Event-Centric Commonsense Reasoning for Contextual Response Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing task-oriented dialogue systems face challenges in incorporating scalable and diverse commonsense knowledge, particularly event-centric knowledge, due to the limitations of symbolic knowledge graphs, which are costly to construct and maintain, and do not scale well across domains and languages, leading to restricted reasoning capabilities and language expression mismatches.
Innovation Solution
A computer-implemented method and system that leverages a generative knowledge graph through an encoder-decoder architecture, utilizing neural event-centric commonsense reasoning to guide input data towards output responses, enabling scalable and meaningful human-machine interaction by inferring event-centric knowledge as an intermediate hop in multi-hop reasoning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If symbolic knowledge graphs are used to represent external knowledge, then knowledge structure is clear and interpretable, but scalability to other domains and languages is poor
Solution Approach 1:
The patent replaces the mechanical symbolic knowledge graph system with a neural embedding-based system. Instead of using explicit symbolic representations (triples, graphs), the system uses continuous vector embeddings to represent knowledge, enabling scalable cross-lingual and cross-domain applications while maintaining interpretability through attention mechanisms and visualization tools.
Solution Approach 2:
The patent changes the parameter representation from discrete symbolic values to continuous embedding vectors. This transformation allows the knowledge representation to capture nuanced relationships and enables scaling to multiple domains and languages through learned embeddings rather than hand-crafted symbolic structures.
2Quantity of substance
If entity-centric symbolic knowledge graphs are used, then encyclopedic knowledge is well-represented, but reasoning capabilities are restricted
Solution Approach 1:
The patent creates a universal knowledge representation that serves multiple functions simultaneously. The neural embedding system can represent encyclopedic knowledge while also enabling complex reasoning tasks, event detection, and relationship inference through the same unified framework, rather than requiring separate systems for each function.
Solution Approach 2:
The patent adds dimensional richness to knowledge representation by using multi-dimensional embedding vectors instead of flat symbolic structures. This allows the system to capture entity attributes, relationships, and contextual information in a unified vector space, enabling both knowledge storage and reasoning operations.
3Loss of information
If fine-grained conditioned knowledge is added to knowledge bases, then knowledge completeness improves, but knowledge base complexity increases
Solution Approach 1:
The patent merges multiple knowledge aspects (entities, events, conditions, relationships) into a unified neural embedding representation. Instead of maintaining separate knowledge bases for different types of information, the system integrates them all into a single continuous vector space, reducing overall system complexity while preserving fine-grained knowledge.
4Quantity of substance
If symbolic knowledge graphs are incorporated into task-oriented dialogue, then external knowledge is available, but language expression mismatches occur
Solution Approach 1:
The patent replaces the rigid symbolic matching mechanism with flexible neural embedding-based similarity computation. Instead of requiring exact symbol matches between language expressions and knowledge graph nodes, the system uses vector similarity measures that can handle paraphrases, synonyms, and varied language expressions, greatly improving ease of operation.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented system and method relate to natural language processing and knowledge representation and reasoning. A first dataset is created that includes input data and situational data. The situational data provides context for the input data. An encoder is configured to generate an encoded representation of the first dataset. The encoder includes at least an encoding network of a first pre-trained generative machine learning model, which relates to a generative knowledge graph. A decoder includes a decoding network of a second pre-trained generative machine learning model. The decoder is configured to generate response data based on the first dataset by decoding the encoded representation. The decoder is also configured to generate event-centric knowledge based on the first dataset by decoding the encoded representation. The input data and the response data are connected to the same event-centric knowledge via the generative knowledge graph. For example, the event-centric knowledge includes goal data, which is inferred from the input data and the situational data.