Ordinal Vector Encoding for Temporal Sequence and Cyclic Pattern Representation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional vector space modeling lacks structural information, particularly in temporal contexts, failing to encode the sequential and cyclic nature of ordinal entities like months and hours, which limits natural interpretation and processing of temporal information in Natural Language Processing (NLP) scenarios.
Innovation Solution
Generating vectors that represent ordinal entities in a manner akin to their representation on a number line, incorporating structural information such as sequence and cyclic patterns, using a method where child vectors are derived from parent vectors with controlled correlation and dimensionality to encode temporal relations, allowing neural networks to learn temporal patterns.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If traditional vector space modeling is used to represent entities, then semantic similarity can be measured, but structural information and sequential relationships are lost
Solution Approach 1:
The vector representation is segmented into multiple components: a base vector capturing semantic meaning and additional ordinal components encoding positional information. This segmentation allows the vector to simultaneously represent both semantic similarity and structural/sequential relationships without losing either type of information.
Solution Approach 2:
The invention adds an ordinal dimension to the traditional semantic vector space. By incorporating positional encoding along an additional dimension, the method preserves structural information while maintaining the original semantic representation, effectively moving from a flat vector space to a multi-dimensional representation that captures both aspects.
2Loss of information
If vectors are generated to capture semantic similarity, then entities with similar meanings are close, but temporal or sequential order cannot be inferred
Solution Approach 1:
The invention introduces an intermediary ordinal encoding mechanism that mediates between semantic similarity and sequential order. This intermediary layer transforms positional information into vector components that can be combined with semantic vectors, allowing both semantic distance and temporal/sequential distance to be measured simultaneously with proper interpretation.
3Ease of operation
If traditional word vectors are used, then computational efficiency is maintained, but natural interpretation of temporal relationships is not achieved
Solution Approach 1:
The invention changes the parameters of vector representation by incorporating ordinal position as an additional parameter alongside semantic meaning. This parameter change enables natural interpretation of temporal relationships while maintaining computational efficiency, as the extended vector structure allows direct mathematical operations to capture temporal patterns without requiring complex post-processing.
Data Source
AI summary
A method and system to generate vectors that represent linearly progressing entities like time are disclosed. Traditional methods of vectorisation account for semantic or associative similarity of the entities. Thus, vectors conveying semantic information do not convey structural relations between such entities. The method allows for the representation of such structural information, for example the months in a year. The vectors generated by the invention encode this relation between the months such that one can interpret the sequence of the months, the difference between then and their cyclic nature. The method works in a manner similar to a genetic code, where subsequent “child” vectors are generated by related “parents”, thus encoding the similarity and the distance of the sequential entities. An object of the inventions to allow algorithms in machine learning to easily learn over temporal entities its natural text.