Neural Network Scoring for Character-Limited Summaries
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing automatic summarization models using recurrent neural networks (RNNs) fail to effectively suppress the generation of words exceeding the character limit in summaries, leading to summaries that often exceed the intended length.
Innovation Solution
A learning method that combines a first score for each word based on context with a second score calculated from the word length and remaining characters, to generate a distribution of generation probabilities, ensuring that the model generates words closer to the character limit without exceeding it.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the model uses only context-based scoring (first score), then word generation accuracy is improved, but the character limit is exceeded
Solution Approach 1:
The patent changes the scoring parameters by introducing a second score that incorporates word length and remaining character count. This transforms the single-dimensional context-based scoring into a multi-dimensional scoring system that simultaneously considers semantic relevance and length constraints, resolving the contradiction between generation accuracy and character limit adherence.
Solution Approach 2:
The patent creates a composite scoring mechanism by combining the first score (context-based) and second score (length-based) into a combined score. This composite approach integrates multiple factors—semantic meaning, word length, and remaining characters—into a unified decision-making process that balances accuracy and length constraints.
2Length of moving object
If the model prioritizes word length constraints, then character limit adherence is improved, but word generation accuracy deteriorates
Solution Approach 1:
The patent modifies the scoring parameters by assigning appropriate weights to both the first score (context-based) and second score (length-based). This parameter adjustment ensures that neither accuracy nor length constraint is overly prioritized, instead achieving a balanced optimization where both factors contribute meaningfully to the final word selection.
Solution Approach 2:
The patent develops a composite scoring system that integrates both accuracy-related and length-related factors. By combining the first score and second score into a unified combined score, the system avoids the pitfall of prioritizing one factor over the other, instead achieving a harmonious balance between word generation accuracy and character limit adherence.
3Reliability
If the model generates words with similar meanings, then summarization relevance is improved, but the character limit is exceeded
Solution Approach 1:
The patent introduces a new parameter dimension by incorporating word length and remaining character count into the scoring process. This parameter change enables the model to distinguish between words with similar meanings that have different lengths, allowing selective generation of shorter words when the character limit is approaching, thus maintaining relevance while adhering to length constraints.
Solution Approach 2:
The patent creates a composite scoring mechanism that combines semantic relevance (first score) with length considerations (second score). This composite approach allows the model to evaluate multiple words with similar meanings and select the most appropriate one based on both relevance and length, preventing character limit exceedance while maintaining summarization quality.
Data Source
AI summary
A computer-implemented learning method includes combining a first score for a word included in a dictionary of a model and a second score for the word calculated based on a length of the word and a number of remaining characters up to an upper limit number of characters of a summary, the first score being output in response to input of input text to the model, calculating, based on a combined score generated by the combining, a distribution of generation probabilities of the word, and updating, based on the calculated distribution of the generation probabilities, parameters in the model.