Learning Apparatus for NLP via Document Image Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Natural language processing using machine learning faces challenges in data augmentation for text data, as simple image augmentation techniques like rotation or inversion are not applicable, making it difficult to prepare large amounts of training data for discrete character or word values.
Innovation Solution
A learning apparatus converts documents into images and their associated tags into tag images, using these as input and ground truth data to train a network model, allowing for semantic segmentation and data augmentation by varying document composition, such as font, spacing, and arrangement, to generate diverse training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation is applied to text data using simple image techniques like rotation or inversion, then training data quantity increases, but the text data structure and meaning are distorted
Solution Approach 1:
The patent creates synthetic document images by copying and compositing text images with different layouts, fonts, and arrangements. This allows generation of diverse training data without distorting the original text structure, as the synthetic images are constructed from valid text representations rather than randomly transforming existing images.
Solution Approach 2:
The patent varies parameters such as font type, character spacing, line spacing, and document layout to generate diverse training data. These parameter changes create variations in the visual representation of text while preserving the underlying text content and structure, allowing data augmentation without distortion.
2Adaptability or versatility
If text data is processed as discrete character or word values, then natural language processing can be performed, but data augmentation becomes difficult
Solution Approach 1:
The patent introduces an intermediate representation layer where text is converted to visual document images before processing. This intermediary form allows the application of image processing techniques for data augmentation while maintaining the discrete text nature needed for NLP, thus bridging the gap between text processing and data augmentation capabilities.
Solution Approach 2:
The patent transforms text data from a one-dimensional sequence of characters/words into a two-dimensional visual representation as document images. This dimensional change enables the use of image augmentation techniques while preserving the text information, effectively solving the data augmentation problem for text data.
3Measurement precision
If a large amount of training data is prepared, then inference accuracy improves, but data preparation time and complexity increase
Solution Approach 1:
The patent implements a system where the data augmentation process is automated and self-service oriented. The system automatically generates synthetic training data by compositing text images with different parameters, eliminating the need for manual data preparation and reducing time investment while maintaining high inference accuracy through comprehensive training data.
Data Source
AI summary
According to one embodiment, a learning apparatus includes a processor. The processor acquires a document to which a tag is added. The processor converts the document into an image to generate a document image, and converts the tag into an image according to composition of the document image to generate a tag image. The processor trains a network model using the document image as input data and the tag image as ground truth data to generate a trained model.