Chart Caption Paraphrasing Using Placeholder Entity Names
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data captioning systems struggle with maintaining accuracy and computational efficiency when generating naturally phrased captions for complex or sophisticated sentence structures, particularly due to the use of multi-word entity names and complex sentence structures, and existing systems fail to accurately paraphrase data in graphs or charts.
Innovation Solution
The insight augmentation system replaces complex entity names with placeholder names, utilizing a renaming map, and trains a distilled insight model to generate augmented insights, which are then transformed back to include the original entity names, thereby improving accuracy and reducing computational requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If template-based captions with multi-word entity names are used for data captioning, then the captions can represent complex data structures, but the accuracy of paraphrasing deteriorates due to hallucination rates
Solution Approach 1:
The patent introduces placeholder names as intermediary elements that replace complex multi-word entity names during the paraphrasing process. These placeholders serve as mediators between the original complex entities and the language model, reducing hallucination rates while preserving the ability to represent complex data structures. The placeholders are later mapped back to original entity names to maintain accuracy.
2Manufacturing precision
If large natural language models are used to generate naturally phrased insights, then the quality of paraphrasing improves, but computational resource consumption increases
Solution Approach 1:
The patent segments the caption processing into distinct stages: (1) extracting entity names from template-based captions, (2) replacing them with placeholder names, (3) generating paraphrases using the language model, and (4) mapping placeholders back to original entities. This segmentation allows for optimized resource usage at each stage while maintaining overall paraphrasing quality.
Solution Approach 2:
The patent performs preliminary actions by pre-processing the template-based captions to extract and replace entity names with placeholders before feeding them to the language model. This pre-processing step simplifies the input for the language model, reducing computational complexity while preserving the essential information needed for accurate paraphrasing.
3Loss of information
If complex entity names are retained in captions, then the captions maintain contextual information, but the model prediction accuracy deteriorates for complicated sentence structures
Solution Approach 1:
Placeholder names act as intermediaries that preserve the structural and positional information of complex entity names without introducing the computational complexity that reduces model prediction accuracy. The placeholders maintain the necessary contextual relationships while simplifying the input for the language model, enabling more accurate predictions for complicated sentence structures.
Data Source
AI summary
The present disclosure relates to systems, methods, and non-transitory computer readable media for generating augmented insights which paraphrase complicated captions for data charts or data graphs in natural language utilizing a natural language model. In some embodiments, the insight augmentation system generates a modified caption by replacing an entity name within the template-based caption with a placeholder name utilizing a renaming map. Based on the modified caption and utilizing a large language model, in some cases, the insight augmentation system generates a placeholder insight describing the data chart in natural language using the placeholder name. Furthermore, in some embodiments, the insight augmentation system generates an augmented insight describing the data chart in natural language by replacing the placeholder name in the placeholder insight with the entity name.