System and method for enhanced chatflow application

Patent metadata

Field	Value
Publication number	US-11095590-B2
Application number	US-201615279248-A
Country	US
Kind code	B2
Filing date	Sep 28, 2016
Priority date	Sep 28, 2016
Publication date	Aug 17, 2021
Grant date	Aug 17, 2021

Abstract

Embodiments provide a computer implemented method, in a data processing system including a processor and a memory including instructions which are executed by the processor to cause the processor to train an enhanced chatflow system, the method including: ingesting a corpus of information including at least one user input node corresponding to a user question and at least one variation for each user input node; for each user input node: designating the node as a class; storing the node in a dialog node repository; designating each of the at least one variations as training examples for the designated class; converting the classes and the training examples into feature vector representations; training one or more training classifiers using the one or more feature vector representations of the classes; and training classification objectives using the one or more feature vector representations of the training examples.

First claim

What is claimed is:

1. A computer implemented method, in a data processing system comprising a processor and a memory comprising instructions which are executed by the processor to cause the processor to train an enhanced chatflow system, the method comprising:
ingesting a corpus of information comprising at least one user input node corresponding to a user question and a plurality of variations for each user input node, wherein each variation is a different semantical reordering of the user question, wherein the at least one user input node is derived from previously collected user questions in an already deployed rule-based dialog system;
for each user input node:
designating the node as a class;
storing the node in a dialog node repository;
designating the plurality of variations as training examples for the designated class;

converting the classes and the training examples into fixed-length feature vector representations through a first conversion approach, wherein in the first conversion approach, variable length sentences are transformed to the fixed-length feature vector representations using convolutional neural networks;
training a classifier using the fixed-length feature vector representations of the classes;
training at least one classification objective using the fixed-length feature vector representations of the training examples;
providing a new user input to the trained classifier;
performing regular expression and exact matching to determine if the new user input corresponds to a known dialog node;
returning a known dialog node confidence score;
converting the new user input into a user input vector representation through a second conversion approach different than the first conversion approach, wherein in the second conversion approach, the new user input is converted by directly combining a word vector of each word in the new user input;
using the trained classifier, determining the probability that the user input vector representation belongs to a particular dialog node;
returning a matching dialog node confidence score based on the probability;
performing a fusion analysis of the known dialog node confidence score and the matching dialog node confidence score using a recall weighted score; and
determining a chosen dialog node based upon the fusion analysis.

2. The method as recited in claim 1 , further comprising:
converting the probability into a confidence score.

3. The method as recited in claim 1 , further comprising:
using the classifier, calculating the probability that the user input vector representation corresponds to a direct response trigger; and
converting the probabilities into one or more confidence scores.

4. The method as recited in claim 1 , further comprising:
performing regular expression matching and exact matching to determine if the user input corresponds to a known dialog node;
if the user input corresponds to a known dialog node:
processing the user input in accordance with the known dialog node;

if the user input does not correspond to a known dialog node:
converting the user input into a user input vector representation;
determining a classifier corresponding to the user input vector representation;
using the classifier, determining the probability that the user input vector representation belongs to a particular dialog node; and
processing the user input in accordance with the particular dialog node.

5. The method as recited in claim 4 , further comprising:
if the user input does not correspond to a known dialog node:
processing the user input in accordance with the particular dialog node having the highest determined probability.

6. A computer program product for enhanced chatflow, the computer program product comprising a computer readable storage medium having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:
ingest a corpus of information comprising at least one user input node corresponding to a user question and a plurality of variations for each user input node, wherein each variation is a different semantical reordering of the user question, wherein the at least one user input node is derived from previously collected user questions in an already deployed rule-based dialog system;
for each user input node:
designate the node as a class;
store the node in a dialog node repository;
designate the plurality of variations as training examples for the designated class;

convert the classes and the training examples into fixed-length feature vector representations through a first conversion approach, wherein in the first conversion approach, variable length sentences are transformed to the fixed-length feature vector representations using convolutional neural networks;
train a classifier using the fixed-length feature vector representations of the classes;
train at least one classification objective using the fixed-length feature vector representations of the training examples;
provide a new user input to the trained classifier;
perform regular expression and exact matching to determine if the new user input corresponds to a known dialog node;
return a known dialog node confidence score;
convert the new user input into a user input vector representation through a second conversion approach different than the first conversion approach, wherein in the second conversion approach, the new user input is converted by directly combining a word vector of each word in the new user input;
using the trained classifier, determine the probability that the user input vector representation belongs to a particular dialog node;
return a matching dialog node confidence score based on the probability;
perform a fusion analysis of the known dialog node confidence score and the matching dialog node confidence score using a recall weighted score; and
determine a chosen dialog node based upon the fusion analysis.

7. The computer program product as recited in claim 6 , wherein the processor is further caused to:
convert the probability into a confidence score.

8. The computer program product as recited in claim 6 , wherein the processor is further caused to:
using the classifier, calculate the probability that the user input vector representation corresponds to a direct response trigger; and
convert the probabilities into one or more confidence scores.

9. The computer program product as recited in claim 6 , wherein the processor is further caused to:
if the user input does not correspond to a known dialog node:
process the user input in accordance with the particular dialog node having the highest determined probability.

10. The computer program product as recited in claim 6 , wherein the processor is further caused to:
perform regular expression and exact matching to determine if the user input corresponds to a known dialog node;
return a known dialog node confidence score;
convert the user input into a user input vector representation;
determine the classifier corresponding to the user input vector representation;
using the classifier, determine the probability that the user input vector representation belongs to a particular dialog node;
return a matching dialog node confidence score;
perform a fusion analysis of the known dialog node confidence score and the matching dialog node confidence score using a recall weighted score; and
determine a chosen dialog node based upon the fusion analysis.

11. An enhanced chatflow system, comprising:
a processor configured to:
ingest a corpus of information comprising at least one user input node corresponding to a user question and a plurality of variations for each u

Assignees

• IBM

Inventors

• Bakis Raimo
• Kunc Ladislav
• Nahamoo David
• Polymenakos Lazaros
• Zakos John

Classifications

• H04L51/216 Primary

  Handling conversation history, e.g. grouping of messages in sessions or threads · CPC title

• H04L51/18 Primary

  Commands or executable codes · CPC title

• G06N20/00

  Machine learning · CPC title

• H04L51/16 Primary

  Electricity · mapped topic