Input Combination Strategies for Multi-Source Transformer Decoder
This work addresses multi-source sequence-to-sequence tasks for NLP applications like translation, but it is incremental as it extends existing methods from recurrent architectures to Transformers.
The paper tackled the problem of integrating multiple sources in Transformer-based sequence-to-sequence tasks by proposing four input combination strategies (serial, parallel, flat, and hierarchical) for encoder-decoder attention. The result showed that these models improved over single-source baselines in multimodal translation and multi-source language translation tasks.
In multi-source sequence-to-sequence tasks, the attention mechanism can be modeled in several ways. This topic has been thoroughly studied on recurrent architectures. In this paper, we extend the previous work to the encoder-decoder attention in the Transformer architecture. We propose four different input combination strategies for the encoder-decoder attention: serial, parallel, flat, and hierarchical. We evaluate our methods on tasks of multimodal translation and translation with multiple source languages. The experiments show that the models are able to use multiple sources and improve over single source baselines.