CNN Architecture for Query-Response Relevance Ranking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current Convolutional Neural Networks (CNNs) face challenges in learning deep semantic knowledge from conversation data and efficiently evaluating query-response pairs due to textual overlap and processing inefficiencies, which affects their ability to provide relevant responses.
Innovation Solution
A CNN architecture is developed with a two-branch configuration, utilizing dynamic pooling and k-max pooling to process queries and responses independently, generating sentence representations, and ranking pairs based on conversational relevance using multilayer perceptron (MLP) layers for improved processing efficiency and relevance evaluation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a CNN model breaks the query into textual parts to evaluate semantic features, then semantic feature evaluation is enabled, but processing efficiency deteriorates due to textual overlap and redundancy
Solution Approach 1:
The patent segments the query into multiple parts using a splitter component, then processes each segment independently through the CNN model. This segmentation allows the model to evaluate semantic features of each part while avoiding redundant processing of overlapping text, thereby improving processing efficiency without sacrificing semantic evaluation accuracy
Solution Approach 2:
The patent extracts and removes redundant textual information from the query processing pipeline. By identifying and eliminating redundant parts through the splitter and processor components, the system maintains semantic feature evaluation capability while reducing processing time and computational resources
2Productivity
If a CNN model processes query and response independently to improve efficiency, then processing efficiency is improved, but conversational relevance evaluation deteriorates
Solution Approach 1:
The patent merges the independently processed query and response representations in the MLP (Multi-Layer Perceptron) layer. This combination allows the model to evaluate conversational relevance between the query and response while maintaining the processing efficiency benefits of independent CNN processing. The MLP integrates the separate representations to produce relevance scores that capture contextual relationships
3Measurement precision
If a CNN model uses deep learning to learn semantic knowledge from conversation data, then semantic knowledge is improved, but the ability to provide relevant responses deteriorates due to challenges in utilizing learned knowledge
Solution Approach 1:
The patent introduces an intermediary MLP layer that bridges the CNN model's learned semantic knowledge and the final response selection. This intermediary component processes the deep semantic representations from the CNN and transforms them into relevance scores that directly indicate response appropriateness, making the learned knowledge more actionable for response selection
Solution Approach 2:
The patent implements a feedback mechanism where the MLP layer continuously refines the relevance assessment based on the interaction between query and response representations. This feedback loop allows the model to adjust its relevance scoring to better reflect actual conversational context, improving response relevance while maintaining deep semantic knowledge
Data Source
AI summary
Non-limiting examples of the present disclosure describe a convolutional neural network (CNN) architecture configured to evaluate conversational relevance of query-response pairs. A CNN model is provided that can include a first branch, a second branch, and multilayer perceptron (MLP) layers. The first branch includes convolutional layers with dynamic pooling to process a query. The second branch includes convolutional layers with dynamic pooling to process candidate responses for the query. The query and the candidate responses are processed in parallel using the CNN model. The MLP layers are configured to rank query-response pairs based on conversational relevance.