Word Grouping Using Multiple Models and User Annotations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional word grouping methods using n-gram and Word2Vec struggle to accurately judge similarity between words, as they do not effectively differentiate between preferred and non-preferred words, leading to inefficient dictionary creation.
Innovation Solution
A computer-implemented method that calculates similarity scores for each word relative to a seed word using multiple models, updates model weights based on user annotations of preferred and non-preferred words, and groups words accordingly to enhance similarity judgment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional methods (n-gram and Word2Vec) are used to group words, then the process is simple, but the similarity judgment accuracy is insufficient
Solution Approach 1:
The patent combines multiple word embedding models (n-gram, Word2Vec, and other models) into a unified framework where their similarity scores are aggregated through weighted sums. This merging approach leverages the strengths of different models to improve overall similarity judgment accuracy while managing complexity through systematic integration.
Solution Approach 2:
The patent creates a universal word grouping system that can work with multiple different embedding models simultaneously. The framework is designed to accommodate various models and adaptively combine their outputs, making the system multi-functional and robust across different linguistic contexts and word types.
2Measurement precision
If multiple models are used to calculate similarity scores, then the similarity judgment accuracy improves, but the computational complexity increases
Solution Approach 1:
The patent performs preliminary actions by pre-calculating and storing word embedding vectors for all words in the vocabulary. These pre-computed vectors are then reused during similarity calculations, avoiding redundant computations. The system also pre-identifies candidate words based on initial filtering criteria before applying the full multi-model similarity assessment.
Solution Approach 2:
The patent implements a two-stage approach where not all words undergo full multi-model similarity calculation. Instead, a candidate word selection stage filters the vocabulary to identify only those words worth detailed analysis. This partial action approach applies the computationally intensive multi-model comparison only to a subset of promising candidates, reducing overall computational power requirements.
3Measurement precision
If user annotations are incorporated to update model weights, then the word grouping accuracy improves, but the processing time increases
Solution Approach 1:
The patent implements a feedback mechanism where user annotations about preferred and non-preferred word groupings are collected and used to update the weights of different embedding models. This feedback loop allows the system to learn from user preferences and continuously improve its word grouping accuracy. The weight updates are performed incrementally, allowing the system to adapt to user needs over time.
Solution Approach 2:
The patent performs preliminary actions by pre-processing user annotations and organizing them into training data structures before weight updates. The system aggregates annotation information across multiple user interactions, preparing consolidated weight update signals in advance. This preliminary organization of feedback data reduces the computational burden during actual weight update operations.
4Productivity
If candidate words are filtered and selected, then the dictionary creation efficiency improves, but the risk of excluding relevant words increases
Solution Approach 1:
The patent applies partial action by implementing a multi-stage filtering process where different criteria are applied at different stages. The first stage uses loose criteria to identify candidate words, ensuring minimal exclusion of potentially relevant terms. Subsequent stages apply progressively stricter criteria to refine the selection. This staged approach balances efficiency gains from filtering with reliability preservation through multiple checkpoints.
Solution Approach 2:
The patent incorporates feedback mechanisms in the candidate word selection process by using user annotations to validate and adjust selection criteria. When users indicate that a filtered-out word should have been included, or when they confirm the exclusion of certain words, this feedback is used to refine the filtering thresholds and criteria. This iterative refinement improves both the efficiency and reliability of word selection over time.
Data Source
AI summary
A computer-implemented method is provided. The method includes acquiring a seed word; calculating a similarity score of each of a plurality of words relative to the seed word for each of a plurality of models to calculate a weighted sum of similarity scores for each of the plurality of words; outputting a plurality of candidate words among the plurality of words; acquiring annotations indicating at least one of preferred words and non-preferred words among the plurality of the candidate words; updating weights of the plurality of models in a manner to cause weighted sums of similarity scores for the preferred words to be relatively larger than the weighted sums of the similarity scores for the non-preferred words, based on the annotations; and grouping the plurality of candidate words output based on the weighted sum of similarity scores calculated with updated weights of the plurality of models.