Simultaneous translation device and computer program
AssigneeNational Institute of Information and Communications Technology
InventorsXiaolin Wang, Masao Uchiyama
Life of this patent
Estimating the term…
Abstract
A simultaneous translation system includes: an encoder encoding an input word sequence to an intermediate language representation; a chunk-end detecting device detecting an end of a chunk in the word sequence; a word vector reading unit inputting a partial word sequence up to the chunk-end detected by the chunk-end detecting device to the encoder; a decoder and a translated word searching unit receiving the intermediate language representation from encoder as an input, for outputting a translation word sequence corresponding to the partial word sequence; and a translated word sequence storage unit storing the translation word sequences output by decoder and translated word searching unit.
Claims5
A simultaneous translation device, comprising:
- a processor;
- and
- a program memory connected to the processor and configured to store a computer program that, when executed by the processor, causes the processor to perform functions of:
- an encoding function for encoding an input word sequence to an intermediate language representation;
- a chunk-end detecting function for detecting an end of each chunk in said word sequence on a real-time basis;
- a word sequence input function for inputting a partial word sequence to said encoding function each time said chunk-end detecting function detects an end of a chunk, said partial word sequence consisting of a part of said input word sequence from a head of said input word sequence up to the chunk of which end is detected by said chunk-end detecting function;
- a decoding function for outputting a translation word sequence of a prescribed language corresponding to said partial word sequence in response to each said intermediate language representation outputted by said word sequence input function;
- and
- a translation word sequence storage function for storing said translation word sequence output from said decoding function;
- wherein
- said decoding function includes a sequential decoding function responsive to said intermediate language representation output from said encoding function for searching for a translation word sequence by following trees of translated word candidate sequences based on a probability distribution which quantifies a likelihood of each word appearing given a preceding word or words, and for sequentially outputting translation word sequences in the prescribed language corresponding to said partial word sequence, wherein
- said sequential decoding function selects only those translated word candidate sequences that start with said translation word sequence stored in said translation word sequence storage function, and
- wherein said chunk-end detection function is performed by a neural network trained using word vector sequences in which a word designated to be a chunk-end is labeled with a chunk-end label as teacher data.
Dependencies are inferred from the claim text.