Aristotelis Charalampous, Sotirios Chatzis
Sequence-to-sequence (encoder-decoder) models with attention constitute a cornerstone of deep learning research, as they have enabled unprecedented sequential data modeling capabilities. This effectiveness largely stems from the capacity of these models to infer salient temporal dynamics over long horizons; these are encoded into the obtained neural attention (NA) distributions. However, existing NA formulations essentially constitute point-wise selection mechanisms over the observed source sequences; that is, attention weights computation relies on the assumption that each source sequence element is independent of the rest. Unfortunately, although convenient, this assumption fails to account for higher-order dependencies which might be prevalent in real-world data. This paper addresses these limitations by leveraging Quantum-Statistical modeling arguments. Specifically, our work broadens the notion of NA, by attempting to account for the case that the NA model becomes inherently incapable of discerning between individual source elements; this is assumed to be the case due to higher-order temporal dynamics. On the contrary, we postulate that in some cases selection may be feasible only at the level of pairs of source sequence elements. To this end, we cast NA into inference of an attention density matrix (ADM) approximation. We derive effective training and inference algorithms, and evaluate our approach in the context of a machine translation (MT) application. We perform experiments with challenging benchmark datasets. As we show, our approach yields favorable outcomes in terms of several evaluation metrics.