Verbal Product Entity Extraction via Neural Tokenization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition systems are limited in their ability to accurately identify multiple product names from a single verbal utterance without delimiters, due to ambiguity and complexity in voice utterances, which prevents users from efficiently ordering multiple items at once.
Innovation Solution
The implementation of machine learning-based systems using Entity Transformer neural network architectures and sequence-to-sequence modeling, along with featurizer and classifier models, to recognize and extract multiple product entities from utterances, even without explicit delimiters, by training models on user-specific data to enhance accuracy and reduce processing requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional speech recognition systems are used to identify product names from verbal utterances, then the system implementation is simple, but the accuracy of identifying multiple product names is limited
Solution Approach 1:
The system segments the verbal utterance into multiple candidate product name entities using delimiters (commas, semicolons, periods) and processes each segment separately through entity recognition models, enabling accurate identification of multiple products while managing system complexity through modular processing
Solution Approach 2:
The system introduces an intermediary processing layer that includes featurizer models, entity transformer models, and classifier/extractor models between the raw utterance and final product identification, allowing complex multi-product recognition through staged processing while maintaining manageable system architecture
2Measurement precision
If machine learning models are trained on user-specific data to enhance accuracy, then the identification accuracy improves, but the computation and memory requirements increase
Solution Approach 1:
The system applies user-specific training data locally to customize entity recognition for individual users while using pre-trained general models for common product names, achieving high accuracy for user-specific products without requiring complete retraining of all models, thus balancing accuracy with computational efficiency
Solution Approach 2:
The system performs preliminary processing including tokenization, featurization, and candidate generation before final entity classification, allowing user-specific data to be applied selectively at critical stages rather than throughout the entire processing pipeline, reducing overall computational requirements while maintaining accuracy
Data Source
AI summary
Some embodiments provide retail product ordering systems comprising: a user computing device comprising an application executed by a device control circuit to: receive an audible utterance; controls a product identifier application interface to: apply a tokenizer model and obtain a set of individual search words; apply a series of featurizer models to the search words to generate features; and apply a classifier and extractor model based on the features and generate multiple requested product entities each comprising a respective sub-set of the position labeled product terms; wherein the device control circuit is further configured to access a purchase history database, confirm an accuracy of each of requested product entities relative to a purchase history, generate a listing of determined product identifiers corresponding to the confirmed set of the multiple requested product entities, and control a display system of the user computing device to render the listing of determined product identifiers.