Context-Aware Virtual Agent Reply Generation Using RNN Encoding
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current turn-based interaction systems between customers and virtual agents fail to capture the context of conversations, leading to sub-optimal response quality, as they do not consider previous conversational inputs when generating replies.
Innovation Solution
A computer-implemented method and apparatus that uses a recurrent neural network (RNN) based model to encode and decode multiple turns of conversation, including customer and agent interactions, within a virtual bounding box to generate context vector representations, allowing for context-aware virtual agent replies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If trained responses are identified based on current customer input alone, then response generation is simple and fast, but response quality is sub-optimal and lacks contextual awareness
Solution Approach 1:
The system performs preliminary encoding of conversational inputs into context vectors before response generation. The RNN encoder processes previous customer inputs and agent responses to create contextual representations in advance, which are then used during response generation to improve quality without adding complexity at the decision point
Solution Approach 2:
Context vectors serve as intermediary representations between raw conversational inputs and final responses. The RNN encoder transforms sequences of inputs into compressed contextual vectors that capture essential information, acting as a mediator that enables contextual awareness while maintaining system efficiency
2Adaptability or versatility
If context information is incorporated into response generation, then response relevance improves, but computational complexity increases
Solution Approach 1:
The system changes the parameter representation from raw text sequences to compressed context vectors. By transforming the input format through RNN encoding, the system captures contextual information in a condensed vector form that is more efficient to process during response generation
Solution Approach 2:
The conversational context is segmented into discrete inputs that are sequentially processed by the RNN encoder. Each conversational turn is encoded independently into a context vector, allowing the system to manage complex contextual information through modular processing
3Measurement precision
If multiple previous conversational inputs are processed, then contextual accuracy improves, but processing time increases
Solution Approach 1:
Context vectors are computed in advance during the encoding phase, before response generation begins. This preliminary processing allows the system to have contextual representations ready when needed, reducing processing time during actual interaction
Solution Approach 2:
The RNN encoder processes sequences of varying lengths and can discard less relevant previous inputs while retaining essential contextual information. The fixed-size context vector recovers only the most important contextual features, filtering out redundant information
Data Source
AI summary
A method and apparatus for facilitating a turn-based interaction between a virtual agent and a customer of an enterprise are disclosed. The method includes receiving a conversational input provided by the customer during a turn-based interaction between the customer and the agent. One or more conversational inputs exchanged between the customer and the agent prior to the customer's conversational input are identified by positioning a virtual bounding box of fixed width over textual representation of the turn-based interaction. The conversational input and the one or more conversational inputs configure a set of conversational inputs. At least one context vector representation is generated based on an encoding of the set of conversational inputs. Each word of a virtual agent reply is predicted based on the at least one context vector representation. The virtual agent reply is provided to the customer in response to the conversational input of the customer.