Four deep learning trends from ACL 2017


Though attention often plays the role of word alignment in NMT, Koehn and Knowles note that it learns to play other, harder-to-understand roles too; thus it is not always as understandable as we might hope. In Parameter Free Hierarchical Graph-Based Clustering for Analyzing Continuous Word Embeddings, Trost and Klakow perform clustering on word embeddings, then cluster those clusters, and so on to obtain a hierarchical tree-like structure. Neural networks are powerful because they can learn arbitrary continuous representations, but humans find discrete information – like language itself – easier to understand. These systems should ideally produce a proof or derivation of the answer – for a semantic parsing question answering system, this might be the semantic parse itself, or a relevant excerpt from the knowledge base.