Teacher Model Pseudo-Labeling for Named Entity Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems exhibit poor recognition of named entities due to limited training data and varying pronunciations across accents and regional differences, leading to incorrect labeling and reduced accuracy.
Innovation Solution
A system and method that trains teacher models to transcribe unlabeled audio samples and predict pseudo-labeled datasets, with corrections using user-specific textual data, and retraining based on the most accurate teacher model predictions to improve named entity recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional ASR systems use limited training data for named entities, then training time and resources are reduced, but recognition accuracy deteriorates
Solution Approach 1:
The system enables models to self-improve by generating their own training data through pseudo-labeling. Teacher models transcribe unlabeled audio samples and create pseudo-labeled datasets that are then used to retrain student models, forming a self-service loop that continuously improves named entity recognition without requiring additional manual annotation resources
Solution Approach 2:
The system combines multiple data sources including unlabeled audio samples, user-specific textual data, and corrections from multiple teacher models to create a composite training dataset. This composite approach integrates diverse information sources to improve recognition accuracy beyond what any single data source could provide
2Measurement precision
If multiple teacher models are trained to improve prediction accuracy, then named entity recognition improves, but device complexity increases
Solution Approach 1:
The system segments the training process into distinct phases: initial teacher model training on available data, pseudo-label generation from unlabeled audio, correction using user-specific data, and retraining of student models. This segmentation allows complex multi-model training to be managed through modular, sequential steps rather than monolithic processing
Solution Approach 2:
Pseudo-labeled datasets serve as an intermediary between teacher models and student models. The teacher models generate pseudo-labels from unlabeled audio, these pseudo-labels are corrected using user-specific textual data, and the corrected pseudo-labeled datasets are then used to train student models. This intermediary mechanism facilitates knowledge transfer while managing complexity
3Measurement precision
If user-specific corrections are applied to transcribed NE labels, then recognition accuracy improves, but processing time increases
Solution Approach 1:
User-specific named entity textual data is collected and prepared in advance before the pseudo-labeling and correction process. This preliminary preparation of user data allows corrections to be applied efficiently during the training process without adding significant processing time during inference or operation
Solution Approach 2:
The system applies corrections selectively to named entity labels rather than processing entire transcriptions uniformly. By focusing correction efforts specifically on NE labels where accuracy is most critical, the system achieves improved transcription accuracy for named entities without proportionally increasing overall processing time for complete audio transcriptions
Data Source
AI summary
A method includes training a set of teacher models. Training the set of teacher models includes, for each individual teacher model of the set of teacher models, training the individual teacher model to transcribe unlabeled audio samples and predict a pseudo labeled dataset having multiple labels. At least some of the unlabeled audio samples contain named entity (NE) audio data. At least some of the labels include transcribed NE labels corresponding to the NE audio data. The method also includes correcting at least some of the transcribed NE labels using user-specific NE textual data. The method further includes retraining the set of teacher models based on the pseudo labeled dataset from a selected one of the teacher models, where the selected one of the teacher models predicts the pseudo labeled dataset more accurately than other teacher models of the set of teacher models.