Statistics-based machine translation method, apparatus and electronic device

An apparatus comprising:

• one or more processors;
• memory;
• an acquisition unit stored in the memory and executable by the one or more processors to obtain parallel corpus;
• a parsing unit stored in the memory and executable by the one or more processors to parse a plurality of sentences included in the parallel corpus, and obtain a plurality of words included in the parallel corpus;
• a setting unit stored in the memory and executable by the one or more processors to:
• set a plurality of word vectors for the plurality of words included in the parallel corpus in a first preset dimension to form correspondence relationships between the plurality of words and the plurality of word vectors, and
• randomly set a plurality of sentence vectors for the plurality of sentences in a second preset dimension;
• a construction unit stored in the memory and executable by the one or more processors to construct first neural net language model samples corresponding to first N-gram phrases included in source language sentences and second neural net language model samples corresponding to first N-gram phrases included in target language sentences in the parallel corpus based at least in part on the correspondence relationships between the plurality of words and the plurality of word vectors, first sentence vectors corresponding to the source language sentences and second sentence vectors corresponding to the target language sentences in the parallel corpus, a first preset neural net language model corresponding to a source language and a second preset neural net language model corresponding to a target language;
• and
• a learning unit stored in the memory and executable by the one or more processors to train the neural net language model corresponding to the source language and the neural net language model corresponding to the target language by maximizing a sum of likelihood probabilities of the first neural net language model samples corresponding to the first N-gram phrases included in the source language sentences in the parallel corpus and a sum of likelihood probabilities of the second neural net language model samples corresponding to the second N-gram phrases comprised in the target language sentences in the parallel corpus, while minimizing a sum of distances between the first sentence vectors of the source language sentences and the second sentence vectors of the target language sentences in pairs in the parallel corpus as a training goal to obtain the correspondence relationships between the plurality of words and the plurality of word vectors.