Deep Clustering Platform for Mining User Intentions
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Designing and refining Intelligent Virtual Assistants (IVAs) for customer service and sales support requires significant human effort to mine latent user intentions from large volumes of customer service texts, as existing methods face challenges in text representation, clustering, and the unknown number of clusters, leading to inefficiencies in automating customer service tasks.
Innovation Solution
An analysis platform combining unsupervised and semi-supervised approaches, integrating language models fine-tuning, distributed K-nearest neighbor graph building, and community detection techniques, which includes a language models fine-tuning module, a KNN-graph building module, and a clustering module using K-means and Louvain algorithms to mine intentions and topics from texts.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If human analysts manually mine customer service texts to identify user intentions, then the quality and accuracy of intent identification is maintained, but the time and effort required increases significantly
Solution Approach 1:
The patent introduces an analysis platform as an intermediary system between customer service texts and intent identification. This platform uses unsupervised and semi-supervised deep clustering approaches with language models (BERT, RoBERTa) to automatically mine and organize user intentions from text data, reducing manual analyst involvement while maintaining identification accuracy through sophisticated NLP techniques
Solution Approach 2:
The patent replaces the mechanical manual analysis process with an automated computational system. The analysis platform uses deep learning models and clustering algorithms to substitute human analysts in the intent mining process, achieving both time efficiency and maintained accuracy through advanced machine learning techniques
2Device complexity
If traditional text clustering methods are used to organize customer service texts, then the process is simpler, but the clustering quality and ability to identify latent intentions deteriorates
Solution Approach 1:
The patent changes key parameters of the clustering process by using deep learning language models (BERT, RoBERTa) to generate high-quality text embeddings, combining unsupervised and semi-supervised approaches, and implementing a two-stage clustering process (initial clustering followed by refinement). These parameter changes significantly improve clustering quality for identifying latent intentions while managing complexity through modular architecture
Solution Approach 2:
The patent creates a composite clustering approach by integrating multiple techniques: unsupervised deep clustering, semi-supervised learning with language models, and community detection algorithms. This composite methodology combines the strengths of different approaches to achieve superior clustering quality that can identify latent user intentions effectively
3Adaptability or versatility
If the number of clusters is unknown in the clustering process, then the method is more adaptable to different datasets, but the difficulty of determining the correct number of clusters increases
Solution Approach 1:
The patent implements a self-service approach where the clustering system automatically determines the optimal number of clusters without requiring manual specification. The semi-supervised deep clustering algorithm autonomously identifies the appropriate cluster count by analyzing data characteristics and applying community detection techniques, making the system adaptable to different datasets while eliminating the difficulty of manual cluster number determination
4Measurement precision
If language models are fine-tuned for specific domains to improve text representation, then the text understanding accuracy improves, but the time and computational resources required increase
Solution Approach 1:
The patent applies partial fine-tuning of language models by selectively adapting pre-trained models (BERT, RoBERTa) to specific domains using domain-specific customer service texts. Rather than complete retraining, the system performs targeted fine-tuning on relevant portions of the model, achieving improved text representation accuracy for intent identification while significantly reducing the time and computational resources compared to full model training
Data Source
AI summary
An analysis platform combines unsupervised and semi-supervised approaches to quickly surface and organize relevant user intentions from conversational text (e.g., from natural language inputs). An unsupervised and semi-supervised pipeline is provided that integrates the fine-tuning of high performing language models via a language models fine-tuning module, a distributed KNN-graph building method via a KNN-graph building module, and community detection techniques for mining the intentions and topics from texts via an intention mining module.