Question Answer Matching Using Convolutional Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional question and answer matching methods in smart transaction scenarios rely heavily on manual rule design and feature extraction, leading to inefficiencies, inconsistent feature expression, and limited generalization capabilities, particularly due to the accuracy limitations of earlier stage models like Named Entity Recognition (NER).
Innovation Solution
A question answer matching system and method utilizing a convolutional neural network that transforms user queries and suggested answers into vector representations using pre-trained word vectors, performs convolutional operations to extract features, and maps these into a sample annotating space, with distinct configuration parameters for query and answer operations, trained end-to-end with positive and negative samples from a constructed dataset.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If manual rule design and feature extraction methods are used, then the system can be implemented with conventional techniques, but the manufacturing precision and reliability of matching results are limited
Solution Approach 1:
The patent replaces manual rule-based systems with a neural network-based automated system. Specifically, it substitutes the mechanical process of manual feature extraction and rule design with an automated neural network model that learns features directly from data, thereby improving matching accuracy while reducing reliance on manual intervention and complex rule engineering
Solution Approach 2:
The patent changes the fundamental parameters of the matching system by transitioning from discrete feature extraction to continuous vector representations. It uses word2vec to convert text into continuous vector spaces and applies different convolutional filter configurations (kernel sizes, strides) to capture various n-gram features automatically, allowing the system to adaptively learn relevant features without manual specification
2Adaptability or versatility
If conventional feature extraction methods are used, then the system can process data with simple structures, but the adaptability and generalization capabilities for mass data are insufficient
Solution Approach 1:
The patent creates a universal matching system that can handle diverse question-answer pairs through a unified neural network architecture. The model uses shared embedding layers for both questions and answers, and applies the same convolutional processing framework to different types of text data, enabling it to generalize across various domains and data types without requiring domain-specific customization
Solution Approach 2:
The patent performs preliminary action by pre-training word embeddings using word2vec before the main matching task. This pre-processing step creates rich semantic representations of words that capture contextual relationships, which then serve as input features for the convolutional neural network. This preliminary feature learning enables the system to better handle unseen data and improve generalization performance
3Measurement precision
If the same configuration parameters are shared between query and answer convolutional operations, then the device complexity is reduced, but the measurement precision of matching results deteriorates
Solution Approach 1:
The patent applies local quality by using different convolutional configurations for questions and answers. Specifically, it employs distinct filter kernel sizes, strides, and potentially different numbers of filters for each input type. This allows the model to optimize feature extraction independently for each modality, capturing question-specific and answer-specific patterns with higher precision rather than forcing a one-size-fits-all configuration
Data Source
AI summary
The specification discloses a question answer matching method, system and computer storage medium. The method comprises: transforming the user query and one of one or more suggested answers corresponding to the user query by using a pre-trained word vector to obtain vector representations of the user query and the one of one or more suggested answers corresponding to the user query; performing a convolutional operation on the vector representations of the user query and the one of one or more suggested answers, respectively, to extract features; and mapping convolution results of the vector representations of the user query and the vector expression of the one of one or more suggested answers into a sample annotating space, to obtain a matching result of the user query.