Conformal Predictor for Zero-Shot Text Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Pre-trained language models for zero-shot text classification face inefficiencies due to high computational costs and increased carbon footprint, particularly when dealing with large transformer-based models and numerous target labels, as they require multiple passes over text and full self-attention computations for each label.
Innovation Solution
The Conformal Predictor framework reduces the number of target class labels by employing a fast base classifier to filter out unlikely labels within a pre-defined error rate, using a calibration dataset to generate a subset of classification labels that are then used with larger NLI/NSP-based models for prediction, thereby limiting the computational burden.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If zero-shot text classification models employ cross-encoder architecture and make forward passes for each label-text pair separately, then classification accuracy is maintained, but computational cost increases linearly with the number of labels
Solution Approach 1:
The patent segments the classification process into two distinct stages: (1) a fast base classifier performs initial filtering to identify a reduced subset of candidate labels, and (2) the computationally intensive zero-shot model is applied only to this reduced subset. This segmentation resolves the contradiction by separating the broad scanning function from the detailed analysis function, maintaining accuracy while reducing overall computational cost.
Solution Approach 2:
The patent applies preliminary action by using a fast base classifier to pre-filter and reduce the label space before the main zero-shot classification model processes the text. This preliminary filtering step identifies a small subset of candidate labels that are most likely to be correct, allowing the computationally expensive model to focus only on relevant candidates rather than evaluating all possible labels.
2Adaptability or versatility
If the number of target labels grows, then classification coverage is improved, but the computational cost for text classification grows accordingly
Solution Approach 1:
The patent segments the label set into two groups: a reduced subset of candidate labels identified by the fast base classifier, and the remaining labels that are filtered out. This segmentation allows the system to maintain adaptability to a large number of potential labels while improving productivity by limiting the expensive inference process to only the most promising candidates.
Solution Approach 2:
The patent applies partial action by performing complete zero-shot classification only on a partial subset of labels (the candidate labels identified by the base classifier) rather than on all possible labels. This partial processing maintains classification coverage for the full label space while significantly improving inference speed by avoiding unnecessary computations on unlikely candidates.
3Reliability
If larger transformer-based models are used for zero-shot classification, then model performance is improved, but carbon footprint increases
Solution Approach 1:
The patent segments the computational workload between a lightweight base classifier and a larger transformer-based zero-shot model. The base classifier handles the bulk of the filtering work with minimal environmental impact, while the larger model is invoked only for the reduced subset of candidate labels, thereby maintaining high model performance while significantly reducing the overall carbon footprint.
Solution Approach 2:
The patent introduces a fast base classifier as an intermediary between the input text and the larger transformer-based zero-shot model. This intermediary performs preliminary filtering to identify candidate labels, reducing the number of times the carbon-intensive large model must be executed, thus maintaining reliability while mitigating harmful environmental effects.
Data Source
AI summary
Embodiments described herein provide a Conformal Predictor (CP) that reduces the number of likely target class labels CP. Specifically, the CP provides a model agnostic framework to generate a label set, instead of a single label prediction, within a pre-defined error rate. The CP employs a fast base classifier which may be used to filter out unlikely labels from the target label set, and thus restrict the number of probable target class labels while ensuring the candidate class labels set meets the pre-defined error rate.