One-Hot Token Vectors for 3D Scene Graphs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current tokenization techniques, such as Byte Pair Encoding (BPE), are inefficient for tokenizing XML files of 3D virtual environments, resulting in unfeasibly large numbers of tokens and loss of modality-specific information.
Innovation Solution
A custom tokenization method that generates one-hot token vectors for tags and metadata vectors for metadata in a data file describing a 3D virtual environment, reducing the number of tokens needed to represent the scene while preserving modality-specific information.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If Byte Pair Encoding (BPE) is used to tokenize XML files of 3D virtual environments, then the tokenization process is simple and automated, but the number of tokens becomes unfeasibly large (approximately 28,000 tokens)
Solution Approach 1:
The patent segments the tokenization process into two distinct parts: (1) generating one-hot token vectors for XML tags using a predefined vocabulary, and (2) separately processing metadata values. This segmentation allows each part to be optimized independently, preventing the exponential token expansion that occurs with BPE on entire XML documents.
Solution Approach 2:
The patent changes the fundamental parameter of token representation from variable-length BPE tokens to fixed-length one-hot vectors. By mapping each XML tag to a predefined index in a vocabulary list and representing it as a one-hot vector, the system eliminates the token multiplication effect of BPE while maintaining automated processing capability.
2Ease of manufacture
If standard tokenization methods are used on XML files, then the process is straightforward, but modality-specific information is lost
Solution Approach 1:
The patent extracts metadata values from their XML context and processes them separately from the tag structure. By pulling out metadata (such as floating-point values, identifiers, and other attribute data) and handling them as distinct entities with appropriate vector representations, the system preserves modality-specific information that would otherwise be lost in generic tokenization.
Solution Approach 2:
The patent introduces an intermediary metadata vector representation layer between the XML structure and the final token sequence. This intermediary layer preserves the semantic meaning and modality-specific characteristics of metadata values while enabling their integration with tag tokens in a unified format suitable for machine learning models.
3Quantity of substance
If one-hot token vectors with metadata vectors are generated, then the number of tokens is reduced (to around 519 tokens), but the tokenization process becomes more complex
Solution Approach 1:
The patent performs preliminary actions by pre-defining a vocabulary of XML tags and their corresponding one-hot vector representations before processing the actual XML file. This preprocessing step creates lookup tables and metadata schemas that simplify the main tokenization process, reducing runtime complexity despite the initial setup effort.
4Extent of automation
If BPE tokenization is used, then automation is maintained, but the tokens are unfeasibly large for machine-learning models
Solution Approach 1:
The patent introduces dynamic adaptability by allowing the system to adjust metadata vector dimensions and configurations based on the specific requirements of different machine learning models. This dynamic approach enables optimization for model processing efficiency while maintaining automated tokenization, as the token structure can be tailored to match model input requirements without manual intervention.
Data Source
AI summary
A token engine may receive a data file that describes a three-dimensional (3D) virtual environment using tags for attributes in the 3D virtual environment. The token engine generates a set of one-hot token vectors from the tags in the data file and a set of metadata vectors from metadata in the data file, where one or more metadata vectors in the set of metadata vectors correspond to one or more one-hot token vectors in the set of one-hot token vectors. The token engine combines the set of one-hot token vectors and the set of metadata vectors. The token engine provides a combined set of one-hot token vectors and metadata vectors as input to a deep-learning model.