Differences

This shows you the differences between two versions of the page.

--- nlp:beam_search [2021/04/20 16:44] – [History] jmflanig
+++ nlp:beam_search [2024/08/25 23:53] (current) – [Papers] jmflanig
@@ Line 4: / Line 4: @@
 ===== Papers =====
   * **Beam search for RNNs**: [[https://arxiv.org/pdf/1509.00685.pdf|Rush et al 2015 - A Neural Attention Model for Abstractive Sentence Summarization]]  This may be the first paper to use beam search in seq2seq models (as far as Jeff knows)
+  * **[[https://arxiv.org/pdf/1708.00111|Goyal et al 2017 - A Continuous Relaxation of Beam Search for End-to-end Training of Neural Sequence Models]]**
 ===== History =====
 Invented for decoding in speech recognition, beam search was the de-facto decoding algorithm for statistical machine translation and was also used for statistical parsing.   Historical references are here:
+  * Speech Recognition
+    * [[http://jpk.pku.edu.cn/course/sjjg/chapter2/resource/A%20fast%20sequential%20decoding%20algorithm%20using%20a%20stack.pdf|Jelinek 1969 - Fast Sequential Decoding Algorithm Using a Stack]] This is the standard stack-based beam search decoder, used extensively in statistical MT. Cited and used by [[https://aclanthology.org/1988.tmi-1.19.pdf|Brown et al 1988]] for MT.  Didn't use the term beam search however.
   * Machine translation, see [[Statistical Machine Translation]]
     * [[https://www.aclweb.org/anthology/N03-1017.pdf|Och et al 2003 - Statistical Phrase-Based Translation]]