Utterance-Level Graph Relation Extraction for Multi-Speaker Dialogue
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing natural language processing systems struggle to effectively process conversation data with multiple speakers and utterances in colloquial language, as they lack a specialized methodology to handle cross-references and length variations, making it difficult to determine the appropriate focus for task-solving.
Innovation Solution
A dialogue relation extraction method using utterance-level graph computation, employing a graph convolutional network (GCN) and a prompt-based learning methodology to generate a graph from conversation data, incorporating adjacency and node feature matrices, and utilizing a pre-trained language model to derive relations between argument pairs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If sentence-level or single document-level data is used for relation extraction, then the processing is simple, but it is difficult to handle conversation data with multiple speakers and utterances
Solution Approach 1:
The patent segments the conversation data into multiple utterances, each treated as a separate node in the graph structure. This allows the system to process conversation data with multiple speakers and utterances by maintaining discrete, manageable units while capturing their relationships through graph connections.
Solution Approach 2:
The patent transitions from traditional sentence-level or document-level processing to an utterance-level graph structure. This dimensional change enables the system to handle conversation data by representing utterances as nodes with features and relationships, adding a structural dimension that captures multi-speaker dynamics.
2Adaptability or versatility
If conversation data with cross-references and length variations is used, then the system can process real dialogue, but it is difficult to determine which part to focus on for task solving
Solution Approach 1:
The patent introduces an advisor node as an intermediary element in the graph structure. This advisor node helps the system determine which parts of the conversation to focus on by mediating between the multiple utterances and the task objectives, enabling effective handling of colloquial language with cross-references.
3Reliability
If graph convolutional network with prompt-based learning is used, then the relation extraction performance is improved, but the system complexity increases
Solution Approach 1:
The patent merges the graph convolutional network architecture with prompt-based learning methodology into a unified system. This combination integrates the relational reasoning capabilities of GCN with the knowledge utilization of prompt-based learning, improving relation extraction performance by combining complementary strengths.
Solution Approach 2:
The patent creates a multi-functional system where the graph convolutional network handles relational reasoning, the prompt-based learning handles knowledge utilization, and the advisor node handles focus determination. This universal approach allows a single system to perform multiple functions related to dialogue relation extraction.
Data Source
AI summary
Disclosed is a device and method for dialogue relation extraction using utterance-level graph computation. The dialogue relation extraction method refers to a dialogue relation extraction method performed by a computing device including at least a processor and includes receiving a target conversation that includes a plurality of utterances and an argument pair that is a target of relation extraction; generating a graph (G=(A, X)) that includes an adjacency matrix (A) and a node feature matrix (X) based on the target conversation and the argument pair; and deriving a relation between subject and object included in the argument pair by inputting the graph to a graph convolutional network (GCN) trained to infer the relation of the argument pair.