{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/backpropagating-through-structured-argmax","title":"Backpropagating through Structured Argmax using a SPIGOT","arxiv_id":"1805.04658","date":"2018-05-12","proceeding":"ACL 2018 7","authors":["Hao Peng","Sam Thomson","Noah A. Smith"],"abstract":"We introduce the structured projection of intermediate gradients optimization\ntechnique (SPIGOT), a new method for backpropagating through neural networks\nthat include hard-decision structured predictions (e.g., parsing) in\nintermediate layers. SPIGOT requires no marginal inference, unlike structured\nattention networks (Kim et al., 2017) and some reinforcement learning-inspired\nsolutions (Yogatama et al., 2017). Like so-called straight-through estimators\n(Hinton, 2012), SPIGOT defines gradient-like quantities associated with\nintermediate nondifferentiable operations, allowing backpropagation before and\nafter them; SPIGOT's proxy aims to ensure that, after a parameter update, the\nintermediate structure will remain well-formed.\n  We experiment on two structured NLP pipelines: syntactic-then-semantic\ndependency parsing, and semantic parsing followed by sentiment classification.\nWe show that training with SPIGOT leads to a larger improvement on the\ndownstream task than a modularly-trained pipeline, the straight-through\nestimator, and structured attention, reaching a new state of the art on\nsemantic dependency parsing.","url_abs":"http://arxiv.org/abs/1805.04658v1","url_pdf":"http://arxiv.org/pdf/1805.04658v1.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"backpropagating-through-structured-argmax","repo_url":"https://github.com/Noahs-ARK/SPIGOT","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"none","reach":{"status":"ok","spdx":"MIT"}}],"tasks":[{"task_slug":"dependency-parsing","task_name":"Dependency Parsing"},{"task_slug":"reinforcement-learning","task_name":"Reinforcement Learning"},{"task_slug":"semantic-dependency-parsing","task_name":"Semantic Dependency Parsing"},{"task_slug":"semantic-parsing","task_name":"Semantic Parsing"},{"task_slug":"sentiment-analysis","task_name":"Sentiment Analysis"},{"task_slug":"sentiment-classification","task_name":"Sentiment Classification"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1805.04658","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}