User-Log Intent Discovery Without Known Intent Counts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing intent detection systems face challenges in handling evolving user needs and new functionality due to the continuous increase in the number of intents over time, with unsupervised approaches leading to noisy clusters and semi-supervised methods requiring accurate estimates of new intents, which can result in manual effort and poor semantic utterance representations.
Innovation Solution
A deep semi-supervised contrastive clustering method that iteratively performs representation learning and clustering using a combination of semi-supervised contrastive loss and cross-entropy loss, followed by a cluster merger algorithm to refine clusters and interpret them as intents, leveraging both labeled and unlabeled utterances.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If unsupervised clustering approaches are used for intent discovery, then manual labeling effort is reduced, but cluster purity deteriorates due to noisy clusters
Solution Approach 1:
The patent applies preliminary action by pre-training a language model on large amounts of unlabeled text data before performing intent clustering. This pre-training establishes a solid foundation for semantic understanding, enabling the subsequent clustering to achieve higher purity without requiring extensive manual labeling. The model learns general language patterns and semantic relationships in advance, which then guide accurate intent discovery.
Solution Approach 2:
The patent introduces an intermediary language model that mediates between the raw user logs and the final clustering results. This intermediate representation layer captures semantic meaning from unlabeled data, allowing the clustering algorithm to operate on enriched features rather than raw text, thereby improving cluster purity while maintaining automation.
2Manufacturing precision
If semi-supervised clustering approaches are used for intent discovery, then cluster purity improves, but the system complexity increases due to requirement for accurate intent number estimation
Solution Approach 1:
The patent applies self-service by enabling the system to automatically determine the number of intents through clustering algorithms without requiring external input or manual specification. The system self-determines the appropriate number of clusters based on the data structure and semantic patterns, eliminating the need for complex manual configuration while maintaining high cluster purity.
3Loss of information
If existing semi-supervised approaches are used for intent discovery, then labeled data utilization improves, but performance deteriorates when number of new intents is unknown
Solution Approach 1:
The patent applies dynamics by implementing a flexible clustering framework that can adapt to any number of intents dynamically. The system uses iterative clustering algorithms that can identify and separate known intents from unknown intents automatically, allowing the model to handle varying numbers of categories without retraining or manual configuration. This dynamic capability maintains high performance whether the number of intents is fixed or unknown.
4Productivity
If k-means clustering is used for intent discovery, then computational efficiency improves, but clustering accuracy deteriorates due to incorrect cluster alignment over epochs
Solution Approach 1:
The patent applies feedback by implementing iterative alignment mechanisms that continuously refine cluster assignments based on semantic coherence and data characteristics. The system provides feedback loops that monitor cluster quality and adjust assignments, ensuring accurate alignment of clusters with actual intent categories while maintaining computational efficiency through automated optimization.
Data Source
AI summary
Existing semi-supervised and unsupervised approaches for intent discovery require an estimate of the number of new intents present in the user logs. The present disclosure receives labeled utterances from known intents and update parameters of a pre-trained language model (PLM). Representation learning and clustering is performed iteratively using labeled and unlabeled utterances from known intents and unlabeled utterances from unknown intents to fine-tune PLM and a plurality of clusters is generated. Cluster merger algorithm is executed iteratively on generated plurality of clusters. A query cluster is obtained by randomly selecting one cluster from the plurality of clusters and by obtaining a corresponding plurality of nearest neighbors based on a cosine-similarity. A response for merging the query cluster and corresponding plurality of nearest neighbors is obtained, and a new cluster is created. The corresponding cluster representation is recalculated and each of the new cluster is interpreted as an intent.