Neural Network Machine Translation Using Prefix Tree Candidate Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Neural network machine translation methods face significant computational challenges during the decoding process, particularly when determining conditional probabilities for a large number of words, leading to reduced translation speed.
Innovation Solution
The method employs a pre-built prefix tree based on a target sentence database to determine candidate objects, using a Gated Recurrent Unit (GRU) model for encoding and the self-normalization algorithm for calculating conditional probabilities, along with parallel matrix operations and hardware acceleration using GPUs or FPGAs to enhance translation speed.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conditional probability is calculated for all words in the word library during decoding, then translation accuracy is improved, but translation speed deteriorates due to huge computing quantity
Solution Approach 1:
The patent pre-builds a prefix tree structure from the target sentence database before translation. During decoding, candidate objects are quickly retrieved from the pre-built prefix tree using the current translation prefix as a query key, avoiding the need to calculate conditional probabilities for all words in the word library. This preliminary preparation significantly reduces computational complexity while maintaining translation accuracy by focusing calculations only on relevant candidate words.
Solution Approach 2:
The patent extracts and utilizes only the necessary subset of words from the complete word library by building a prefix tree from the target sentence database. During translation, only words that match the current prefix (candidate objects) are considered for conditional probability calculation, rather than processing all words in the word library. This extraction approach maintains translation quality while dramatically reducing the computational burden.
2Adaptability or versatility
If a complete word library is processed during decoding, then comprehensive translation coverage is achieved, but computational complexity increases
Solution Approach 1:
The patent segments the large word library into a hierarchical prefix tree structure organized by target sentences. During decoding, the search space is further segmented by using the current translation prefix to quickly identify and retrieve only the relevant candidate objects from the prefix tree. This segmentation approach maintains comprehensive translation coverage while reducing computational complexity by eliminating the need to process the entire word library for each decoding step.
Data Source
AI summary
A neural network machine translation method comprises: obtaining a to-be-translated source sentence; converting the source sentence into a vector sequence; determining candidate objects corresponding to the vector sequence according to a prefix tree which is pre-obtained and built based on a target sentence database, and determining a target sentence as a translation result according to the candidate objects.