{"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/a-continuous-relaxation-of-beam-search-for","title":"A Continuous Relaxation of Beam Search for End-to-end Training of Neural Sequence Models","arxiv_id":"1708.00111","date":"2017-08-01","proceeding":null,"authors":["Kartik Goyal","Graham Neubig","Chris Dyer","Taylor Berg-Kirkpatrick"],"abstract":"Beam search is a desirable choice of test-time decoding algorithm for neural\nsequence models because it potentially avoids search errors made by simpler\ngreedy methods. However, typical cross entropy training procedures for these\nmodels do not directly consider the behaviour of the final decoding method. As\na result, for cross-entropy trained models, beam decoding can sometimes yield\nreduced test performance when compared with greedy decoding. In order to train\nmodels that can more effectively make use of beam search, we propose a new\ntraining procedure that focuses on the final loss metric (e.g. Hamming loss)\nevaluated on the output of beam search. While well-defined, this \"direct loss\"\nobjective is itself discontinuous and thus difficult to optimize. Hence, in our\napproach, we form a sub-differentiable surrogate objective by introducing a\nnovel continuous approximation of the beam search decoding procedure. In\nexperiments, we show that optimizing this new training objective yields\nsubstantially better results on two sequence tasks (Named Entity Recognition\nand CCG Supertagging) when compared with both cross entropy trained greedy\ndecoding and cross entropy trained beam decoding baselines.","url_abs":"http://arxiv.org/abs/1708.00111v2","url_pdf":"http://arxiv.org/pdf/1708.00111v2.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":[],"tasks":[{"task_slug":"ccg-supertagging","task_name":"CCG Supertagging"},{"task_slug":"motion-segmentation","task_name":"Motion Segmentation"},{"task_slug":"named-entity-recognition-1","task_name":"Named Entity Recognition"},{"task_slug":"named-entity-recognition-ner","task_name":"Named Entity Recognition (NER)"},{"task_slug":"named-entity-recognition","task_name":"named-entity-recognition"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[{"leaderboard":"/sota/motion-segmentation-on-hopkins155","task":"Motion Segmentation","dataset":"Hopkins155","model":"T-Linkage","rank_in_archive_order":3,"of":4,"metrics":{"Classification Error":"1.97"},"uses_additional_data":false}],"syntology":{"syntology_url":"https://syntology.ai/paper/1708.00111","atlas_url":"https://app.syntology.ai/?focus=1708.00111","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}