Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2020-09-28 Cooperative journals: 《计算机应用研究》
Abstract： To improve the performance of multi-accent Mandarin speech recognition task, this paper present a method for hybrid end-to-end automatic speech recognition(ASR) by combining Connectionist Temporal Classification (CTC) and MutiHead Attention by using a multiobjective training and joint decoding. Our analysis shows that hybrid model with lower CTC weight and deeper encoder layers performance better learning capacity. And we trained a very deep models with up to 48 layers for encode-decoder Architecture, which outperform all previous end-to-end ASR approaches on Aidatatang 200h multi-accent dataset, achieve 5.6% Character Error Rate(CER) and 26.2% Sentence Error Rate(SER) . The experiment proves that the recognition rate of the end-to-end model proposed in this paper exceeds the general end-to-end model, and it has certain advancedness in solving the Mandarin recognition with accents.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2019-05-10 Cooperative journals: 《计算机应用研究》
Abstract： In sound event detection tasks, the target event was susceptible to background noise, and was not present in a significantly high portion of time frames of each signal. To solve the problem, this paper proposed a new method of sound event detection based on one-dimensional Dense Convolutional Network(DenseNet) with multi-level attention mechanism. Firstly, it used the one-dimensional DenseNet for frame-wise detection, which was effective in finding the precise onset and offset time. Then, it embedded the multi-level attention mechanism in the one-dimensional DenseNet model, which made the attention-aware features from different modules change adaptively as layers went deeper. Therefore, the model could automatically select and attend on important frames for the targets while ignoring the unrelated parts (e. g. , the background noise segments) . Finally, this work evaluated the model using DCASE 2017 Task 2 development dataset. Results show that the overall performance of the method has further improved than the conventional deep learning method.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2019-04-01 Cooperative journals: 《计算机应用研究》
Abstract： In recent years, the development of social networks had promoted research in many fields, such as public opinion monitoring, advertising recommendation and opinion leader identification etc. The influence measurement of social network users is the basis of the above research. This paper integrated the basic attributes of user, interaction behavior of user and user’s microblog content into the PageRank algorithm, therefore, it proposed a multi-dimensional user influence measurement algorithm:MDIR(multi-dimension influence rank). The experiment shows that, the MDIR can reflect the actual influence of microblog users more comprehensively and realistically than other five commonly used influence measurement algorithms.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-05-24 Cooperative journals: 《计算机应用研究》
Abstract： This paper proposed a collaborative topic regression model based on the preference for frequent topic sets to address the item-cold-start problem in academic paper recommendation. The algorithm takes into account the user's preference for research hotspots when selecting academic papers, and uses frequent topic sets to represent research hotspots. So, user's preference for research hotspots is expressed as the user's preference for frequent topic sets. Firstly, the papers-topic probability distribution matrix is obtained through LDA algorithm and filter out the topics with higher probability in the paper. Then, the algorithm mines the frequently-occurring topic sets and gets the relationships between papers and frequent topic sets. Finally, the user's preference for frequent topic sets is used for the prediction of unknown scores. Experiments on CiteULike datasets show that the algorithm improves the recall, accuracy and RMSE over the matrix factorization model and the collaborative topic regression model.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-04-24 Cooperative journals: 《计算机应用研究》
Abstract： Aiming at the problem of cold start caused by the addition of new item in the recommendation system, This paper proposed a collaborative filtering recommendation model USPTMF-CFIA based on matrix factorization model, which combines the similarity of item image and category attributes . First, it used the matrix factorization model based on users’ preference and time weight to predict and fill the missing item. Then, it used the VGG16 neural network to extract the features of the item images and combines category attributes to calculate the similarity between the new item and the historical items, then got the item’s neighbors. Finally, the new item is predicted based on the similarity between the new item and the neighbors, and the first N items with high score are recommended to the correspond user. The experiment on the dataset provided by GroupLens proved that the proposed accuracy rate of this model. The recommended accuracy of this model is 0.006~0.015 higher than the MAP-BPR model , 0.02~0.028 higher than the traditional collaborative filtering model and 0.001 ~ 0.003 higher than that of the USPTMF-CFA model without image similarity0.001~0.002 higher than ACMF model.
Hits
Hits
Downloads
Comment