Dismiss
InnovationQ will be updated on Sunday, Oct. 22, from 10am ET - noon. You may experience brief service interruptions during that time.
Browse Prior Art Database

Method and System for Providing Semi-Supervised User Intent Detection for Multi-Domain Dialogues

IP.com Disclosure Number: IPCOM000249786D
Publication Date: 2017-Apr-05
Document File: 5 page(s) / 123K

Publishing Venue

The IP.com Prior Art Database

Related People

Aasish Pappu: INVENTOR [+3]

Abstract

A method and system is disclosed for providing semi-supervised user intent detection for multi-domain dialogues by enabling an agent to learn an inventory of intents from a small set of task-oriented user utterances and automatically clustering the intents. The method and system also enables the agent to reliably recognize the complex user intent on previously unseen user activities or by observing sequence of applications using graph-based semi-supervised learning methods.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 34% of the total text.

Method and System for Providing Semi-Supervised User Intent Detection for Multi-Domain Dialogues

Abstract

A method and system is disclosed for providing semi-supervised user intent detection for multi-domain dialogues by enabling an agent to learn an inventory of intents from a small set of task-oriented user utterances and automatically clustering the intents.  The method and system also enables the agent to reliably recognize the complex user intent on previously unseen user activities or by observing sequence of applications using graph-based semi-supervised learning methods.

Description

Users interact with mobile applications (apps) to fulfill certain intents such as, but not limited to, finding restaurants.  Some intents are complex and may span multiple apps such as, but not limited to, a restaurant app, a messenger app and a calendar app, in order to plan a dinner with friends.

In order to reduce the effort from third-party for building specific apps to handle such complex intents such as dinner planning, there is a need for an intelligent agent that actively learns to understand these intents, thereby providing timely assistance when needed.

Disclosed is a method and system for providing semi-supervised user intent detection for multi-domain dialogues by enabling an agent to learn an inventory of intents from a small set of task-oriented user utterances and automatically clustering the intents.  The method and system also enables the agent to reliably recognize the complex user intent on previously unseen user activities or by observing sequence of applications using graph-based semi-supervised learning methods.

The agent learns an inventory of complex user intents each time when enough new observations are accumulated and the agent updates the inventory with the following outcomes such as, but not limited to, new intents that can be learned and new examples that can be added to existing intents.  In a nutshell, this can be achieved by automatically clustering interactions into groups. 

In order to form an array of intents, the following resources are utilized to mine user intents.

‘R1’ denotes sequences of apps users interact with on smartphones.  This is noisy from the perspective of multi-app activities since some intents are fulfilled by single apps.  Besides, some apps may be background noise.

‘R2’ denotes user knowledge of the intent-embedded sequence structures and the nature of the intents.

‘R3’ denotes speech commands to perform actions that compose complex intents.

The figure below illustrates a process flow of the method and system in accordance with an embodiment.

Figure

As illustrated in the figure, in order to cluster seen activities into groups, a vector representation for each data point and a clustering algorithm are required.

Sequences in R2 are represented as , where and are (fixed-size) vector representations of app sequence and intent description respectively.

Similarly, activities in R3 are rep...