Text Position Encoding with Adaptive Binning for Document AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing layout language models, such as LayoutLM, are prone to overfitting on specific bounding box positions and are not customizable, leading to reduced generalizability and susceptibility to noise, which affects the accuracy of document image understanding tasks.
Innovation Solution
A method to encode the bounding box of a textual object with customizable relative location, allowing variation within a given range, using a parameter search or automatic determination based on word density, and applying bin sizes to discretize the relative position, enhancing the model's generalizability and noise tolerance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If 2D position embeddings are used to encode bounding box location information, then the model can capture spatial relationships, but the model becomes prone to overfitting on specific locations and loses generalizability
Solution Approach 1:
The patent transforms the continuous 2D position coordinates into discrete bin indices by changing the parameter representation from continuous floating-point values to discrete integer categories. This discretization process modifies the parameter space, allowing the model to learn positional relationships in a more generalized manner while reducing sensitivity to exact location variations.
Solution Approach 2:
The patent divides the continuous position space into discrete bins along both x and y axes, segmenting the spatial domain into manageable discrete units. This segmentation approach converts the infinite continuous position space into a finite discrete grid, enabling the model to handle positional information in a categorical manner that improves generalizability.
2Device complexity
If fixed 2D position embeddings are used, then the encoding process is simple, but the model becomes susceptible to noise and lacks tolerance to position variations
Solution Approach 1:
The patent performs preliminary discretization of position coordinates into bins before feeding them to the model. This preliminary action of categorizing positions into discrete bins beforehand allows the model to be more robust to noise and small variations, as the discretization process inherently smooths out minor positional fluctuations.
3Measurement precision
If continuous position coordinates are used, then the location information is precise, but the model overfits to specific locations and dilutes the textual signal
Solution Approach 1:
The patent changes the parameter type from continuous coordinates to discrete bin indices, which reduces the information overload from precise continuous positions. This parameter transformation maintains sufficient positional discrimination while preventing the model from overfitting to exact location details, thereby preserving the relative importance of textual signals.
Data Source
AI summary
A computer-implemented method for adaptively discretizing a position of a textual object in a document includes receiving, by a computer system, an image of the document and determining, by the computer system, an absolute position of the textual object in the image of the document. The method further includes normalizing, by the computer system, the absolute position to determine a relative position of the textual object. The method also includes calculating, by the computer system, a bin size such that at least one axis of the image is divided into a plurality of separate bins, wherein a distance between each bin along the at least one axis and its adjacent bin equals the bin size. The method includes discretizing, by the computer system, the relative position based on the bin size to determine a discretized position of the textual object; and providing, by the computer system, the discretized position and a textual content of the textual object as an input to a machine learning model.