Tree Kernel Learning for Intent Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current text classification systems face challenges in accurately determining the intent of text due to reliance on keyword statistics, which are insufficient as intent classes are weakly correlated with keywords, and suffer from low accuracy due to limited training sets and inconsistency across domains.
Innovation Solution
The system creates a communicative discourse tree from the utterance, combines it with a parse tree to form a parse thicket, and applies a classification model to determine intent from a predefined list of classes, using support vector machines or tree-kernel learning, and iteratively adjusts the model based on feedback for improved accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If keyword-based classification is used, then the system is simple to implement, but the classification accuracy deteriorates because intent classes are weakly correlated with keywords
Solution Approach 1:
The patent transforms the classification approach from keyword-based statistical parameters to discourse structure-based parameters. By creating discourse trees that represent rhetorical relationships and communicative actions, the system changes the fundamental parameters used for classification from surface-level keywords to deep structural features, thereby improving accuracy while maintaining reasonable implementation complexity
Solution Approach 2:
The patent adds a new dimension to text classification by incorporating discourse structure analysis. Instead of relying solely on keyword frequency and co-occurrence, the system introduces discourse trees that capture hierarchical relationships, rhetorical functions, and communicative intentions, effectively moving from a one-dimensional keyword space to a multi-dimensional structural space
2Ease of manufacture
If template matching with limited training sets is used, then the system is easier to train, but the accuracy deteriorates due to limited training set size and domain inconsistency
Solution Approach 1:
The patent creates a universal discourse tree structure that can be applied across multiple domains and intent classes. The discourse tree framework captures fundamental rhetorical relationships and communicative actions that are domain-independent, allowing the same structural approach to work across diverse domains without requiring domain-specific retraining, thereby improving both accuracy and training efficiency
3Speed
If keyword classification is used, then the processing speed is fast, but the accuracy deteriorates because keyword classification does not consider phrasing, style, or document structure
Solution Approach 1:
The patent segments text into elementary discourse units (EDUs) and builds hierarchical discourse trees from these segments. This segmentation approach breaks down complex texts into manageable rhetorical units while preserving their structural relationships, enabling the system to capture phrasing and style information efficiently without sacrificing processing speed
Solution Approach 2:
The patent performs preliminary discourse parsing to create discourse trees before classification. By pre-processing text into structured discourse representations that capture rhetorical relationships and communicative actions, the system prepares the data in advance, making the subsequent classification more accurate while the modular architecture maintains reasonable processing speed
Data Source
AI summary
Systems, devices, and methods of the present invention are related to determining an intent of an utterance. For example, an intent classification application accesses a sentence with fragments. The intent classification application creates a parse tree for the sentence. The intent classification application generates a discourse tree that represents rhetorical relationships between the fragments. The intent classification application matches each fragment that has a verb to a verb signature, thereby creating a communicative discourse tree. The intent classification application creates a parse thicket by combining the communicative discourse tree and the parse tree. The intent classification application determines an intent of the sentence from a predefined list of intent classes by applying a classification model to the parse thicket.