Lexical Constraint Generation via Input Sequence Embedding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional lexically constrained machine translation technologies, such as grid beam search, face challenges with increased search time and decreased translation accuracy when dealing with long constraint vocabularies.
Innovation Solution
The proposed technology includes a generation apparatus that generates an output sequence by inputting an extended input sequence, which combines the source language sentence and constraint vocabulary, into a sequence conversion model. This apparatus performs a constrained search using grid beam search to ensure the output sequence includes all constraint information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If grid beam search is used to ensure lexical constraints are satisfied, then translation accuracy is improved, but search time increases significantly
Solution Approach 1:
The constraint vocabulary is integrated into the input sequence before feeding it to the sequence conversion model. By preparing the input sequence in advance with constraint information embedded, the model can process constraints during normal translation rather than requiring separate post-processing search steps, thereby reducing overall search time while maintaining accuracy
Solution Approach 2:
The patent combines the input sequence and constraint vocabulary into a single extended input sequence that is fed to the sequence conversion model. This merging of constraints into the input allows the model to handle both translation and constraint satisfaction simultaneously in one pass, eliminating the need for separate constraint-checking search phases
2Adaptability or versatility
If constraint vocabulary is extended to cover more terms, then lexical constraint coverage is improved, but search time increases
Solution Approach 1:
The sequence conversion model is designed to handle both general translation tasks and constraint satisfaction tasks simultaneously through a single unified architecture. The model processes the extended input sequence (containing both source text and constraint vocabulary) and generates translations that satisfy constraints as a natural part of its operation, rather than requiring separate specialized processing for constraints
Data Source
AI summary
A generation apparatus is provided for generating an output sequence from an input sequence, wherein the input sequence is a sequence of information, and the output sequence is a sequence of another piece of information. The generation apparatus includes a processor; and a memory storing instructions that cause the processor to execute a process. The process includes generating an input sequence with constraint information based on the input sequence and constraint information; generating output information by inputting the input sequence with constraint information to a sequence conversion model; and generating the output sequence by performing a constrained search using the output information such that the output sequence includes the constraint information.