Google Inc.

Amazon Technologies, Inc.

Sanqiang Zhao

Yang Song

Raghav Gupta

Dengyong Zhou

Robinson Piramuthu

Gokhan Tur

Yang Liu

Seokhwan Kim

Eshan Bhatnagar

Taehwan Kim

Provided is a knowledge distillation technique for training a student language model that, relative to a larger teacher language model, has a significantly smaller vocabulary, lower embedding dimensions, and/or hidden state dimensions. Specifically, aspects of the present disclosure are directed to a dual-training mechanism that trains the teacher and student language models simultaneously to obtain optimal word embeddings for the student vocabulary. In some implementations, this approach can be combined with learning shared projection matrices that transfer layer-wise knowledge from the teacher language model to the student language model. Example experimental results have also demonstrated higher compression efficiency and accuracy when compared with other state-of-the-art compression techniques, including the ability to compress the BERTmodel by more than 60×, with only a minor drop in downstream task metrics, resulting in a language model with a footprint of under 7 MB.

Extreme language model compression with optimal sub-words and shared projections

Techniques for rendering visual content, in response to one or more utterances, are described. A device receives one or more utterances that define a parameter(s) for desired output content. A system (or the device) identifies natural language data corresponding to the desired content, and uses natural language generation processes to update the natural language data based on the parameter(s). The system (or the device) then generates an image based on the updated natural language data. The system (or the device) also generates video data of an avatar. The device displays the image and the avatar, and synchronizes movements of the avatar with output of synthesized speech of the updated natural language data. The device may also display subtitles of the updated natural language data, and cause a word of the subtitles to be emphasized when synthesized speech of the word is being output.

Virtual conversational companion

Provided is a knowledge distillation technique for training a student language model that, relative to a larger teacher language model, has a significantly smaller vocabulary, lower embedding dimensions, and/or hidden state dimensions. Specifically, aspects of the present disclosure are directed to a dual-training mechanism that trains the teacher and student language models simultaneously to obtain optimal word embeddings for the student vocabulary. In some implementations, this approach can be combined with learning shared projection matrices that transfer layer-wise knowledge from the teacher language model to the student language model. Example experimental results have also demonstrated higher compression efficiency and accuracy when compared with other state-of-the-art compression techniques, including the ability to compress the BERTmodel by more than 60&#xd7;, with only a minor drop in downstream task metrics, resulting in a language model with a footprint of under 7 MB.

Growing community of inventors

Pittsburgh, PA, United States of America

Sanqiang Zhao