Graph-Based NLP Embeddings for Structured Document Analysis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing NLP systems struggle to effectively process documents with complex structures and formats, such as financial documents with data tables, as they fail to capture spatial information and semantic relationships between text spans, leading to loss of important logical information.
Innovation Solution
A graph-based NLP system that generates vector embeddings by representing documents as geometric graphs, where nodes denote text spans and edges represent spatial relationships, allowing for the joint encoding of textual and spatial semantics, thereby capturing the spatial and semantic relationships throughout the document.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional sequential text input methods are used for language models, then processing speed and simplicity are improved, but the ability to capture spatial information and semantic relationships in structured documents deteriorates
Solution Approach 1:
The patent transforms the traditional one-dimensional sequential text input into a two-dimensional spatial graph representation. Nodes represent text spans and edges represent spatial relationships, preserving the geometric layout information that is lost in sequential processing. This dimensional transformation allows the language model to access both sequential and spatial information simultaneously.
Solution Approach 2:
The patent introduces a graph-based intermediary representation layer between the structured document and the language model. This graph serves as a mediator that encodes spatial relationships and semantic connections, translating complex document structures into a format that preserves information while being processable by sequential language models.
2Adaptability or versatility
If documents are converted to sequential text input, then compatibility with existing language models is improved, but the accuracy of data extraction from structured elements deteriorates
Solution Approach 1:
The patent segments the document into discrete text spans represented as nodes in a graph, where each node can be independently processed and then recombined. This segmentation allows the system to maintain compatibility with token-based language models while preserving the structural relationships through graph edges, thereby improving data extraction accuracy from tables and other structured elements.
Solution Approach 2:
The patent creates a composite representation combining sequential text tokens with spatial graph structures. This composite approach integrates the strengths of both sequential processing (compatibility with language models) and spatial reasoning (accuracy in structured document understanding), enabling accurate data extraction while maintaining model compatibility.
3Device complexity
If spatial layout information is ignored in favor of sequential text processing, then system complexity is reduced, but the ability to understand document structure and semantics deteriorates
Solution Approach 1:
The patent introduces a graph-based intermediary that captures spatial relationships without requiring complex modifications to the language model architecture. The graph structure serves as a lightweight mediator that encodes document hierarchy, proximity, and layout information, preserving structural semantics while maintaining relative system simplicity.
Solution Approach 2:
The patent adds a spatial dimension to the traditional sequential text representation by constructing a graph where nodes are text spans and edges encode spatial relationships. This dimensional enhancement preserves document structure and semantics without requiring fundamentally complex system changes, as the graph can be constructed and processed using standard NLP techniques.
Data Source
AI summary
Embodiments described herein provide a mechanism that encodes a text document into a geometric graph, where the nodes of the graph represent bits of text from the document and the edges of the graph represent relationships among the bits of text as laid out on a page of the document. Each node of the graph is encoded into a vector representation that contains information of the node and the local sub-graph (including the node and all edges branching out from the node). In this way, the vector representations of the document contain information of the inner-relationship between words, sentences and paragraphs of the document, instead of just mapping the text in the document as a string of input tokens to a vector representation.