I have implemented an LSTM network for transliteration that is using Seq2Seq approach. But its only giving me one output sequence for a given input sequence. I can not see how can I obtain more than one output sequence and their associated probability values using an LSTM network. Or if it is even possible to obtain multiple solutions using this approach. I have a training data in form of English words and transliterations of these words in a second language other than English. I can see that people have been using LSTM for problems like next word suggestion where more than one answer is possible. So it should be possible to achieve the same for transliteration also.
我已经使用Seq2Seq方法实现了用于音译的LSTM网络。但它只给出了给定输入序列的一个输出序列。我看不出如何使用LSTM网络获得多个输出序列及其相关的概率值。或者，如果甚至可以使用这种方法获得多种解决方案。我有一个英语单词形式的培训数据，这些单词的音译用英语以外的第二语言。我可以看到人们一直在使用LSTM来解决诸如下一个单词建议之类的问题，其中可能有多个答案。所以音译也应该可以达到同样的效果。
1 个解决方案
#1
1
You can use "beam search" method to generate variable results.
您可以使用“波束搜索”方法生成变量结果。
There are many materials that introduce it well and here is one written by CMU, "Neural Machine Translation and Sequence-to-sequence Models: A Tutorial", Chapter 7.2.
有许多材料可以很好地介绍它，这里有一个由CMU编写的“神经机器翻译和序列到序列模型：教程”，第7.2章。
Recently, tensorflow seq2seq model implemented the beam search clearly, you can refer to Tensorflow-seq2seq-BeamSearchDecoder.
最近，tensorflow seq2seq模型清楚地实现了光束搜索，你可以参考Tensorflow-seq2seq-BeamSearchDecoder。
#1
1
You can use "beam search" method to generate variable results.
您可以使用“波束搜索”方法生成变量结果。
There are many materials that introduce it well and here is one written by CMU, "Neural Machine Translation and Sequence-to-sequence Models: A Tutorial", Chapter 7.2.
有许多材料可以很好地介绍它，这里有一个由CMU编写的“神经机器翻译和序列到序列模型：教程”，第7.2章。
Recently, tensorflow seq2seq model implemented the beam search clearly, you can refer to Tensorflow-seq2seq-BeamSearchDecoder.
最近，tensorflow seq2seq模型清楚地实现了光束搜索，你可以参考Tensorflow-seq2seq-BeamSearchDecoder。