Granted Patent B2

US 10,853,579 · App. 16/170,034 · Granted Dec 1, 2020

Inventors: Srivatsan Laxman (Palo Alto, CA); Devang Savita Ram Mohan (Bangalore, IN); Supriya Rao (Palo Alto, CA)

G06F40/30G06N3/08

Abstract

Claims (30)

1. A method useful for goal-oriented dialog automation comprising:

receiving an input message;

implementing an entity tagging operation on the input message, wherein the entity tagging operation implements a hierarchical sequence labelling based entity tagger to label and tag the input message;

tagging the message context of the input message to generate a tagged message context;

with a semantic frame extractor, implementing semantic frame extraction from the tagged message context, wherein the semantic frame extractor extracts a semantic frame from a conversation dialog of the input message, wherein the input message has been annotated with labels and tags by the entity tag operation, and wherein the semantic frame extractor implements hierarchical sequence labelling to infer frames from the conversational dialog, and wherein the semantic frame comprises a directed acyclic graph (DAG), and wherein the DAG comprises a set of nodes that recursively resolve to a collection of entity values such that when a node of the DAG contains an entity value, the node becomes a leaf node of the DAG;

implementing an entity interpretation on the extracted frame;

accessing a database to determine a business schedule and a client profile, wherein the business schedule and the client profile are related to the input message;

implementing a retrieval engine, wherein the retrieval engine obtains one or more response templates;

generating a ranked list of candidate templates from the output of the retrieval engine by computing a rank-score score using an Long short-term memory (LSTM)-based encoder-decoder architecture with a hierarchical attention mechanism, wherein the LSTM-based encoder-decoder architecture comprises a set of LSTM units for layers of a recurrent neural network (RNN), wherein the LSTM units are stored in an LSTM network, wherein a LSTM unit is composed of a cell, an input gate, an output gate and a forget gate, and wherein the cell is used to remember a set of values over an arbitrary time intervals;

based on the output of the entity interpretation, the business schedule and the client profile, and the ranked list of candidate templates, implementing a candidate eliminator; and

based on the output of the candidate eliminator, providing a set of recommended responses, wherein each recommend response is associated with a confidence score.

2. The method of claim 1 , wherein the input message comprises a voice message.

3. The method of claim 1 , wherein the input message comprises a text message.

4. The method of claim 3 further comprising:

passing the ranked list of candidate templates to a candidate extractor, wherein the candidate extractor filters the responses to ensure that the response is semantically consistent with the semantic frame and the availability returned by a relevant database (DB) does not violate the business rules.

5. A computerized system useful for goal-oriented dialog automation, comprising:

at least one processor configured to execute instructions;

a memory containing instructions when executed on the processor, causes the at least one processor to perform operations that:

receive an input message;

implement an entity tagging operation on the input message, wherein the entity tagging operation implements a hierarchical sequence labelling based entity tagger to label and tag the input message;

tag the message context of the input message to generate a tagged message context;

with a semantic frame extractor, implementing semantic frame extraction from the tagged message context, wherein the semantic frame extractor extracts a semantic frame from a conversation dialog of the input message, wherein the input message has been annotated with labels and tags by the entity tag operation, and wherein the semantic frame extractor implements hierarchical sequence labelling to infer frames from the conversational dialog, and wherein the semantic frame comprises a directed acyclic graph (DAG), and wherein the DAG comprises a set of nodes that recursively resolve to a collection of entity values such that when a node of the DAG contains an entity value, the node becomes a leaf node of the DAG;

implement an entity interpretation on the extracted frame;

access a database to determine a business schedule and a client profile, wherein the business schedule and the client profile are related to the input message;

implement a retrieval engine, wherein the retrieval engine obtains one or more response templates;

generate a ranked list of candidate templates from the output of the retrieval engine by computing a rank-score score using an Long short-term memory (LSTM)-based encoder-decoder architecture with a hierarchical attention mechanism, wherein the LSTM-based encoder-decoder architecture comprises a set of LSTM units for layers of a recurrent neural network (RNN), wherein the LSTM units are stored in an LSTM network, wherein a LSTM unit is composed of a cell, an input gate, an output gate and a forget gate, and wherein the cell is used to remember a set of values over an arbitrary time intervals;

based on the output of the entity interpretation, the business schedule and the client profile, and the ranked list of candidate templates, implement a candidate eliminator; and

based on the output of the candidate eliminator, provide a set of recommended responses, wherein each recommend response is associated with a confidence score.

6. The computerized system of claim 5 , wherein the input message comprises a voice message.

7. The computerized system of claim 5 , wherein the input message comprises a text message.

Assignments (5)

Continuity (1)

Cited By (14)

US 12,354,021 US 12,444,412 US 12,488,785 US 12,505,288 US 12,525,224 US 12,562,150 US 12,592,220 US 12,592,239 US 12,664,190 US 12,670,325 US 12,676,144 US 12,682,183 US 12,706,098 US 12,711,951