Context Decoding End Token Control via Model Replacement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for obtaining sentences from context information can generate excessively long sentences due to the absence of an end token, leading to inefficiencies in decoding processes.
Innovation Solution
An electronic device and method that replace decoding data with alternative data when a certain number of words are output without an end token, increasing the probability of obtaining the end token, thereby preventing the generation of infinitely long sentences. This involves using a processor to determine when to replace language models or decoding algorithms based on reference values, ensuring the end token is output efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If decoding is repeatedly performed in sequence to obtain sentences from context information, then the sentence generation capability is improved, but excessively long sentences are generated when the end token is not output
Solution Approach 1:
The patent changes the parameters of the decoding process by replacing the language model or decoding algorithm when the output sentence exceeds a reference length. This parameter change allows the system to switch to alternative models that are more likely to generate sentences with proper end tokens, thereby resolving the contradiction between maintaining sentence generation capability and controlling sentence length.
Solution Approach 2:
The patent implements a feedback mechanism where the system monitors the length of generated sentences and the presence of end tokens. When the sentence length exceeds a reference value or the end token is not output, the system feeds this information back by replacing the current language model or decoding algorithm with alternative data,从而防止生成过长的句子。
2Reliability
If the end token is not output due to errors or repeated word output, then decoding continues, but the sentence becomes considerably long
Solution Approach 1:
The patent applies preliminary anti-action by proactively replacing the language model or decoding algorithm before the sentence becomes excessively long. When the sentence length reaches a reference threshold, the system preemptively switches to alternative models that are more likely to output the end token, thereby preventing the generation of harmful overly long sentences while maintaining decoding continuity.
Solution Approach 2:
The patent prepares alternative language models and decoding algorithms in advance as a cushioning mechanism. When decoding errors occur or the end token is not output, these pre-prepared alternative models are immediately deployed to correct the decoding process and prevent excessively long sentences, thus cushioning against potential harmful outcomes.
3Reliability
If alternative data is replaced to increase the probability of obtaining the end token, then the end token output probability is improved, but the system complexity increases
Solution Approach 1:
The patent implements a self-service mechanism where the system automatically monitors sentence length and end token presence, and autonomously decides when to replace the language model or decoding algorithm. This self-service approach increases end token output probability while managing system complexity through automated decision-making based on simple length thresholds and model comparison.
Data Source
AI summary
A method of obtaining, by an electronic device, a sentence corresponding to context information, including obtaining first output information including at least one word output by decoding the context information based on at least one data; based on detecting that a first token is not included in the first output information, determining whether a number of words included in the first output information is greater than or equal to a reference value; based on a result of the determining, replacing the at least one data with other data; and obtaining the sentence corresponding to the context information based on at least one output information obtained by decoding the context information based on the other data.