Paraphrasing Candidate Sentence Generation for Translation Corpus Expansion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine translation systems lack sufficient usable example sentences for performance improvement, necessitating a more efficient method to create translation corpora with increased pairs of sentences.
Innovation Solution
A method and device for generating paraphrasing candidate sentences by rephrasing fragments of an original sentence and identifying equivalent meaning sentences, then creating new pairs with translated sentences to enhance the translation corpus.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If more sentence pairs are accumulated in the translation corpus, then machine translation performance improves, but the cost and time for collecting sentence pairs increases
Solution Approach 1:
The original sentence is divided into multiple fragments, and each fragment is independently paraphrased to generate diverse candidate sentences. This segmentation allows systematic expansion of the corpus without manually collecting entire sentence pairs.
Solution Approach 2:
Instead of manually collecting new sentence pairs, the system creates copies of existing sentence pairs by generating paraphrased versions of the source sentence. These paraphrased sentences are then paired with the translated sentence to create new training examples, effectively multiplying the utility of each original pair.
2Reliability
If more sentence pairs are accumulated in the translation corpus, then machine translation performance improves, but the cost for collecting sentence pairs increases
Solution Approach 1:
The system generates multiple paraphrased versions of the source sentence fragments and combines them with the translated sentence to create new training pairs. This copying approach allows exponential expansion of corpus size from a single original pair, dramatically reducing the need for additional manual collection resources.
Solution Approach 2:
The system performs self-service by automatically generating paraphrased sentences using NLP techniques rather than relying on external manual collection. The paraphrasing process is automated through computational methods, eliminating the need for additional human annotators or data collectors.
3Productivity
If paraphrasing is applied to generate more candidate sentences, then the translation corpus size increases, but the complexity of identifying equivalent meaning sentences increases
Solution Approach 1:
The system employs a scoring mechanism that evaluates each paraphrased candidate sentence against the original sentence and its translation. This feedback loop automatically filters candidates based on meaning equivalence scores, managing the complexity of identification through systematic evaluation rather than manual review.
Data Source
AI summary
A translation corpus creation method of the present disclosure includes generating plural paraphrasing candidate sentences for a first original sentence in a first language by paraphrasing one or plural fragments among plural fragments included in the first original sentence into other expressions in the first language by a paraphrasing candidate sentence generation unit, identifying one or plural paraphrasing candidate sentences in the same meaning as the meaning of the first original sentence from the plural paraphrasing candidate sentences as one or plural paraphrasing sentences by a paraphrasing sentence identification unit, and generating a new set of sentences by setting the one or plural identified paraphrasing sentences and a second original sentence translated from the first original sentence as a set of sentences to create a translation corpus with the generated and new set of sentences by a translation corpus creation unit.