Utilizing a dynamic memory network for state tracking

Patent metadata

Field	Value
Publication number	US-2021118430-A1
Application number	US-202017135629-A
Country	US
Kind code	A1
Filing date	Dec 28, 2020
Priority date	Sep 19, 2018
Publication date	Apr 22, 2021
Grant date	—

Abstract

The present disclosure relates to generating digital responses based on digital dialog states generated by a neural network having a dynamic memory network architecture. For example, in one or more embodiments, the disclosed system provides a digital dialog having one or more segments to a dialog state tracking neural network having a dynamic memory network architecture that includes a set of multiple memory slots. In some embodiments, the dialog state tracking neural network further includes update gates and reset gates used in modifying the values stored in the memory slots. For instance, the disclosed system can utilize cross-slot interaction update/reset gates to accurately generate a digital dialog state for each of the segments of digital dialog. Subsequently, the system generates a digital response for each segment of digital dialog based on the digital dialog state.

First claim

What is claimed is:

1 . A non-transitory computer-readable medium storing instructions thereon that, when executed by at least one processor, cause a computing device to:
provide a segment of digital dialog to a dynamic memory network comprising a plurality of memory slots, wherein each memory slot corresponds to a designated dialog state characteristic; generate, utilizing the dynamic memory network, a digital dialog state for the segment of digital dialog by:
determining a value corresponding to a first designated dialog state characteristic for a first memory slot from the plurality of memory slots based on the segment of digital dialog; and
determining the digital dialog state based on the value of the first memory slot that corresponds to the first designated dialog state characteristic; and
generate a digital response to the segment of digital dialog based on the digital dialog state.

2 . The non-transitory computer-readable medium of claim 1 ,
further comprising instructions that, when executed by the at least one processor, cause the computing device to determine a previous value stored in the first memory slot, the previous value corresponding to at least one previous segment of digital dialog, wherein determining the value corresponding to the first designated dialog state characteristic for the first memory slot from the plurality of memory slots based on the segment of digital dialog comprises determining the value based on the segment of digital dialog and the previous value stored in the first memory slot.

3 . The non-transitory computer-readable medium of claim 1 , wherein determining the value corresponding to the first designated dialog state characteristic for the first memory slot comprises determining the value utilizing at least one of a reset gate or an update gate.

4 . The non-transitory computer-readable medium of claim 3 , wherein determining the value utilizing the at least one of the reset gate or the update gate comprises:
determining, utilizing a cross-slot interaction reset gate, a cross-slot interaction reset value based on cross-slot interactions between the first memory slot and at least one other memory slot from the plurality of memory slots of the dynamic memory network; and determining the value corresponding to the first designated dialog state characteristic for the first memory slot based on the segment of digital dialog and the cross-slot interaction reset value.

5 . The non-transitory computer-readable medium of claim 3 , wherein determining the value utilizing the at least one of the reset gate or the update gate comprises:
determining, utilizing a cross-slot interaction update gate, a cross-slot interaction update value based on cross-slot interactions between the first memory slot and at least one other memory slot from the plurality of memory slots of the dynamic memory network; and determining the value corresponding to the first designated dialog state characteristic for the first memory slot based on the segment of digital dialog and the cross-slot interaction update value.

6 . The non-transitory computer-readable medium of claim 1 ,
further comprising instructions that, when executed by the at least one processor, cause the computing device to further utilize the dynamic memory network to generate the digital dialog state for the segment of digital dialog by determining an additional value corresponding to a second designated dialog state characteristic for a second memory slot from the plurality of memory slots based on the segment of digital dialog, the second designated dialog state characteristic different from the first designated dialog state characteristic, wherein determining the digital dialog state based on the value of the first memory slot that corresponds to the first designated dialog state characteristic comprises determining the digital dialog state based on the value of the first memory slot and the additional value of the second memory slot.

7 . The non-transitory computer-readable medium of claim 1 , wherein determining the value corresponding to the first designated dialog state characteristic for the first memory slot comprises determining the value corresponding to one of a dialog topic, a location, an entity, or an action.

8 . The non-transitory computer-readable medium of claim 1 ,
further comprising instructions that, when executed by the at least one processor, cause the computing device to receive an audio representation of the segment of digital dialog, wherein the instructions, when executed by the at least one processor, cause the computing device to generate the digital response to the segment of digital dialog by generating an audio response comprising the digital response.

9 . A system comprising:
at least one processor; and at least one non-transitory computer-readable medium comprising instructions that, when executed by the at least one processor, cause the system to:
provide a segment of digital dialog to a dynamic memory network comprising a plurality of memory slots and a plurality of gating mechanisms, wherein each memory slot is associated with a designated dialog state characteristic and each gating mechanism corresponds to a memory slot;
generate, utilizing one or more neural network layers of the dynamic memory network, a feature vector comprising a representation of the segment of digital dialog;
determine, utilizing the plurality of gating mechanisms of the dynamic memory network, values for the plurality of memory slots by, for a first memory slot:
determining a previous value stored in the first memory slot, the previous value corresponding to a previous segment of digital dialog; and
generating, utilizing a gating mechanism corresponding to the first memory slot, a value corresponding to a first designated dialog state characteristic associated with the first memory slot based on at least one of the feature vector for the segment of digital dialog or the previous value stored in the first memory slot;

determine, utilizing one or more additional neural network layers of the dynamic memory network, a digital dialog state based on the values for the plurality of memory slots; and
generate a digital response to the segment of digital dialog based on the digital dialog state.

10 . The system of claim 9 , wherein generating, utilizing the gating mechanism corresponding to the first memory slot, the value corresponding to the first designated dialog state characteristic associated with the first memory slot based on the at least one of the feature vector for the segment of digital dialog or the previous value stored in the first memory slot comprises:
determining, utilizing an update gate of the gating mechanism, an update value; and generating the value for the first memory slot based on the update value and the at least one of the feature vector for the segment of digital dialog or the previous value stored in the first memory slot.

11 . The system of claim 10 , wherein generating the value for the first memory slot based on the update value and the at least one of the feature vector for the segment of digital dialog or the previous value stored in the first memory slot comprises:
determining an impact of the feature vector for the segment of digital dialog on the value utilizing the update value; and generating the value for the first memory slot utilizing the previous value stored in the first memory slot and the feature vector for the segment of digital dialog based on the impact of the feature vector for the segment of digital dialog.

12 . The system of claim 10 , wherein generating, utilizing the gating mechanism correspo

Assignees

• Adobe Inc

Inventors

• Kim Seokhwan
• Chang Walter

Classifications

• G06N3/045

  Combinations of networks · CPC title

• G06N3/044 Primary

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/048

  Activation functions · CPC title

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/09

  Supervised learning · CPC title