A9.com, Inc.

Amazon Technologies, Inc.

Inderjit Dhillon

Hsiang-Fu Yu

Angeliki Metallinou

Rahul Goel

Po-Wei Wang

Anish Acharya

Vijai Mohan

Wei-Cheng Chang

Rajat Sen

A contextual label compression framework is presented that uses trained sequence to sequence models. A set of training data including received queries and related content can be processed to generate sequences of semantic encodings. These sequences can be used to train the sequence to sequence models, in order to be able to predict queries for instances of content when the relevant information for those instances is processed by the model. When such information is received for an instance, that information can be processed to generate a semantic encoding sequence which can then be processed by the model. A resulting semantic sequence output by the model can be segmented and decoded to produce a set of relevant queries for the instance of content. This information can then be provided to an entity associated with the instance of content for purposes in managing aspects relating to that content.

Extreme multi-label classification

Large scale time series forecasting models are described that leverage deep learning. This can include the utilization of temporal convolution networks and long short-term memory (LTSM) units of recurrent neural networks (RNNS). The model architectures can handle very large data sets with a large number of time series. Diverse scaling is provided through use of a scale-free leveling network architecture, and sparse time-series data is managed using a gating approach. A deep temporally regularized matrix factorization approach to time-series forecasting is utilized that can leverage correlations between the time series during both training and prediction.

High dimensional time series forecasting

Various embodiments of systems and methods allow unsupervised, deep learning, character-level language models to complete and correct search engine queries, given arbitrary search prefixes. Embodiments use a multi-layer, character-level, recurrent neural network trained on query logs to generate query suggestions for the user. Integrated is an error correction framework. More particularly, approaches disclosed herein for query error correction and completion combine the character-level language models with an edit-distance-based potential function calibrated to generate spelling corrections, linking the two using a tree-based beam search algorithm that can exploit the potential functions and efficiently rank the suggestions. Optimizations are made to the predictive system, and efficient processor-based computations complete the queries, with error correction, in real-time. The embodiments substantially increase hit rate over standard approaches and are capable of handling tail queries.

Neural query auto-correction and completion

A natural language understanding (NLU) system that uses a reduced dimensionality of word embedding features to configure compressed NLU models that use reduced computing resources for NLU tasks. A modified NLU model may include a compressed vocabulary data structure of word embedding data vectors that include a set of values corresponding to a reduced dimensionality of the original word embedding features, resulting in a smaller sized vocabulary data structure and reduced size of the vocabulary data structure. Further components of the modified NLU model perform matrix operations to expand the dimensionality of the reduced word embedding data vectors up to the expected dimensionality of later layers of the NLU model. Additional training and reweighting can adjust for potential loses in performance resulting from reductions in the word embedding features. Thus the modified NLU model can achieve similar performance to an original NLU model with reductions in use of computing resources.

Growing community of inventors

Berkeley, CA, United States of America

Inderjit Dhillon