Attention-based neural joint source-channel coding of text for point to point and broadcast channel

In this work, we consider the transmissions of structured data such as text over a noisy channel and correlated texts over a broadcast channel. As the separate source-channel coding principle no longer holds in such scenarios, we propose a joint source-channel coding scheme which is based on deep learning architecture. In order to enhance the convergence speed, we adopt the bidirectional gated recurrent unit at the encoder. For the decoder, to improve the recovery quality, we propose the following two types of strategies: (1) After a unidirectional neural network based decoder is used, a generative adversarial network is applied to train the whole joint source-channel coding framework and pointwise mutual information is added to the objective function of beam search process; (2) Rather than using a unidirectional neural network-based decoder, we develop a bidirectional neural network based and bidirectional attention mechanism integrated decoder to utilize past and future information. Experiments under different types of channels show that our schemes are superior to the existing deep learning joint source-channel coding method and in the case of low bit budget, long sentence length and small channel signal to noise ratio, our models are significantly superior to those of separate source-channel coding. In addition, we extend the proposed unidirectional and bidirectional decoders to the broadcast channel. Additionally, to improve the performance of unidirectional decoding, we utilize not only the correlation between adjacent words in the same text but also the correlation between words in different languages with the same meaning in the beam search process.