Search

Sentiment analysis (SA) is the study of people's emotions and attitudes toward a particular topic. It is beneficial for monitoring and analyzing social media text in order to gather public opinion. Despite the fact that there are SA applications for monolingual text such as English and non-English languages like Hindi, Chinese and French, the Malay language has far fewer works, not to mention the mixed language such as Malay-English (also known as Manglish). Other than comments and posts from websites and social media, the emoji used by internet users can also help to provide better insights into how they truly feel about a particular topic. Our work focuses on Malay-English mixed language comments and posts on how Malaysians feel about daily new cases of Covid-19 in Malaysia. We proposed a neural network framework to perform SA on languages spoken by Malaysians, namely Malay, English, and Malay-English, by also taking into account the emoji used by internet users. The data was pre-processed to remove noises and then transformed into word vector representation using word embedding technique. Then we propose a framework that involves training and testing mixed language textual data along with emoji analysis by using bidirectional Long Short Term Memory (biLSTM) neural network. To compare with the proposed method, several machine learning models and Long Short Term Memory (LSTM) with word vectorization was used. Finally, compared to the machine learning model such as Naïve Bayes and Logistic Regression, neural networks such as LSTM, the proposed method; biLSTM with tuned hyper-parameter for Malay-English mixed language achieved the highest accuracy of 76.6%, and macro F1-score of 69.6%. [ABSTRACT FROM AUTHOR]

In this paper, a novel single channel source separation using two-dimensional nonnegative matrix factorization (NMF2D) is proposed. In NMF2D, the time-frequency (TF) profile of each source is modeled as two-dimensional convolution of the temporal code and the spectral basis. The proposed model used Beta-divergence as a cost function and updated by maximizing the joint probability of the mixing spectral basis and temporal codes using the multiplicative update rules. Results have concretely shown the effectiveness of the algorithm in blindly separating the audio sources from single channel mixture. [ABSTRACT FROM AUTHOR]