Search

Fairness in machine learning has received increasing attention in recent years. This study focuses on a particular type of machine learning fairness, namely sentimental bias, in text sentiment analysis. Sentimental bias occurs on words (or phrases) when they are distributed distinctly in positive and negative corpora. It results in that an excessively proportion of words carry negative/positive sentiment in learned models. This study proposed a new attention mechanism, called polar attention, to mitigate sentimental biases. It consists of two modules, namely polar flipping and distance measurement. The first module explicitly models word sentimental polarity and can prevent that neutral words flip positively or negatively. The second module is used to attend negative/positive words. In the experiments, three benchmark data sets are used, and supplementary testing sets are compiled. Experimental results verify the effectiveness of the proposed method.

Network compression coding technology is a research hotspot in the field of digital steganography and image synthesis. How to improve image quality while achieving short compression time is a problem currently faced. Based on the improved wavelet neural network theory, this paper constructs a digital steganography and image synthesis model. The model first tracks the contour of the digit to be recognized, then equalizes and resamples the contour to make it translation-invariant and scaling-invariant, and then uses multi-wavelet neural network clusters to stretch the contour shell to obtain orders of magnitude multi-resolution and its average, and finally, these shell coefficients are fed into a feedforward neural network cluster to identify this handwritten digit, solving the problem of multi-resolution decomposition of contour shells while having a high sampling rate. In the simulation process, the classification model that a single pixel is a text/non-text pixel is trained on the original gray value of the gray pixel and its neighboring pixels, and the classified text pixels are connected to a text area through an adaptive MeanShift method. The experimental results show that it is feasible to use multi-wavelet features for handwritten digit recognition. The model combines the neural network and the genetic algorithm, making full use of the advantages of both, so that the new algorithm has the learning ability and robustness of the neural network. The compression ratio after compression by ordinary wavelet coding, PSNR, MSE, and compression time are 8.4, 25 dB, 210, and 7 s, respectively. The values are 11.7, 24 dB, 207, and 11 s, respectively. At the same time, the peak signal-to-noise ratio is higher and the mean square error is lower, that is, the compression quality is better, and the accuracy of digital steganography and image synthesis is effectively improved.

Evaluating academic papers and groups is important in scholar evaluation and literature retrieval. However, current evaluation indices, which pay excessive attention to the citation number rather than the citation importance and unidirectionality, are relatively simple. This study proposes new evaluation indices for papers and groups. First, an improved PageRank (PR) algorithm introducing citation importance is proposed to obtain a new citation-based paper index (CPI) via a pre-ranking and fine-tuning strategy. Second, to evaluate the paper’s influence inside and outside its research field, the focus citation-based paper index (FCPI) and diversity citation-based paper index (DCPI) are proposed based on topic similarity and diversity, respectively. Third, aside from the statistical indices for academic papers, we propose a foreign academic degree of dependence (FAD) to characterise the dependence between two academic groups. Finally, artificial intelligence (AI) papers from 2005 to 2019 are utilised for a case study.

The traditional literature analysis mainly focuses on the literature metadata such as topics, authors, keywords, references, and rarely pays attention to the main content of papers. However, in many scientific domains such as science, computing, engineering, the methods and datasets involved in the papers published carry important information and are quite useful for domain analysis and recommendation. Method and dataset entities have various forms, which are more difficult to recognize than common entities. In this paper, we propose a novel Method and Dataset Entity Recognition model (MDER), which is able to effectively extract the method and dataset entities from the main textual content of scientific papers. The model is the first to combine rule embedding, a parallel structure of Convolutional Neural Network (CNN) and a two-layer Bi-directional Long Short-Term Memory (BiLSTM) with the self-attention mechanism. We evaluate the proposed model on datasets constructed from the published papers of different research areas in computer science. Our model performs well in multiple areas and features a good capacity for cross-area learning and recognition. Ablation study indicates that building different modules collectively contributes to the good entity recognition performance as a whole. The data augmentation positively contributes to model training, making our model much more robust. We finally apply the proposed model on PAKDD papers published from 2009–2019 to mine insightful results over a long time span. 1