Text Classifier Construction Using Adaptive Phrase Weighting
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing text classification methods are inefficient in accurately classifying documents into specific topics, particularly when relying on linear classifiers that require large amounts of human-labeled data and struggle with document length variations.
Innovation Solution
The method involves generating and filtering candidate phrases using n+k-grams, assigning weights based on initial classifier scores, and iteratively refining the classifier using supervised and unsupervised techniques to improve accuracy with minimal human-labeled data, accounting for document length and using search engine queries to validate phrase relevance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional linear classifiers are used for document classification, then classification can be performed with simple methodology, but large amounts of human-labeled data are required and accuracy is insufficient
Solution Approach 1:
The patent applies preliminary action by pre-processing documents to extract n-gram phrases and automatically generating initial classifier weights using unsupervised techniques before human labeling. This preliminary phase creates a foundation that reduces the amount of human-labeled data needed for achieving high classification accuracy.
Solution Approach 2:
The system implements self-service through automated phrase extraction from document corpora and self-training mechanisms where the classifier iteratively improves using unlabeled data. The system serves itself by automatically generating training features and refining weights without continuous human intervention, thereby reducing dependency on large amounts of manually labeled data.
2Adaptability or versatility
If conventional linear classifiers are used, then implementation is straightforward, but the classifiers struggle with document length variations
Solution Approach 1:
The patent applies local quality by normalizing phrase scores based on document length, giving different weights to phrases depending on the local context of document length. This allows the classifier to adapt to varying document lengths while maintaining the overall linear classifier structure, balancing adaptability with implementation simplicity.
Solution Approach 2:
The system implements dynamics by making the classifier adaptive to document length variations through iterative re-training and weight adjustment. The classifier dynamically adjusts its parameters based on the characteristics of the input documents, allowing it to handle varying lengths effectively while maintaining a relatively simple linear structure.
3Quantity of substance
If automated phrase extraction is used to reduce human-labeled data, then data requirements are reduced, but erroneously included candidate phrases are introduced
Solution Approach 1:
The patent applies feedback through iterative evaluation cycles where extracted phrase candidates are scored using the classifier, and poor-performing phrases are identified and removed. This feedback loop continuously refines the phrase candidate set, maintaining high reliability while reducing dependency on human-labeled data for phrase selection.
Solution Approach 2:
The system replaces the mechanical process of manual phrase verification with automated evaluation using the classifier itself. Instead of relying on human experts to filter candidate phrases, the system uses computational scoring and iterative refinement to automatically identify and remove erroneous phrases, thereby reducing human-labeled data requirements while maintaining phrase accuracy.
4Measurement precision
If iterative classifier refinement is performed to improve accuracy, then classification precision is enhanced, but computational time and complexity increase
Solution Approach 1:
The patent applies partial action by performing iterative refinement only to the extent necessary to achieve satisfactory accuracy thresholds. The system stops iteration when performance plateaus or when diminishing returns are detected, avoiding excessive computational time while still achieving improved classification accuracy through targeted iterative improvements.
Data Source
AI summary
Methods, systems, and apparatus, including computer program products, for constructing text classifiers. The method includes receiving a collection of candidate phrases for a given topic; filtering the received candidate phrases to remove erroneously included candidate phrases; assigning weights to the candidate phrases including scoring each candidate phrase using an initial classifier and assigning weights to the candidate phrases based on the scores; and generating a linear classifier using the filtered and weighted candidate phrases, where the linear classifier varies the weights for each phrase candidate depending on the length of the document being classified.