Global Lexical Selection for Machine Translation Reordering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine translation methods rely heavily on local associations between source and target words, which can ignore other beneficial sources of information and struggle with lexical reordering, especially between languages with significantly different word orders.
Innovation Solution
A novel approach that associates target words with the entire source sentence rather than individual source words, using a global lexical selection method that generates a target 'bag of words' and reconstructs the sentence considering all permutations with a conditional probability threshold, without relying on local word alignments.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If local word-alignment algorithms are used to provide local associations between source and target words, then lexical selection precision is improved, but the ability to handle lexical reordering between languages with different word orders deteriorates
Solution Approach 1:
The patent segments the translation process into two independent components: (1) global lexical selection that generates a target bag of words from the entire source sentence, and (2) separate lexical reordering that handles word order adjustments. This segmentation allows each component to specialize - lexical selection focuses on accuracy while reordering handles structural differences between languages.
Solution Approach 2:
The patent introduces a target language bag of words as an intermediary representation between source sentence and target sentence. This bag of words serves as a mediator that decouples the lexical selection process from the reordering process, allowing words to be selected based on global context and then freely rearranged to match target language word order patterns.
2Adaptability or versatility
If phrase-level alignments are used to increase the extent of local associations, then some lexical reordering is captured, but the complexity of the alignment model increases
Solution Approach 1:
The patent extracts the reordering function from the alignment model entirely. Instead of using complex phrase-level alignments to capture reordering patterns, the system uses simple word-level alignments only for lexical selection, and handles all reordering separately through the target language model and bag-of-words approach, thereby simplifying the alignment model.
Solution Approach 2:
Instead of using complex alignments to directly produce reordered target phrases, the patent inverts the approach: it uses simple alignments to select target words, then generates all possible permutations of these words, and finally uses a target language model to score and select the most natural ordering. This inversion simplifies the alignment component while achieving the same goal.
3Adaptability or versatility
If a global lexical selection approach is used to associate target words with the entire source sentence, then lexical reordering capability is improved, but the precision of local word associations deteriorates
Solution Approach 1:
The patent merges multiple source words that align to the same target word into a single global selection decision. Instead of making independent local decisions for each source-target word pair, the system combines all alignment evidence for each target word across the entire source sentence, then makes a unified selection decision based on global context, thereby improving reordering capability while maintaining precision through aggregated evidence.
Data Source
AI summary
Disclosed are systems, methods, and computer-readable media for performing translations from a source language to a target language. The method comprises receiving a source phrase, generating a target bag of words based on a global lexical selection of words that loosely couples the source words/phrases and target words/phrases, and reconstructing a target phrase or sentence by considering all permutations of words with a conditional probability greater than a threshold.