the physical network. We implemented our algorithm in C++ and performed side-by-side comparison with prior algorithm. Extensive simulations show that our algorithm can significantly reduce the energy cost by up to 16% over the state-of-the-art algorithm, while maintaining near the same revenue.

This paper studies the problem of energy-efficient scheduling for parallel tasks in high-performance computing systems, such as clusters and data centers. Our goal is to minimize the energy consumption of parallel tasks within a deadline constraint. Among the existing techniques that reduce the energy for computing systems, dynamic voltage and frequency scaling (DVFS) is generally considered as a promising technique that can strike a balance between the energy consumption and the performance for tasks. By using the DVFS technique, the main line of research is to slack the non-critical path tasks to reduce energy consumption of parallel tasks. However, the existing studies slack the tasks greedily and could not efficiently utilize the slack-room, i.e. the idle time of the processors. In this paper, we develop a novel slacking concept, partial optimal slacking (POS), which can take full advantage of the slack-room by slack-sharing. Our formal analysis shows that POS can lead to optimum energy reduction in the partial task set. Based on the POS concept, we propose a new scheduling algorithm for parallel tasks, namely enhanced an energy-efficient scheduling (EES) algorithm. Through extensive evaluation studies, the results demonstrate that the EES algorithm can further improve the energy efficiency of parallel tasks while meeting the deadline constraint.

Abstract Communication systems utilize the distributed hash table (DHT) approach to build the network infrastructure for advantages of even distribution of workload, high scalability, and cost‐effectiveness. Although DHT is undoubtedly applicative in such architectures, some practical distinctions still should be considered to meet the performance requirements of communication infrastructures. This paper focuses on two features of the distributed communication system, the real‐time response and dynamic network maintenance, and proposes the Comb, which is a hierarchical DHT lookup service. Comb's overlay is organized as a two‐layered architecture, workload is distributed evenly among nodes, and most queries can be routed in no more than two hops. Comb is capable to scale to large systems and resilient to fluctuate, it provides a self‐managing and self‐healing mechanism for supporting system recovery from inconsistence. Comb performs effectively with low bandwidth consumption and satisfactory fault tolerance even in a continuously changing environment. Both theoretical analyses and experimental results demonstrate that the two‐layered architecture of Comb is resilient and efficient. Comb improves the performances on routing delay and lookup failure rates with high scalability and availability. Copyright © 2014 John Wiley & Sons, Ltd.

Technologies to support the Machine-to-Machine (M2M) is becoming more important as the need to better understand our environments and make them smart increases. As a result it is predicted that intelligent devices and networks, such as wireless network, will not be isolated but connected and integrated composing computer networks. So far, to enable an End-to-end M2M service, WebSocket has attracted lots of attentions because of its unique full-duplex communications features. Besides, ZigBee technology has widely been deployed in short-range wireless communication systems with its low-power dissipation and high transmission speed. In this paper, we focus on the emerging M2M gateway development for home and industry applications. Specifically, by providing the detailed system architecture and user cases, we give a specific analysis on environmental monitoring implemented with WebSocket and ZigBee technology. The ZigBee sensor network is used to collect the temperature and humidity information. The foreground of the system shows the related data through B/S (Browser/Server) mode by utilizing WebSocket to push the information received by a web server to the client browser.

Energy consumption has become a major concern to the widespread deployment of cloud data centers. The growing importance for parallel applications in the cloud introduces significant challenges in reducing the power consumption drawn by the hosted servers. In this paper, we propose an enhanced energy-efficient scheduling (EES) algorithm to reduce energy consumption while meeting the performance-based service level agreement (SLA). Since slacking non-critical jobs can achieve significant power saving, we exploit the slack room and allocate them in a global manner in our schedule. Using random generated and real-life application workflows, our results demonstrate that EES is able to reduce considerable energy consumption while still meeting SLA.

How to achieve energy efficiency to run a cloud data center is a major challenge in the era of rising electricity cost and environmental protection. Various techniques have been devised to help reduce energy consumption for cloud data centers that consist of a large number of identical servers, including dynamic allocation of active servers, consolidating diverse applications to run on them, and adjusting the CPU speed of an active server. Leveraging these techniques, we use an Online Coloring Bin Packing problem to model the consolidation problem and devise an effective application-aware approximation algorithm to find a near-optimal solution. We show a 1.7 asymptotic approximation ratio. We then apply a Predictive Bayesian Network model to identify daily workload patterns and adjust resource provisioning accordingly. We evaluate our approaches using traces collected from a real data center and demonstrate that (1) our prediction algorithm is effective in estimating future demands, (2) our coordinated approaches can provide significant savings of energy and operational costs close to the near-optimal offline solution, and (3) our approaches incur little reliability costs in term of wear-and-tear of server components.

In light of rapid increase of electricity cost, many business organizations have to find new ways to cut the electricity bill. This paper studies how to reduce the electricity cost in geographical inter-domain virtual network embedding, which embeds virtual networks requested by users to multiple geographically distributed substrate networks run by an infrastructure provider. Previous researches have primarily focused on finding embedding methods to increase revenues by accommodating more virtual network requests, with little attention to reducing the electricity cost. To bridge this gap, we formulate an electricity cost model and design an efficient cost-aware virtual network embedding algorithm by exploiting the location-varying and time-varying diversities of the electricity price and optimizing the energy consumption. Through extensive simulations, we show that our algorithm can significantly reduce the electricity cost by up to 21% over the existing cost-oblivious algorithm, while maintaining nearly the same revenues for the infrastructure provider.

Bulk data transfers account for a large portion of inter-datacenter traffic, such as backups, propagation of bulky updates and migration of data. These bulk data transfers not only consume massive inter-datacenter bandwidth, but also increase the transmission cost of datacenters. To solve this problem, we first employ the max-min fairness to the design of optimal multiple bulk data transfers scheduling algorithm, which leverages the delay tolerance nature of bulk data and reuses dynamic leftover bandwidth to complete multiple bulk data transfers. Then we apply time-expanded technique to transform the problem under a dynamic network into a static network multi-flow problem, and solve it simultaneously from both routing assignment and bandwidth allocation through iterative linear programming approach. Extensive simulations are conducted on a real datacenter topology to demonstrate that our solutions can: 1) improve the network resource utilization; 2) minimize the average bulk data transfer completion time.

Along with the improvement of cloud technologies, we envision that thousands of web applications whether owned by individuals or enterprises will be migrated into cloud. These web applications share the cloud resources while require different quality of service guarantees and management policies. It is challenging while with abundant innovation opportunities given the easy and dynamic provisioning of cloud resources. In order to facilitate practitioners to evaluate, experiment, or enhance existing functionalities of distributed web application management in cloud environment, a componentized distributed service foundation named TANSO is proposed. TANSO targets serving as platform to manage thousands of web application server instances and provide plenty of interfaces for user to extend and experiment new ideas. TANSO has implemented framework components to enable fundamental management. It also provides interfaces to admit and foster user's innovative algorithms and designs for managing large-scale web applications.

Free-riding phenomenon is overwhelming in nowadays P2P network which causes researchers to investigate and develop many approaches to combat it. However, almost all the studies neglect the role of relative contribution of peers, namely willingness of contribution (WoC) in our paper. The ignorant to the ratio of peer's actually contribution to its physical capability would undoubtedly lead to unfairly resource allocation and discourage many peers, as well as loose their support in long term. Therefore, in this paper we present a novel and effective approach Cowtra to address such problem. Our algorithm guarantees that the bandwidth of a source node is distributed properly according to the absolute contribution of the competing peers. Then we adjust the amount of competing peers' received bandwidth by utilizing the WoC in the second phase, such that peers with higher WoC obtain more resource while peers with lower WoC otherwise. At last, simulation results demonstrate the superiority of our algorithm in terms of fairness and efficiency.

The paper realizes a design of digital fuzzy adaptive tuning PID electronic governor. It is based on VHDL description and implemented on FPGA. The paper proposes to detect the speed with duty cycle calculation and use fuzzy-control rules to amend the PID parameters on line to achieve the fuzzy adaptive tuning PID control. On this basis, the paper introduces the implementation of the system in detail, including speed detection, fuzzy logic inference, and fuzzy adaptive tuning PID algorithm. The experiment proves that digital wide-range electronic speed governor implemented on control chip FPGA can be used widely and fuzzy adaptive tuning PID control algorithm can improve the dynamic response of the system, which makes the electronic speed governor under good control in a wider work scope.

The paper studies a digital electronic governor with TMS320F2812 for a gas engine. The governor employed chopped current angular subdivision to realize feeding oil control. Modifying PID parameters automatically by fuzzy control rules, a fuzzy adaptive PID control was achieved. Fuzzy self-adjusted PID control improves dynamic response while the load change, and keeps the governor performing well in a wide operating range