2020/10/23 by Chunchuan Lyu, Shay B. Cohen, Lyu, Chunchuan +3 · 1 citation
Biochemistry, Genetics and Molecular Biology · Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning in Bioinformatics #Natural Language Processing Techniques #Topic Modeling
paper · pdf · doi:10.48550/arxiv.2010.12676
openalex publication_date 2020/10/23 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
Abstract Meaning Representations (AMR) are a broad-coverage semantic\nformalism which represents sentence meaning as a directed acyclic graph. To\ntrain most AMR parsers, one needs to segment the graph into subgraphs and align\neach such subgraph to a word in a sentence; this is normally done at\npreprocessing, relying on hand-crafted rules. In contrast, we treat both\nalignment and segmentation as latent variables in our model and induce them as\npart of end-to-end training.\n As marginalizing over the structured latent variables is infeasible, we use\nthe variational autoencoding framework.\n To ensure end-to-end differentiable optimization, we introduce a\ndifferentiable relaxation of the segmentation and alignment problems. We\nobserve that inducing segmentation yields substantial gains over using a\n`greedy' segmentation heuristic. The performance of our method also approaches\nthat of a model that relies on the segmentation rules of\n citetlyu-titov-2018-amr, which were hand-crafted to handle individual AMR\nconstructions.\n