Processing natural language using machine learning to determine slot values based on slot descriptors

Abstract

Determining slot value(s) based on received natural language input and based on descriptor(s) for the slot(s). In some implementations, natural language input is received as part of human-to-automated assistant dialog. A natural language input embedding is generated based on token(s) of the natural language input. Further, descriptor embedding(s) are generated (or received), where each of the descriptor embeddings is generated based on descriptor(s) for a corresponding slot that is assigned to a domain indicated by the dialog. The natural language input embedding and the descriptor embedding(s) are applied to layer(s) of a neural network model to determine, for each of the slot(s), which token(s) of the natural language input correspond to the slot. A command is generated that includes slot value(s) for slot(s), where the slot value(s) for one or more of the slot(s) are determined based on the token(s) determined to correspond to the slot(s).

1 . A method implemented by one or more processors, comprising:
 receiving natural language input generated based on user interface input during a human-to-automated assistant dialog;   generating a token embedding of tokens determined based on the natural language input;   selecting a domain based on the human-to-automated assistant dialog;   determining at least one slot descriptor embedding for at least one textual descriptor of a slot assigned to the selected domain, the at least one textual slot descriptor embedding determined based on the at least one slot descriptor or the slot descriptor embedding being assigned to the selected domain in one or more computer readable media;   determining, based on application of the token embedding and the slot descriptor embedding to a trained neural network model, that one or more of the tokens correspond to the slot assigned to the selected domain;   generating an agent command that includes a slot value for the slot that is based on the token determined to correspond to the slot; and   transmitting the agent command to an agent over one or more networks, wherein the agent command causes the agent to generate responsive content and transmit the responsive content over one or more networks.   
     
     
         2 . The method of  claim 1 , wherein selecting the domain comprises selecting the agent based on the human-to-automated assistant dialog, and wherein the at least one slot descriptor embedding is determined based on the at least one slot descriptor or the slot descriptor embedding being assigned to the agent. 
     
     
         3 . The method of  claim 1 , further comprising:
 receiving the responsive content generated by the agent.   
     
     
         4 . The method of  claim 3 , further comprising:
 transmitting, to a client device at which the user interface input was provided, output that is based on the responsive content generated by the agent.   
     
     
         5 . The method of  claim 1 , wherein determining, based on application of the token embedding and the slot descriptor embedding to the trained neural network model, that one or more of the tokens correspond to the slot assigned to the selected domain comprises:
 applying both the token embedding and the slot descriptor embedding to a combining layer of the trained neural network model.   
     
     
         6 . The method of  claim 5 , wherein the combining layer is a feed forward layer. 
     
     
         7 . The method of  claim 1 , wherein generating the token embedding of the tokens of the natural language input comprises:
 applying the tokens to a memory layer of the trained neural network model to generate the token embedding.   
     
     
         8 . The method of  claim 7 , wherein the memory layer is a bi-directional memory layer comprising a plurality of memory units. 
     
     
         9 . The method of  claim 7 , wherein generating the token embedding of the tokens of the natural language input further comprises:
 applying one or more annotations of one or more of the tokens to the memory layer to generate the token embedding.   
     
     
         10 . The method of  claim 7 , wherein the combining layer is downstream from the memory layer, and upstream from one or more additional layers of the neural network model. 
     
     
         11 . The method of  claim 10 , wherein the one or more additional layers include at least one of:
 an additional memory layer; and   an affine layer.   
     
     
         12 . A system, comprising:
 memory including instructions;   one or more processors operable to execute the instructions to:
 receive natural language input generated based on user interface input during a human-to-automated assistant dialog; 
 generate a token embedding of tokens determined based on the natural language input; 
 select a domain based on the human-to-automated assistant dialog; 
 determine at least one slot descriptor embedding for at least one textual descriptor of a slot assigned to the selected domain, the at least one textual slot descriptor embedding determined based on the at least one slot descriptor or the slot descriptor embedding being assigned to the selected domain in one or more computer readable media; 
 determine, based on application of the token embedding and the slot descriptor embedding to a trained neural network model, that one or more of the tokens correspond to the slot assigned to the selected domain; 
 generate an agent command that includes a slot value for the slot that is based on the token determined to correspond to the slot; and 
 transmit the agent command to an agent over one or more networks, wherein the agent command causes the agent to generate responsive content and transmit the responsive content over one or more networks. 
   
     
     
         13 . The system of  claim 12 , wherein in selecting the domain one or more of the processors are to select the agent based on the human-to-automated assistant dialog, and wherein the at least one slot descriptor embedding is determined based on the at least one slot descriptor or the slot descriptor embedding being assigned to the agent. 
     
     
         14 . The system of  claim 12 , wherein in executing the instructions or more of the processors are further to:
 receive the responsive content generated by the agent.   
     
     
         15 . The system of  claim 14 , wherein in executing the instructions or more of the processors are further to:
 transmit, to a client device at which the user interface input was provided, output that is based on the responsive content generated by the agent.   
     
     
         16 . The system of  claim 12 , wherein in determining, based on application of the token embedding and the slot descriptor embedding to the trained neural network model, that one or more of the tokens correspond to the slot assigned to the selected domain, one or more of the processors are to:
 apply both the token embedding and the slot descriptor embedding to a combining layer of the trained neural network model.   
     
     
         17 . The system of  claim 16 , wherein the combining layer is a feed forward layer. 
     
     
         18 . The system of  claim 12 , wherein in generating the token embedding of the tokens of the natural language input, one or more of the processors are to:
 apply the tokens to a memory layer of the trained neural network model to generate the token embedding.   
     
     
         19 . The system of  claim 18 , wherein the memory layer is a bi-directional memory layer comprising a plurality of memory units. 
     
     
         20 . The system of  claim 18 , wherein in generating the token embedding of the tokens of the natural language input, one or more of the processors are further to:
 apply one or more annotations of one or more of the tokens to the memory layer to generate the token embedding.