Intent-based conversational knowledge graph for spoken language understanding system
Described are systems, methods, apparatuses, and computer program product embodiments for automatically processing intent-based spoken language for SLU. The disclosed solution uses a scale-free network structured conversational knowledge graph that stores nodes representative of actions, objects, and intent names and edges representative of relationships between the nodes. For all phrases (including a sentence) from the same intent, the system calculates a mean feature vector using a Universal Sentence Embedding (USE) model as a feature element. The system also employs a multi-step intent detection strategy. A graph query technique may be used to match all potential intent nodes from the trained knowledge graph. The system may compute a covariance matrix between the feature element of an input phrase and feature elements of all potential intents. The major component of the covariance matrix along with the maximum covariance may be used to determine the final intent.
1. A method for training and applying a knowledge graph data structure for a spoken language understanding (SLU) system, the method performed by a system comprising one or more processors, the method comprising:
receiving input training data representing one or more phrases;
training a knowledge graph data structure, the training comprising:
extracting word data representing one or more words from the input training data representing one or more phrases;
generating and storing, based on the word data representing the one or more words, a named entity node in the knowledge graph data structure;
generating taxonomy data representing one or more categories in which the one or more phrases of the input training data can be segmented;
generating and storing, based on the taxonomy data, a taxonomy node in the knowledge graph data structure;
generating intent data based on the input training data representing the one or more phrases;
generating and storing, based on the intent data, an intent node in the knowledge graph; and
generating and storing a first edge and a second edge in the knowledge graph,
wherein the first edge forms a first connection between the named entity node and the taxonomy node, and the second edge forms a second connection between the taxonomy node and the intent node;
receiving user input phrase data; and
applying one or more data processing operations, based at least in part on the user input phrase data and the trained knowledge graph data structure comprising the named entity node, the taxonomy node, the intent node, the first edge, and the second edge, to generate output data representing a final intent of the user input phrase data, wherein the one or more data processing operations comprises:
determining one or more potential intents of the user input phrase data; and
based on the determined one or more potential intents and using a covariance matrix, determine the final intent of the user input phrase data, wherein the final intent of the user input phrase data is determined based on a maximum value of the covariance matrix.
2. The method of claim 1 , further comprising:
receiving second user input phrase data;
determining whether the knowledge graph data structure includes an intent node corresponding to the second user input phrase data; and
in accordance with the determination that the knowledge graph data structure does not include the intent node corresponding to the second user input phrase data, adding the intent node and adding one or more additional nodes representative of words in the second user input phrase to the knowledge graph data structure.
3. The method of claim 1 , wherein the training of the knowledge graph data structure comprises:
for each of the one or more phrases in the input training data:
performing lemmatization on the respective phrase to reduce the phrase;
tagging parts of speech in the reduced phrase;
extracting action words and forming an action list;
extracting object words and forming an object list; and
generating and storing one or more nodes in the knowledge graph data structure corresponding to the extracted action words and the extracted object words.
4. The method of claim 1 , wherein the named entity node is of a type selected from the group of nodes comprising: action nodes corresponding to action words or object nodes corresponding to object words.