US Patent Application 18315931. SYSTEM AND METHOD FOR CONTEXT INSERTION FOR CONTRASTIVE SIAMESE NETWORK TRAINING simplified abstract

SYSTEM AND METHOD FOR CONTEXT INSERTION FOR CONTRASTIVE SIAMESE NETWORK TRAINING

Organization Name

SAMSUNG ELECTRONICS CO., LTD.==Inventor(s)==

[[Category:Brendon Christopher Beachy Eby of Chicago IL (US)]]

[[Category:Suhel Jaber of San Jose CA (US)]]

[[Category:Sai Ajay Modukuri of San Francisco CA (US)]]

[[Category:Omar Abdelwahab of Mountain View CA (US)]]

[[Category:Ankit Goyal of Belmont CA (US)]]

SYSTEM AND METHOD FOR CONTEXT INSERTION FOR CONTRASTIVE SIAMESE NETWORK TRAINING - A simplified explanation of the abstract

This abstract first appeared for US patent application 18315931 titled 'SYSTEM AND METHOD FOR CONTEXT INSERTION FOR CONTRASTIVE SIAMESE NETWORK TRAINING

Simplified Explanation

- The patent application describes a method for processing input utterances that are continuations of previous utterances. - The method involves using a trained Siamese network to determine embeddings (representations) of tokens from the input utterance. - These embeddings are then combined with a context token embedding representing the class associated with the previous utterance to generate a representative embedding for the input utterance. - The method also involves determining representative embeddings for multiple possible classes, where each class has first and second threshold boundaries. - Using the Siamese network, similarity scores are calculated for each possible class based on the distance between the representative input utterance embedding and a selected threshold boundary of the representative embedding for that class. - Based on the similarity scores, a class is identified for the input utterance, and an action corresponding to the identified class is performed.

Original Abstract Submitted

A method includes receiving an input utterance that is a continuation of a previous utterance. The method also includes, using a trained Siamese network, determining input utterance embeddings representing tokens from the input utterance, pooling the input utterance embeddings with a context token embedding representing a class associated with the previous utterance to generate a representative input utterance embedding, and determining a representative embedding associated with each of multiple possible classes. Each possible class is associated with first and second threshold boundaries. The method further includes, using the trained Siamese network, determining a similarity score for each possible class based on a distance between the representative input utterance embedding and a selected threshold boundary of the representative embedding for that possible class and identifying a class for the input utterance based on the determined similarity scores. In addition, the method includes performing an action corresponding to the identified class.