Multimodal Deep Co-Learning for Tabular Data Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for ranking tabular data entities do not effectively consider the unique properties and structures of different aspects of tabular data, such as context inference along table axes, which can differ from word context, and often underestimate the significance of titles and structural elements.
Innovation Solution
A system using gated multimodal units to generate joint representations of tabular data aspects, combining separate representations of table properties like descriptions, schema, records, and facets, and applying a joint loss function for training, which supports both query-dependent and query-independent representations for improved ranking accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing neural network based representations (ELMo, BERT, XLNet) are used for natural language phrases, then word sense based on context is improved, but tabular data structure elements are not effectively considered
Solution Approach 1:
The patent segments tabular data into multiple aspects (descriptions, schema, records, facets) and processes each aspect separately through dedicated encoders, then combines them using gated multimodal units. This segmentation allows each aspect to be represented optimally while maintaining the ability to consider all aspects together for ranking.
Solution Approach 2:
The patent creates a universal ranking system that can handle multiple types of tabular data aspects through a common architecture. The gated multimodal unit serves multiple functions by selectively combining different aspect representations, making the system adaptable to various tabular structures while maintaining a unified approach.
2Adaptability or versatility
If Table2Vec embedding method is used, then tabular data structure elements are considered, but they are handled similarly without distinguishing unique properties
Solution Approach 1:
The patent applies local quality by assigning different encoders to different tabular aspects (description encoder, schema encoder, records encoder, facets encoder). Each encoder is optimized for its specific aspect, allowing nuanced representation of unique properties like context inference along table axes that differ from word context.
Solution Approach 2:
The patent creates a composite representation system that combines multiple aspect representations through gated multimodal units. This composite approach integrates the strengths of different aspect encoders while maintaining their individual characteristics, resulting in a richer representation that captures both shared and unique properties of tabular data.
3Measurement precision
If multiple separate representations of tabular aspects are generated and combined, then ranking accuracy is improved, but system complexity increases
Solution Approach 1:
The patent merges multiple separate aspect representations through gated multimodal units that combine encodings from different aspects. This merging process integrates the information from descriptions, schema, records, and facets into unified representations that can be used for ranking, reducing the complexity of handling multiple representations separately.
Solution Approach 2:
The gated multimodal units provide dynamic control over how different aspect representations are combined. The gates allow the system to adaptively weight and select relevant aspects based on the query and data, making the combination process flexible and efficient rather than a fixed complex operation.
Data Source
AI summary
A system and a computer-implemented method for ranking tabular data entities by likelihood of comprising answers for (natural language) queries, based on multimodal descriptions of the tabular data entities, comprising separate representations, which represent different aspects of the tabular data entities. The ranking is based on joint representations, generated from the query representation and separate representations of the tabular data entities' aspects, using gated multimodal units. The computer-implemented method may be used for applications such as web searches, data aggregation, and research tasks.