Conversational Response Generation with Two-Tier Intent Clustering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems struggle to generate dynamic, timely, and pertinent conversational responses due to the sparseness of training data and individual user particularities, making it difficult to determine specific user intents accurately.
Innovation Solution
A two-tier machine learning model approach is employed, where a first model clusters specific intents into groups using unsupervised hierarchical clustering, and a second model selects a subset of these clusters based on feature inputs, allowing users to select the appropriate intent from a group, thereby reducing false positives.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning model is trained to determine specific user intents directly, then the system can generate conversational responses, but the accuracy is reduced due to sparseness of training data and similar feature inputs for different intents
Solution Approach 1:
The patent segments the intent classification problem into two stages: first clustering intents into groups based on feature similarity, then selecting the most likely cluster. This segmentation allows the system to handle sparseness by grouping similar intents together, where each cluster can be represented by a single prototype that captures the essence of multiple intents with similar feature patterns.
Solution Approach 2:
The patent introduces intent clusters as an intermediary between the raw feature inputs and the final intent determination. These clusters serve as intermediate representations that aggregate information from multiple similar intents, allowing the system to infer intent accurately even when training data for specific intents is sparse, by leveraging the collective information within clusters.
2Measurement precision
If the system aggregates detailed information about user actions and circumstances to determine intent, then response pertinence improves, but processing time increases
Solution Approach 1:
The patent performs preliminary clustering of intents into groups during an offline training phase, creating a pre-organized structure of intent clusters with representative prototypes. This preliminary action eliminates the need to process all detailed user action information in real-time, as the system only needs to compare current features against pre-computed cluster prototypes during actual conversation, significantly reducing processing time while maintaining accuracy.
3Reliability
If the system uses multiple machine learning models to cluster and select intents, then false positives are reduced, but system complexity increases
Solution Approach 1:
The patent divides the intent classification task into two sequential steps performed by two simpler models: first a clustering model that groups intents based on feature similarity, then a selection model that picks the most likely cluster. This segmentation into two simpler models is less complex than using a single complex model, as each model can be trained independently and the architecture is more modular and easier to implement.
Data Source
AI summary
Methods and systems are described for generating dynamic conversational responses using two-tier machine learning models. The dynamic conversational responses may be generated in real time and reflect the likely goals and/or intents of a user. The two-tier machine learning model may include a first tier that determines an intent cluster based on a feature input, and a second tier that determines a specific intent from the cluster.