In speech recognition, confidence annotation adopts a single confidence feature or a combination of different features for classification. These confidence features are always extracted from decoding information. However, it is proved that about 30% of knowledge of human speech understanding is mainly derived from high-level information. Thus, how to extract a high-level confidence feature statistically independent of decoding information is worth researching in speech recognition. In this paper, a novel confidence feature extraction algorithm based on latent topic similarity is proposed. Each word topic distribution and context topic distribution in one recognition result is firstly obtained using the latent Dirichlet allocation (LDA) topic model, and then, the proposed word confidence feature is extracted by determining the similarities between these two topic distributions. The experiments show that the proposed feature increases the number of information sources of confidence features with a good information complementary effect and can effectively improve the performance of confidence annotation combined with confidence features from decoding information.

A novel method to simplify decision functions of support vector machines (SVMs) is proposed in this paper. In our method, a decision function is determined first in a usual way by using all training samples. Next those support vectors which contribute less to the decision function are excluded from the training samples. Finally a new decision function is obtained by using the remaining samples. Experimental results show that the proposed method can effectively simplify decision functions of SVMs without reducing the generalization capability.

Text sentiment classification aims to automatically classify subjective documents into different sentiment-oriented categories (e.g. positive/negative). Given the high dimensionality of features describing documents, how to effectively select the most useful ones, referred to as sentiment-bearing features, with a lack of sentiment class labels is crucial for improving the classification performance. This paper proposes an unsupervised sentiment-bearing feature selection method (USFS), which incorporates sentiment discriminant analysis (SDA) into sentiment strength calculation (SSC). SDA applies traditional linear discriminant analysis (LDA) in an unsupervised manner without losing local sentiment information between documents. We use SSC to calculate the overall sentiment strength for each single feature based on its affinities with some sentiment priors. Experiments, performed using benchmark movie reviews, demonstrated the superior performance of USFS.

As scientific and technological resources are experiencing information overload, it is quite expensive to find resources that users are interested in exactly. The personalized recommendation system is a good candidate to solve this problem, but data sparseness and the cold starting problem still prevent the application of the recommendation system. Sparse data affects the quality of the similarity measurement and consequently the quality of the recommender system. In this paper, we propose a matrix factorization recommendation algorithm based on similarity calculation(SCMF), which introduces potential similarity relationships to solve the problem of data sparseness. A penalty factor is adopted in the latent item similarity matrix calculation to capture more real relationships furthermore. We compared our approach with other 6 recommendation algorithms and conducted experiments on 5 public data sets. According to the experimental results, the recommendation precision can improve by 2% to 9% versus the traditional best algorithm. As for sparse data sets, the prediction accuracy can also improve by 0.17% to 18%. Besides, our approach was applied to patent resource exploitation provided by the wanfang patents retrieval system. Experimental results show that our method performs better than commonly used algorithms, especially under the cold starting condition.

To find the exact items from the massive patent resources for users is a matter of great urgency. Although the recommender systems have shot this problem to a certain extent, there are still some challenging problems, such as tracking user interests and improving the recommendation quality when the rating matrix is extremely sparse. In this paper, we propose a novel method called Collaborative Filtering Auto-Encoder for the top-N recommendation. This method employs Auto-Encoders to extract the item's features, converts a high-dimensional sparse vector into a low-dimensional dense vector, and then uses the dense vector for similarity calculation. At the same time, to make the recommendation list closer to the user's recent interests, we divide the recommendation weight into time-based and recent similarity-based weights. In fact, the proposed method is an improved, item-based collaborative filtering model with more flexible components. Experimental results show that the method consistently outperforms state-of-the-art top-N recommendation methods by a significant margin on standard evaluation metrics.

The symbol flipping decoding algorithms based on prediction (SFDP) for non-binary LDPC codes perform well in terms of error performances but converge slowly when compared to other symbol flipping decoding algorithms. In order to improve the convergence rate, we design new flipping rules with two phases for the SFDP algorithms. In the first phase, two or more symbols are flipped at each iteration to allow a quick increase of the objective function. While in the second phase, only one symbol is flipped to avoid the oscillation of the decoder when the objective function is close to its maximum. Simulation results show that the SFDP algorithms with the proposed flipping rules can reduce the average number of iterations significantly, whereas having similar performances when compared to the original SFDP algorithms.

Analog Hopfield neural networks (HNNs) have so far been used to solve many kinds of optimization problems, in particular, combinatorial problems such as the TSP, which can be described by an objective function and some equality constraints. When we solve a minimization problem with equality constraints by using HNNs, however, the constraints are satisfied only approximately. In this paper we propose a circuit which rigorously realizes the equality constraints and whose energy function corresponds to the prescribed objective function. We use the SPICE program to solve circuit equations corresponding to the above circuits. The proposed method is applied to several kinds of optimization problems and the results are very satisfactory.

In this letter, a novel collaborative representation graph based on the local and global consistency label propagation method, denoted as CRLGC, is proposed. The collaborative representation graph is used to reduce the cost time in obtaining the graph which evaluates the similarity of samples. Considering the lacking of labeled samples in real applications, a semi-supervised label propagation method is utilized to transmit the labels from the labeled samples to the unlabeled samples. Experimental results on three image data sets have demonstrated that the proposed method provides the best accuracies in most times when compared with other traditional graph-based semi-supervised classification methods.

In this letter, a new method is presented to suppress fractional pseudocodewords by eliminating small instantons of irregular low-density parity-check (LDPC) codes under the linear programming (LP) decoding over the binary symmetric channel (BSC). By appending several new rows found by the integer linear programming formulation to the original parity-check matrix, the optimal distribution spectrum of BSC-instantons in the modified code is obtained. Simulation results show that the proposed method can improve the fractional distance of parity-check matrices and considerably enhance the error-correcting performance of irregular LDPC codes under the LP decoding at the cost of a slightly loss of the original code rate.

In this letter, we present a new scheme to find small fundamental instantons (SFIs) of regular low-density parity-check (LDPC) codes for the linear programming (LP) decoding over the binary symmetric channel (BSC). Based on the fact that each instanton-induced graph (IIG) contains at least one short cycle, we determine potential instantons by constructing possible IIGs which contain short cycles and additional paths connected to the cycles. Then we identify actual instantons from potential ones under the LP decoding. Simulation results on some typical LDPC codes show that our scheme is effective, and more instantons can be obtained by the proposed scheme when compared with the existing instanton search method.

It inevitably comes out information overload problem with the increasing available data on e-commence websites. Most existing approaches have been proposed to recommend the users personal significant and interesting items on e-commence websites, by estimating unknown rating which the user may rate the unrated item, i.e., rating prediction. However, the existing approaches are unable to perform user prediction and item prediction, since they just treat the ratings as real numbers and learn nothing about the ratings' embeddings in the training process. In this paper, motivated by relation prediction in multi-relational graph, we propose a novel embedding model, namely RPEM, to solve the problem including the tasks of rating prediction, user prediction and item prediction simultaneously for recommendation systems, by learning the latent semantic representation of the users, items and ratings. In addition, we apply the proposed model to cross-domain recommendation, which is able to realize recommendation generation in multiple domains. Empirical comparison on several real datasets validates the effectiveness of the proposed model. The data is available at https://github.com/yuzhaour/da.

Knowledge graph (KG) embedding aims at learning the latent semantic representations for entities and relations. However, most existing approaches can only be applied to KG completion, so cannot identify relations including unseen entities (or Out-of-KG entities). In this paper, motivated by the zero-shot learning, we propose a novel model, namely JointE, jointly learning KG and entity descriptions embedding, to extend KG by adding new relations with Out-of-KG entities. The JointE model is evaluated on entity prediction for zero-shot embedding. Empirical comparisons on benchmark datasets show that the proposed JointE model outperforms state-of-the-art approaches. The source code of JointE is available at https://github.com/yzur/JointE.