Relative Feature Sequences for Structure-Preserving Content Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods struggle to generate new content that maintains the overall structure of existing content, such as songs, due to difficulties in learning and replicating the natural structure and flow of the entire composition.
Innovation Solution
An information processing apparatus that utilizes a trained encoder to extract feature values from partial data, calculates relative feature values between data sets to determine the structure, and generates new content based on these values and arbitrary data to maintain the structure of the original content.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If conventional machine learning methods are used to generate content based on lyrics data, then new content can be automatically generated, but the overall structure and natural flow of the generated content does not match human-composed songs
Solution Approach 1:
The song data is divided into multiple pieces of data representing different segments or sections of the song. The learning unit processes each segment individually to extract structural features, which are then integrated to reconstruct the overall song structure. This segmentation allows the system to capture local patterns while maintaining global structural integrity.
Solution Approach 2:
The patent introduces a hierarchical structure with multiple levels of feature extraction. Instead of processing song data in a single dimension, the system creates a multi-dimensional representation where feature values at different hierarchical levels capture both local and global structural characteristics. This dimensional expansion enables the model to preserve overall song structure while generating new content.
2Measurement precision
If the system learns from individual song features, then specific content characteristics can be captured, but the overall structure and flow of the entire song cannot be maintained
Solution Approach 1:
The learning unit combines feature values from multiple data segments to create a comprehensive representation of the overall song structure. By merging information from individual segments through the hierarchical feature extraction process, the system maintains both detailed local features and global structural consistency in the generated content.
3Productivity
If simple generation methods are used following lyrics data, then generation speed is fast, but the generated melody or code does not maintain natural song structure
Solution Approach 1:
The system performs preliminary learning of song structural features from training data before actual content generation. The learning unit pre-processes and extracts structural patterns from multiple song segments, storing this knowledge in the hierarchical feature representation. During generation, this pre-learned structural knowledge is applied to ensure natural song structure is maintained while generating new content at high speed.
Data Source
Figure 1
Figure 2
Figure 3~4
AI summary
An information processing apparatus (100) according to the present disclosure includes an acquiring unit (132) that uses a trained encoder to acquire a feature value of each set of partial data forming first content; a calculating unit (133) that calculates a relative feature value that is a relative feature value between sets of partial data from the acquired feature values of the sets of partial data to calculate a relative feature value sequence indicating a feature of a structure of the first content; and a generating unit (134) that generates second content based on the relative feature value sequence of the first content and a feature value of optional data.