Slot Recognition via Knowledge Graph Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-round slot recognition strategies face challenges in universality, robustness, maintainability, and expandability, particularly due to the complexity of understanding user intentions and the need for complex combinational logic or rigid manual rules, and limited ability to resolve behavior verbs and descriptive words.
Innovation Solution
A method and apparatus that utilize a knowledge graph to infer entity words by detecting entity slots and relationship-determinative words in queries, replacing demonstrative pronouns with recent session entities, and searching for peripheral knowledge graphs to output matching entity words, reducing the need for state machines and simplifying slot recognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If conventional sequence labeling is used for slot recognition, then the system can identify basic entities, but it cannot recognize richer meanings and requires complex combinational logic for backend search
Solution Approach 1:
The patent introduces an intermediary component called a 'session state machine' that maintains session state information between multiple user interactions. This state machine acts as a mediator that stores and manages slot information across different conversation rounds, allowing the system to resolve anaphoric references without requiring complex combinational logic in the backend search service.
Solution Approach 2:
The system performs preliminary slot recognition and state maintenance during the conversation session itself, before the backend search is triggered. By pre-processing and storing slot information in the session state machine, the system prepares the data structure in advance, eliminating the need for complex real-time logic during the search operation.
2Loss of information
If a session state machine is maintained to record all sessions and replace demonstrative pronouns, then some noun slots can be replaced, but the ability to resolve behavior verbs and descriptive words is limited
Solution Approach 1:
The patent enhances the session state machine to perform multiple functions: it not only replaces demonstrative pronouns but also resolves behavior verbs and descriptive words. The state machine is designed to handle various types of anaphoric references uniformly, making it a universal solution that works across different slot types without requiring separate handling mechanisms.
Solution Approach 2:
The system uses feedback from the session context to improve anaphora resolution. By continuously updating the session state machine with information from each interaction round, the system accumulates context that enables more accurate resolution of various types of references, including behavior verbs and descriptive words that were previously difficult to resolve.
3Productivity
If direct slot sequence labeling is performed without anaphora resolution, then words with reference relationships are directly outputted, but the application party must write complex combinational logic to understand user requirements
Solution Approach 1:
The session state machine serves as an intermediary that automatically performs anaphora resolution and semantic extension, shielding the application party from the complexity of understanding user requirements. The state machine processes and enriches slot information before passing it to the application layer, so developers don't need to write complex combinational logic to handle references.
Solution Approach 2:
The system performs self-service by automatically resolving anaphoric references and enriching slot meanings through the session state machine. This automation eliminates the need for manual complex logic programming, as the system independently handles the understanding of user requirements across multiple conversation turns.
Data Source
AI summary
Embodiments of the present disclosure provide a method and apparatus for outputting information. A specific embodiment of the method includes: in response to receiving a query, detecting whether there is an entity slot in the query; in response to there being an entity slot in the query, adding the detected entity slot to a candidate slot; detecting, in the query, a relationship-determinative word of an entity; searching in a preset knowledge graph for a peripheral knowledge graph of the candidate slot; and inferring on the basis of the peripheral knowledge graph according to the relationship-determinative word, and outputting an entity word matching the relationship-determinative word.