Search

Multivariate time series segmentation (also known as multiple change-point detection), is one of the important basic tasks in time series data mining. A time series is an ordered sequence of elements. A segment is a region of the time series where the observations are similar according to some measure. The aim of time series segmentation is to identify these regions. In the field of wearable device application, time series segmentation is applied to behavior switching discovering, activity recognition, dimension reduction and so on. With the widespread use of wearable devices, large amount of sensor data can be available. To effectively utilize this time series data, it is important to segment the data into homogeneous phases prior to build an appropriate model for the task, and it must be done at large scale. In this paper, a high accuracy and low time complexity time series segmentation method is proposed based on the time series feature extraction and segmented gaussian model, and applies it to the human activity data sets to discover activity switching.

Relation classification is an important ingredient task in the construction of knowledge graph, question answering system and numerous other natural language processing (NLP) tasks. With the application of deep neural networks (DNNs) such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs), relation classification task has achieved satisfactory results. However, many proposed models can not take well advantages of multiple window sizes for filters in CNNs and finally hurt the performance of this task. Moreover, unlike public and general dataset that has a large quantity of instances from natural languages or daily conversations, the performances of many deep neural networks with high complexity are not well enough for a small corpus in specific fields. To work out these problems, we propose a novel CNN model with attention mechanism for multi-window-sized kernels to capture the most important information and test our system not only on a general dataset of SemEval 2010 but also on a small dataset built from Chinese fundamentals of electric circuits textbook artificially. The experimental results show that our system outperforms the baseline systems for the SemEval 2010 relation classification task and validate the effectiveness of CNN on the specific Chinese small corpus relation classification task.

Neural abstractive summarization models, based on the attentional sequence-to-sequence model, can generate simple summarization from source text. However, these models inevitably face with repetition problem, and ignore the rich structure information contained in target summarization. To alleviate these problems, we propose a generation model that combines the coverage mechanism and neural variational inference to make full use of all information in sentences. First, we employ Variation Auto-Encoders(VAEs) for generating latent variables, which obtain the latent structure information in target summarization. Second, we introduced coverage mechanism in final decoding process, which is expected greatly improve the weights of those important parts in the encoder and effectively reduce repetition problem. Abstractive summaries are generate by combining the generated latent structure information and the coverage information. Experimental results on LCSTS dataset show that our model outperforms baseline methods.

Different from the traditional wireless communication network, wireless sensor networks(WSNs) have no infrastructure similar to the cellular network, and nodes can only communicate in the way of "flooding", which is easy to cause the problem of broadcast storm. The energy of wireless sensor nodes is actually very limited, which seriously affects the lifecycle of the whole network. Therefore, the research direction of constructing virtual backbone network in WSN arises at the right moment. In this paper, a minimum connected dominating set(CDS) construction algorithm based on multiple dominating tree is proposed. The algorithm firstly takes the node at the core position in each local area of the network as the root node, and constructs several dominating subtrees in the network. Then, some connector is selected to connect the two subtrees that are not connected to each other. Finally, the redundant dominator is cut off to obtain the final virtual backbone.

Deep neural network has been widely used in a variety of natural language processing (NLP) tasks nowadays. As one of the most import research areas, entity relation extraction applies usual recurrent neural networks (RNNs) and convolutional neural networks (CNNs) and has achieved good results. Most relation extraction tasks are about public and general datasets, they are usually natural languages or daily conversations, and have millions of samples, very few relates to small corpus in a specific field. We hope to construct a knowledge graph about Chinese fundamentals of electric circuits textbook for beginners. The knowledge graph shows students knowledge navigation and consists of important concepts about this field and logical relationships between them. To achieve the goal, the first step is to ensure entities and extract entity relationships from raw corpus automatically. In this paper, a relation extraction dataset is built from Chinese fundamentals of electric circuits textbook artificially and research the relation extraction performance of improved position-enhanced CNN model on this task. The experiment result validates the effectiveness of CNN on specific Chinese small corpus relation extraction task.